1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86ShuffleDecodeConstantPool.h"
22 #include "X86TargetMachine.h"
23 #include "X86TargetObjectFile.h"
24 #include "llvm/ADT/SmallBitVector.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/ADT/StringSwitch.h"
29 #include "llvm/Analysis/EHPersonalities.h"
30 #include "llvm/CodeGen/IntrinsicLowering.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/WinEHFuncInfo.h"
38 #include "llvm/IR/CallSite.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalAlias.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/MC/MCAsmInfo.h"
48 #include "llvm/MC/MCContext.h"
49 #include "llvm/MC/MCExpr.h"
50 #include "llvm/MC/MCSymbol.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Target/TargetOptions.h"
56 #include "X86IntrinsicsInfo.h"
62 #define DEBUG_TYPE "x86-isel"
64 STATISTIC(NumTailCalls, "Number of tail calls");
66 static cl::opt<bool> ExperimentalVectorWideningLegalization(
67 "x86-experimental-vector-widening-legalization", cl::init(false),
68 cl::desc("Enable an experimental vector type legalization through widening "
69 "rather than promotion."),
72 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
73 const X86Subtarget &STI)
74 : TargetLowering(TM), Subtarget(&STI) {
75 X86ScalarSSEf64 = Subtarget->hasSSE2();
76 X86ScalarSSEf32 = Subtarget->hasSSE1();
77 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
79 // Set up the TargetLowering object.
81 // X86 is weird. It always uses i8 for shift amounts and setcc results.
82 setBooleanContents(ZeroOrOneBooleanContent);
83 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
84 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
86 // For 64-bit, since we have so many registers, use the ILP scheduler.
87 // For 32-bit, use the register pressure specific scheduling.
88 // For Atom, always use ILP scheduling.
89 if (Subtarget->isAtom())
90 setSchedulingPreference(Sched::ILP);
91 else if (Subtarget->is64Bit())
92 setSchedulingPreference(Sched::ILP);
94 setSchedulingPreference(Sched::RegPressure);
95 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
96 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
98 // Bypass expensive divides on Atom when compiling with O2.
99 if (TM.getOptLevel() >= CodeGenOpt::Default) {
100 if (Subtarget->hasSlowDivide32())
101 addBypassSlowDiv(32, 8);
102 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
103 addBypassSlowDiv(64, 16);
106 if (Subtarget->isTargetKnownWindowsMSVC()) {
107 // Setup Windows compiler runtime calls.
108 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
109 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
110 setLibcallName(RTLIB::SREM_I64, "_allrem");
111 setLibcallName(RTLIB::UREM_I64, "_aullrem");
112 setLibcallName(RTLIB::MUL_I64, "_allmul");
113 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
117 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
120 if (Subtarget->isTargetDarwin()) {
121 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
122 setUseUnderscoreSetJmp(false);
123 setUseUnderscoreLongJmp(false);
124 } else if (Subtarget->isTargetWindowsGNU()) {
125 // MS runtime is weird: it exports _setjmp, but longjmp!
126 setUseUnderscoreSetJmp(true);
127 setUseUnderscoreLongJmp(false);
129 setUseUnderscoreSetJmp(true);
130 setUseUnderscoreLongJmp(true);
133 // Set up the register classes.
134 addRegisterClass(MVT::i8, &X86::GR8RegClass);
135 addRegisterClass(MVT::i16, &X86::GR16RegClass);
136 addRegisterClass(MVT::i32, &X86::GR32RegClass);
137 if (Subtarget->is64Bit())
138 addRegisterClass(MVT::i64, &X86::GR64RegClass);
140 for (MVT VT : MVT::integer_valuetypes())
141 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
143 // We don't accept any truncstore of integer registers.
144 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
146 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
147 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
148 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
149 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
151 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
153 // SETOEQ and SETUNE require checking two conditions.
154 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
156 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
159 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
161 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
163 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
165 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
167 if (Subtarget->is64Bit()) {
168 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
169 // f32/f64 are legal, f80 is custom.
170 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
173 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
174 } else if (!Subtarget->useSoftFloat()) {
175 // We have an algorithm for SSE2->double, and we turn this into a
176 // 64-bit FILD followed by conditional FADD for other targets.
177 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
178 // We have an algorithm for SSE2, and we turn this into a 64-bit
179 // FILD or VCVTUSI2SS/SD for other targets.
180 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
183 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
185 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
186 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
188 if (!Subtarget->useSoftFloat()) {
189 // SSE has no i16 to fp conversion, only i32
190 if (X86ScalarSSEf32) {
191 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
192 // f32 and f64 cases are Legal, f80 case is not
193 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
196 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
200 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
203 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
205 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
208 if (!Subtarget->useSoftFloat()) {
209 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
210 // are Legal, f80 is custom lowered.
211 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
212 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
214 if (X86ScalarSSEf32) {
215 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
216 // f32 and f64 cases are Legal, f80 case is not
217 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
219 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
220 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
224 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand);
225 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand);
228 // Handle FP_TO_UINT by promoting the destination to a larger signed
230 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
234 if (Subtarget->is64Bit()) {
235 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
236 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
237 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
238 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
240 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
241 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
243 } else if (!Subtarget->useSoftFloat()) {
244 // Since AVX is a superset of SSE3, only check for SSE here.
245 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
246 // Expand FP_TO_UINT into a select.
247 // FIXME: We would like to use a Custom expander here eventually to do
248 // the optimal thing for SSE vs. the default expansion in the legalizer.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
251 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
252 // With SSE3 we can use fisttpll to convert to a signed i64; without
253 // SSE, we're stuck with a fistpll.
254 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
256 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
259 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
260 if (!X86ScalarSSEf64) {
261 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
262 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
263 if (Subtarget->is64Bit()) {
264 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
265 // Without SSE, i64->f64 goes through memory.
266 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
268 } else if (!Subtarget->is64Bit())
269 setOperationAction(ISD::BITCAST , MVT::i64 , Custom);
271 // Scalar integer divide and remainder are lowered to use operations that
272 // produce two results, to match the available instructions. This exposes
273 // the two-result form to trivial CSE, which is able to combine x/y and x%y
274 // into a single instruction.
276 // Scalar integer multiply-high is also lowered to use two-result
277 // operations, to match the available instructions. However, plain multiply
278 // (low) operations are left as Legal, as there are single-result
279 // instructions for this in x86. Using the two-result multiply instructions
280 // when both high and low results are needed must be arranged by dagcombine.
281 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
282 setOperationAction(ISD::MULHS, VT, Expand);
283 setOperationAction(ISD::MULHU, VT, Expand);
284 setOperationAction(ISD::SDIV, VT, Expand);
285 setOperationAction(ISD::UDIV, VT, Expand);
286 setOperationAction(ISD::SREM, VT, Expand);
287 setOperationAction(ISD::UREM, VT, Expand);
289 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
290 setOperationAction(ISD::ADDC, VT, Custom);
291 setOperationAction(ISD::ADDE, VT, Custom);
292 setOperationAction(ISD::SUBC, VT, Custom);
293 setOperationAction(ISD::SUBE, VT, Custom);
296 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
297 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
298 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
299 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
300 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
301 setOperationAction(ISD::BR_CC , MVT::f128, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
303 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
304 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
305 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::f128, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
312 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
313 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
314 if (Subtarget->is64Bit())
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
318 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
319 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
321 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
322 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
323 // is. We should promote the value to 64-bits to solve this.
324 // This is what the CRT headers do - `fmodf` is an inline header
325 // function casting to f64 and calling `fmod`.
326 setOperationAction(ISD::FREM , MVT::f32 , Promote);
328 setOperationAction(ISD::FREM , MVT::f32 , Expand);
331 setOperationAction(ISD::FREM , MVT::f64 , Expand);
332 setOperationAction(ISD::FREM , MVT::f80 , Expand);
333 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
335 // Promote the i8 variants and force them on up to i32 which has a shorter
337 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
338 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
339 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
340 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
341 if (Subtarget->hasBMI()) {
342 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
343 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
344 if (Subtarget->is64Bit())
345 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
347 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
348 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
349 if (Subtarget->is64Bit())
350 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
353 if (Subtarget->hasLZCNT()) {
354 // When promoting the i8 variants, force them to i32 for a shorter
356 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
357 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
358 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
359 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
362 if (Subtarget->is64Bit())
363 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
365 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
366 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
367 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
368 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
369 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
370 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
371 if (Subtarget->is64Bit()) {
372 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
373 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
377 // Special handling for half-precision floating point conversions.
378 // If we don't have F16C support, then lower half float conversions
379 // into library calls.
380 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
381 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
382 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
385 // There's never any support for operations beyond MVT::f32.
386 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
387 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
388 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
389 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
391 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
392 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
393 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
394 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
395 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
396 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
398 if (Subtarget->hasPOPCNT()) {
399 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
401 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
402 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
403 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
404 if (Subtarget->is64Bit())
405 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
408 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
410 if (!Subtarget->hasMOVBE())
411 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
413 // These should be promoted to a larger select which is supported.
414 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
415 // X86 wants to expand cmov itself.
416 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
417 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
418 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
419 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
420 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
421 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
422 setOperationAction(ISD::SELECT , MVT::f128 , Custom);
423 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
424 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
425 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
426 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
427 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
428 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
429 setOperationAction(ISD::SETCC , MVT::f128 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
431 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
432 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
433 if (Subtarget->is64Bit()) {
434 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
435 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
436 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
438 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
439 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
440 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
441 // support continuation, user-level threading, and etc.. As a result, no
442 // other SjLj exception interfaces are implemented and please don't build
443 // your own exception handling based on them.
444 // LLVM/Clang supports zero-cost DWARF exception handling.
445 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
446 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
449 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
450 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
451 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
452 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
453 if (Subtarget->is64Bit())
454 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
455 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
456 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
457 if (Subtarget->is64Bit()) {
458 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
459 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
460 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
461 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
462 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
464 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
465 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
466 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
467 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
468 if (Subtarget->is64Bit()) {
469 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
470 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
471 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
474 if (Subtarget->hasSSE1())
475 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
477 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
479 // Expand certain atomics
480 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
481 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
482 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
483 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
486 if (Subtarget->hasCmpxchg16b()) {
487 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
490 // FIXME - use subtarget debug flags
491 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
492 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
493 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
496 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
497 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
499 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
500 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
502 setOperationAction(ISD::TRAP, MVT::Other, Legal);
503 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
505 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
506 setOperationAction(ISD::VASTART , MVT::Other, Custom);
507 setOperationAction(ISD::VAEND , MVT::Other, Expand);
508 if (Subtarget->is64Bit()) {
509 setOperationAction(ISD::VAARG , MVT::Other, Custom);
510 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
512 // TargetInfo::CharPtrBuiltinVaList
513 setOperationAction(ISD::VAARG , MVT::Other, Expand);
514 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
517 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
518 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
520 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
522 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
523 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
524 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
526 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
527 // f32 and f64 use SSE.
528 // Set up the FP register classes.
529 addRegisterClass(MVT::f32, &X86::FR32RegClass);
530 addRegisterClass(MVT::f64, &X86::FR64RegClass);
532 // Use ANDPD to simulate FABS.
533 setOperationAction(ISD::FABS , MVT::f64, Custom);
534 setOperationAction(ISD::FABS , MVT::f32, Custom);
536 // Use XORP to simulate FNEG.
537 setOperationAction(ISD::FNEG , MVT::f64, Custom);
538 setOperationAction(ISD::FNEG , MVT::f32, Custom);
540 // Use ANDPD and ORPD to simulate FCOPYSIGN.
541 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
542 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
544 // Lower this to FGETSIGNx86 plus an AND.
545 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
546 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
548 // We don't support sin/cos/fmod
549 setOperationAction(ISD::FSIN , MVT::f64, Expand);
550 setOperationAction(ISD::FCOS , MVT::f64, Expand);
551 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
552 setOperationAction(ISD::FSIN , MVT::f32, Expand);
553 setOperationAction(ISD::FCOS , MVT::f32, Expand);
554 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
556 // Expand FP immediates into loads from the stack, except for the special
558 addLegalFPImmediate(APFloat(+0.0)); // xorpd
559 addLegalFPImmediate(APFloat(+0.0f)); // xorps
560 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
561 // Use SSE for f32, x87 for f64.
562 // Set up the FP register classes.
563 addRegisterClass(MVT::f32, &X86::FR32RegClass);
564 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
566 // Use ANDPS to simulate FABS.
567 setOperationAction(ISD::FABS , MVT::f32, Custom);
569 // Use XORP to simulate FNEG.
570 setOperationAction(ISD::FNEG , MVT::f32, Custom);
572 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
574 // Use ANDPS and ORPS to simulate FCOPYSIGN.
575 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
576 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
578 // We don't support sin/cos/fmod
579 setOperationAction(ISD::FSIN , MVT::f32, Expand);
580 setOperationAction(ISD::FCOS , MVT::f32, Expand);
581 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
583 // Special cases we handle for FP constants.
584 addLegalFPImmediate(APFloat(+0.0f)); // xorps
585 addLegalFPImmediate(APFloat(+0.0)); // FLD0
586 addLegalFPImmediate(APFloat(+1.0)); // FLD1
587 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
588 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
590 if (!TM.Options.UnsafeFPMath) {
591 setOperationAction(ISD::FSIN , MVT::f64, Expand);
592 setOperationAction(ISD::FCOS , MVT::f64, Expand);
593 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
595 } else if (!Subtarget->useSoftFloat()) {
596 // f32 and f64 in x87.
597 // Set up the FP register classes.
598 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
599 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
601 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
602 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
603 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
604 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
606 if (!TM.Options.UnsafeFPMath) {
607 setOperationAction(ISD::FSIN , MVT::f64, Expand);
608 setOperationAction(ISD::FSIN , MVT::f32, Expand);
609 setOperationAction(ISD::FCOS , MVT::f64, Expand);
610 setOperationAction(ISD::FCOS , MVT::f32, Expand);
611 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
612 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
614 addLegalFPImmediate(APFloat(+0.0)); // FLD0
615 addLegalFPImmediate(APFloat(+1.0)); // FLD1
616 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
617 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
618 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
619 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
620 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
621 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
624 // We don't support FMA.
625 setOperationAction(ISD::FMA, MVT::f64, Expand);
626 setOperationAction(ISD::FMA, MVT::f32, Expand);
628 // Long double always uses X87, except f128 in MMX.
629 if (!Subtarget->useSoftFloat()) {
630 if (Subtarget->is64Bit() && Subtarget->hasMMX()) {
631 addRegisterClass(MVT::f128, &X86::FR128RegClass);
632 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
633 setOperationAction(ISD::FABS , MVT::f128, Custom);
634 setOperationAction(ISD::FNEG , MVT::f128, Custom);
635 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
638 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
639 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
640 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
642 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
643 addLegalFPImmediate(TmpFlt); // FLD0
645 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
648 APFloat TmpFlt2(+1.0);
649 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
651 addLegalFPImmediate(TmpFlt2); // FLD1
652 TmpFlt2.changeSign();
653 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
656 if (!TM.Options.UnsafeFPMath) {
657 setOperationAction(ISD::FSIN , MVT::f80, Expand);
658 setOperationAction(ISD::FCOS , MVT::f80, Expand);
659 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
662 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
663 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
664 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
665 setOperationAction(ISD::FRINT, MVT::f80, Expand);
666 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
667 setOperationAction(ISD::FMA, MVT::f80, Expand);
670 // Always use a library call for pow.
671 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
672 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
673 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
675 setOperationAction(ISD::FLOG, MVT::f80, Expand);
676 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
677 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
678 setOperationAction(ISD::FEXP, MVT::f80, Expand);
679 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
680 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
681 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
683 // First set operation action for all vector types to either promote
684 // (for widening) or expand (for scalarization). Then we will selectively
685 // turn on ones that can be effectively codegen'd.
686 for (MVT VT : MVT::vector_valuetypes()) {
687 setOperationAction(ISD::ADD , VT, Expand);
688 setOperationAction(ISD::SUB , VT, Expand);
689 setOperationAction(ISD::FADD, VT, Expand);
690 setOperationAction(ISD::FNEG, VT, Expand);
691 setOperationAction(ISD::FSUB, VT, Expand);
692 setOperationAction(ISD::MUL , VT, Expand);
693 setOperationAction(ISD::FMUL, VT, Expand);
694 setOperationAction(ISD::SDIV, VT, Expand);
695 setOperationAction(ISD::UDIV, VT, Expand);
696 setOperationAction(ISD::FDIV, VT, Expand);
697 setOperationAction(ISD::SREM, VT, Expand);
698 setOperationAction(ISD::UREM, VT, Expand);
699 setOperationAction(ISD::LOAD, VT, Expand);
700 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
701 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
702 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
703 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
704 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
705 setOperationAction(ISD::FABS, VT, Expand);
706 setOperationAction(ISD::FSIN, VT, Expand);
707 setOperationAction(ISD::FSINCOS, VT, Expand);
708 setOperationAction(ISD::FCOS, VT, Expand);
709 setOperationAction(ISD::FSINCOS, VT, Expand);
710 setOperationAction(ISD::FREM, VT, Expand);
711 setOperationAction(ISD::FMA, VT, Expand);
712 setOperationAction(ISD::FPOWI, VT, Expand);
713 setOperationAction(ISD::FSQRT, VT, Expand);
714 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
715 setOperationAction(ISD::FFLOOR, VT, Expand);
716 setOperationAction(ISD::FCEIL, VT, Expand);
717 setOperationAction(ISD::FTRUNC, VT, Expand);
718 setOperationAction(ISD::FRINT, VT, Expand);
719 setOperationAction(ISD::FNEARBYINT, VT, Expand);
720 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
721 setOperationAction(ISD::MULHS, VT, Expand);
722 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
723 setOperationAction(ISD::MULHU, VT, Expand);
724 setOperationAction(ISD::SDIVREM, VT, Expand);
725 setOperationAction(ISD::UDIVREM, VT, Expand);
726 setOperationAction(ISD::FPOW, VT, Expand);
727 setOperationAction(ISD::CTPOP, VT, Expand);
728 setOperationAction(ISD::CTTZ, VT, Expand);
729 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
730 setOperationAction(ISD::CTLZ, VT, Expand);
731 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
732 setOperationAction(ISD::SHL, VT, Expand);
733 setOperationAction(ISD::SRA, VT, Expand);
734 setOperationAction(ISD::SRL, VT, Expand);
735 setOperationAction(ISD::ROTL, VT, Expand);
736 setOperationAction(ISD::ROTR, VT, Expand);
737 setOperationAction(ISD::BSWAP, VT, Expand);
738 setOperationAction(ISD::SETCC, VT, Expand);
739 setOperationAction(ISD::FLOG, VT, Expand);
740 setOperationAction(ISD::FLOG2, VT, Expand);
741 setOperationAction(ISD::FLOG10, VT, Expand);
742 setOperationAction(ISD::FEXP, VT, Expand);
743 setOperationAction(ISD::FEXP2, VT, Expand);
744 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
745 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
746 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
747 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
748 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
749 setOperationAction(ISD::TRUNCATE, VT, Expand);
750 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
751 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
752 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
753 setOperationAction(ISD::VSELECT, VT, Expand);
754 setOperationAction(ISD::SELECT_CC, VT, Expand);
755 for (MVT InnerVT : MVT::vector_valuetypes()) {
756 setTruncStoreAction(InnerVT, VT, Expand);
758 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
759 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
761 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
762 // types, we have to deal with them whether we ask for Expansion or not.
763 // Setting Expand causes its own optimisation problems though, so leave
765 if (VT.getVectorElementType() == MVT::i1)
766 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
768 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
769 // split/scalarized right now.
770 if (VT.getVectorElementType() == MVT::f16)
771 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
775 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
776 // with -msoft-float, disable use of MMX as well.
777 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
778 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
779 // No operations on x86mmx supported, everything uses intrinsics.
782 // MMX-sized vectors (other than x86mmx) are expected to be expanded
783 // into smaller operations.
784 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
785 setOperationAction(ISD::MULHS, MMXTy, Expand);
786 setOperationAction(ISD::AND, MMXTy, Expand);
787 setOperationAction(ISD::OR, MMXTy, Expand);
788 setOperationAction(ISD::XOR, MMXTy, Expand);
789 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
790 setOperationAction(ISD::SELECT, MMXTy, Expand);
791 setOperationAction(ISD::BITCAST, MMXTy, Expand);
793 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
795 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
796 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
798 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
799 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
800 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
801 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
802 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
803 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
804 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
805 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
806 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
807 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
808 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
809 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
810 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
811 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
814 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
815 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
817 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
818 // registers cannot be used even for integer operations.
819 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
820 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
821 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
822 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
824 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
825 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
826 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
827 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
828 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
829 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
830 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
831 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
832 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
833 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
834 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
835 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
836 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
837 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
838 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
839 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
840 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
841 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
842 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
843 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
844 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
845 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
846 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
848 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
849 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
850 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
851 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
853 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
854 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
855 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
856 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
858 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
859 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
861 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
864 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
865 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
866 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
867 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
869 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
870 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
871 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
872 // ISD::CTTZ v2i64 - scalarization is faster.
873 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
874 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
875 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
876 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
878 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
879 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
880 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
881 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
882 setOperationAction(ISD::VSELECT, VT, Custom);
883 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
886 // We support custom legalizing of sext and anyext loads for specific
887 // memory vector types which we can load as a scalar (or sequence of
888 // scalars) and extend in-register to a legal 128-bit vector type. For sext
889 // loads these must work with a single scalar load.
890 for (MVT VT : MVT::integer_vector_valuetypes()) {
891 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
892 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
893 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
894 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
895 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
896 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
897 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
898 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
899 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
902 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
903 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
904 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
905 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
906 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
907 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
908 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
909 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
911 if (Subtarget->is64Bit()) {
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
913 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
916 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
917 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
918 setOperationAction(ISD::AND, VT, Promote);
919 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
920 setOperationAction(ISD::OR, VT, Promote);
921 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
922 setOperationAction(ISD::XOR, VT, Promote);
923 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
924 setOperationAction(ISD::LOAD, VT, Promote);
925 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
926 setOperationAction(ISD::SELECT, VT, Promote);
927 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
930 // Custom lower v2i64 and v2f64 selects.
931 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
932 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
933 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
934 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
936 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
937 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
939 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
941 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
942 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
943 // As there is no 64-bit GPR available, we need build a special custom
944 // sequence to convert from v2i32 to v2f32.
945 if (!Subtarget->is64Bit())
946 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
948 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
949 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
951 for (MVT VT : MVT::fp_vector_valuetypes())
952 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
954 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
955 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
956 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
959 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
960 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
961 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
962 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
963 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
964 setOperationAction(ISD::FRINT, RoundedTy, Legal);
965 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
968 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
969 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
970 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
971 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
972 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
973 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
974 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
975 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
977 // FIXME: Do we need to handle scalar-to-vector here?
978 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
980 // We directly match byte blends in the backend as they match the VSELECT
982 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
984 // SSE41 brings specific instructions for doing vector sign extend even in
985 // cases where we don't have SRA.
986 for (MVT VT : MVT::integer_vector_valuetypes()) {
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
988 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
989 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
992 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
997 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
998 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1004 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1005 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1007 // i8 and i16 vectors are custom because the source register and source
1008 // source memory operand types are not the same width. f32 vectors are
1009 // custom since the immediate controlling the insert encodes additional
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1013 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1014 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1018 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1019 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1021 // FIXME: these should be Legal, but that's only for the case where
1022 // the index is constant. For now custom expand to deal with that.
1023 if (Subtarget->is64Bit()) {
1024 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1025 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1029 if (Subtarget->hasSSE2()) {
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1031 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1032 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1034 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1035 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1037 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1038 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1040 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1041 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1043 // In the customized shift lowering, the legal cases in AVX2 will be
1045 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1046 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1048 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1049 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1051 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1052 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1055 if (Subtarget->hasXOP()) {
1056 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1062 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1063 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1066 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1067 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1071 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1072 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1074 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1075 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1076 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1078 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1087 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1088 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1089 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1091 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1100 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1101 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1102 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1104 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1105 // even though v8i16 is a legal type.
1106 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1107 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1108 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1110 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1111 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1112 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1114 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1115 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1117 for (MVT VT : MVT::fp_vector_valuetypes())
1118 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1120 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1121 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1123 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1124 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1126 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1127 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1131 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1132 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1135 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1136 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1139 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1140 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1142 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1143 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1145 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1146 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1148 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1149 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1153 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1154 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1158 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1159 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1162 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1163 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1165 if (Subtarget->hasAnyFMA()) {
1166 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1167 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1168 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1169 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1170 setOperationAction(ISD::FMA, MVT::f32, Legal);
1171 setOperationAction(ISD::FMA, MVT::f64, Legal);
1174 if (Subtarget->hasInt256()) {
1175 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1176 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1177 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1178 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1180 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1181 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1182 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1183 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1185 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1186 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1187 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1188 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1190 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1191 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1192 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1193 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1196 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1197 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1199 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1200 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1202 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1203 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1205 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1206 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1208 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1209 // when we have a 256bit-wide blend with immediate.
1210 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1212 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1217 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1218 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1224 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1225 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1227 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1228 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1229 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1230 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1232 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1233 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1234 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1235 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1237 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1238 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1239 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1240 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1243 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1244 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1246 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1247 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1249 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1250 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1252 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1253 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1256 // In the customized shift lowering, the legal cases in AVX2 will be
1258 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1259 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1261 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1262 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1264 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1265 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1267 // Custom lower several nodes for 256-bit types.
1268 for (MVT VT : MVT::vector_valuetypes()) {
1269 if (VT.getScalarSizeInBits() >= 32) {
1270 setOperationAction(ISD::MLOAD, VT, Legal);
1271 setOperationAction(ISD::MSTORE, VT, Legal);
1273 // Extract subvector is special because the value type
1274 // (result) is 128-bit but the source is 256-bit wide.
1275 if (VT.is128BitVector()) {
1276 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1278 // Do not attempt to custom lower other non-256-bit vectors
1279 if (!VT.is256BitVector())
1282 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1283 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1284 setOperationAction(ISD::VSELECT, VT, Custom);
1285 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1286 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1287 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1288 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1289 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1292 if (Subtarget->hasInt256())
1293 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1295 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1296 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1297 setOperationAction(ISD::AND, VT, Promote);
1298 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1299 setOperationAction(ISD::OR, VT, Promote);
1300 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1301 setOperationAction(ISD::XOR, VT, Promote);
1302 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1303 setOperationAction(ISD::LOAD, VT, Promote);
1304 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1305 setOperationAction(ISD::SELECT, VT, Promote);
1306 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1310 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1311 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1312 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1313 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1314 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1316 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1317 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1318 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1320 for (MVT VT : MVT::fp_vector_valuetypes())
1321 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1323 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1324 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1325 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1326 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1327 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1328 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1329 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1330 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1331 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1332 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1333 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1334 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1336 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1337 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1338 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1339 setOperationAction(ISD::XOR, MVT::i1, Legal);
1340 setOperationAction(ISD::OR, MVT::i1, Legal);
1341 setOperationAction(ISD::AND, MVT::i1, Legal);
1342 setOperationAction(ISD::SUB, MVT::i1, Custom);
1343 setOperationAction(ISD::ADD, MVT::i1, Custom);
1344 setOperationAction(ISD::MUL, MVT::i1, Custom);
1345 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1346 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1348 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1349 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1351 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1355 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1356 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1357 setOperationAction(ISD::FABS, MVT::v16f32, Custom);
1359 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1363 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1364 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1365 setOperationAction(ISD::FABS, MVT::v8f64, Custom);
1366 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1367 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1369 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1371 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1372 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1376 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1377 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1381 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1382 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1383 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1384 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1387 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1388 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1389 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1390 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1391 if (Subtarget->hasVLX()){
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1393 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1394 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1395 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1396 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1399 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1400 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1401 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1402 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1404 setOperationAction(ISD::MLOAD, MVT::v8i32, Custom);
1405 setOperationAction(ISD::MLOAD, MVT::v8f32, Custom);
1406 setOperationAction(ISD::MSTORE, MVT::v8i32, Custom);
1407 setOperationAction(ISD::MSTORE, MVT::v8f32, Custom);
1409 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1410 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1411 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1412 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1413 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1414 if (Subtarget->hasDQI()) {
1415 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1416 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1418 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1419 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1420 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1421 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1422 if (Subtarget->hasVLX()) {
1423 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1424 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1425 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1426 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1427 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1428 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1429 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1430 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1433 if (Subtarget->hasVLX()) {
1434 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1435 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1436 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1437 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1438 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1439 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1440 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1441 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1443 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1444 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1445 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1446 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1447 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1448 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1449 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1452 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1453 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1454 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1455 if (Subtarget->hasDQI()) {
1456 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1457 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1459 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1460 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1461 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1462 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1463 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1464 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1465 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1466 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1467 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1468 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1474 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
1476 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1477 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1479 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1482 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom);
1484 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1485 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1486 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1487 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1490 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1491 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1492 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1494 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1495 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1496 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1497 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1498 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1499 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1500 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1501 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1503 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1504 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1506 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1507 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1509 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1511 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1512 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1514 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1515 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1517 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1518 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1520 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1521 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1522 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1523 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1524 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1525 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1527 if (Subtarget->hasCDI()) {
1528 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1529 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Expand);
1531 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Expand);
1533 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1535 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1536 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1537 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Expand);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Expand);
1539 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Expand);
1540 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Expand);
1542 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1543 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1545 if (Subtarget->hasVLX()) {
1546 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1547 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1548 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1549 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1550 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1551 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1552 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1553 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1555 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1556 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1557 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1558 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1560 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1561 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1562 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1563 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1564 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1565 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1566 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1567 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1569 } // Subtarget->hasCDI()
1571 if (Subtarget->hasDQI()) {
1572 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1573 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1574 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1576 // Custom lower several nodes.
1577 for (MVT VT : MVT::vector_valuetypes()) {
1578 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1580 setOperationAction(ISD::AND, VT, Legal);
1581 setOperationAction(ISD::OR, VT, Legal);
1582 setOperationAction(ISD::XOR, VT, Legal);
1584 if ((VT.is128BitVector() || VT.is256BitVector()) && EltSize >= 32) {
1585 setOperationAction(ISD::MGATHER, VT, Custom);
1586 setOperationAction(ISD::MSCATTER, VT, Custom);
1588 // Extract subvector is special because the value type
1589 // (result) is 256/128-bit but the source is 512-bit wide.
1590 if (VT.is128BitVector() || VT.is256BitVector()) {
1591 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1593 if (VT.getVectorElementType() == MVT::i1)
1594 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1596 // Do not attempt to custom lower other non-512-bit vectors
1597 if (!VT.is512BitVector())
1600 if (EltSize >= 32) {
1601 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1602 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1603 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1604 setOperationAction(ISD::VSELECT, VT, Legal);
1605 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1606 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1607 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1608 setOperationAction(ISD::MLOAD, VT, Legal);
1609 setOperationAction(ISD::MSTORE, VT, Legal);
1610 setOperationAction(ISD::MGATHER, VT, Legal);
1611 setOperationAction(ISD::MSCATTER, VT, Custom);
1614 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1615 setOperationAction(ISD::SELECT, VT, Promote);
1616 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1620 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1621 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1622 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1624 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1625 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1627 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1628 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1629 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1630 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1631 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1632 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1633 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1634 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1635 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1636 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1637 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1640 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1641 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1644 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1645 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1646 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1647 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1648 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1649 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1650 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1651 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1652 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1653 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1655 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1656 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1657 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1660 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1661 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1662 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1663 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1664 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1665 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1666 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1668 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1670 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1671 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1672 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1673 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1674 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1675 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1676 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1677 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1679 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1680 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1681 if (Subtarget->hasVLX())
1682 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1684 if (Subtarget->hasCDI()) {
1685 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1686 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1687 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Expand);
1688 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Expand);
1691 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1692 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1693 setOperationAction(ISD::VSELECT, VT, Legal);
1694 setOperationAction(ISD::SRL, VT, Custom);
1695 setOperationAction(ISD::SHL, VT, Custom);
1696 setOperationAction(ISD::SRA, VT, Custom);
1698 setOperationAction(ISD::AND, VT, Promote);
1699 AddPromotedToType (ISD::AND, VT, MVT::v8i64);
1700 setOperationAction(ISD::OR, VT, Promote);
1701 AddPromotedToType (ISD::OR, VT, MVT::v8i64);
1702 setOperationAction(ISD::XOR, VT, Promote);
1703 AddPromotedToType (ISD::XOR, VT, MVT::v8i64);
1707 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1708 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1709 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1711 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1712 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1713 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1714 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1715 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1716 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1717 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1718 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1719 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1720 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1721 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1722 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1724 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1725 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1726 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1727 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1728 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1729 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1730 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1731 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1733 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1734 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1735 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1736 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1737 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1738 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1739 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1740 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1743 // We want to custom lower some of our intrinsics.
1744 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1745 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1746 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1747 if (!Subtarget->is64Bit()) {
1748 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1749 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1752 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1753 // handle type legalization for these operations here.
1755 // FIXME: We really should do custom legalization for addition and
1756 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1757 // than generic legalization for 64-bit multiplication-with-overflow, though.
1758 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1759 if (VT == MVT::i64 && !Subtarget->is64Bit())
1761 // Add/Sub/Mul with overflow operations are custom lowered.
1762 setOperationAction(ISD::SADDO, VT, Custom);
1763 setOperationAction(ISD::UADDO, VT, Custom);
1764 setOperationAction(ISD::SSUBO, VT, Custom);
1765 setOperationAction(ISD::USUBO, VT, Custom);
1766 setOperationAction(ISD::SMULO, VT, Custom);
1767 setOperationAction(ISD::UMULO, VT, Custom);
1770 if (!Subtarget->is64Bit()) {
1771 // These libcalls are not available in 32-bit.
1772 setLibcallName(RTLIB::SHL_I128, nullptr);
1773 setLibcallName(RTLIB::SRL_I128, nullptr);
1774 setLibcallName(RTLIB::SRA_I128, nullptr);
1777 // Combine sin / cos into one node or libcall if possible.
1778 if (Subtarget->hasSinCos()) {
1779 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1780 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1781 if (Subtarget->isTargetDarwin()) {
1782 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1783 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1784 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1785 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1789 if (Subtarget->isTargetWin64()) {
1790 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1791 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1792 setOperationAction(ISD::SREM, MVT::i128, Custom);
1793 setOperationAction(ISD::UREM, MVT::i128, Custom);
1794 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1795 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1798 // We have target-specific dag combine patterns for the following nodes:
1799 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1800 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1801 setTargetDAGCombine(ISD::BITCAST);
1802 setTargetDAGCombine(ISD::VSELECT);
1803 setTargetDAGCombine(ISD::SELECT);
1804 setTargetDAGCombine(ISD::SHL);
1805 setTargetDAGCombine(ISD::SRA);
1806 setTargetDAGCombine(ISD::SRL);
1807 setTargetDAGCombine(ISD::OR);
1808 setTargetDAGCombine(ISD::AND);
1809 setTargetDAGCombine(ISD::ADD);
1810 setTargetDAGCombine(ISD::FADD);
1811 setTargetDAGCombine(ISD::FSUB);
1812 setTargetDAGCombine(ISD::FNEG);
1813 setTargetDAGCombine(ISD::FMA);
1814 setTargetDAGCombine(ISD::FMINNUM);
1815 setTargetDAGCombine(ISD::FMAXNUM);
1816 setTargetDAGCombine(ISD::SUB);
1817 setTargetDAGCombine(ISD::LOAD);
1818 setTargetDAGCombine(ISD::MLOAD);
1819 setTargetDAGCombine(ISD::STORE);
1820 setTargetDAGCombine(ISD::MSTORE);
1821 setTargetDAGCombine(ISD::TRUNCATE);
1822 setTargetDAGCombine(ISD::ZERO_EXTEND);
1823 setTargetDAGCombine(ISD::ANY_EXTEND);
1824 setTargetDAGCombine(ISD::SIGN_EXTEND);
1825 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1826 setTargetDAGCombine(ISD::SINT_TO_FP);
1827 setTargetDAGCombine(ISD::UINT_TO_FP);
1828 setTargetDAGCombine(ISD::SETCC);
1829 setTargetDAGCombine(ISD::BUILD_VECTOR);
1830 setTargetDAGCombine(ISD::MUL);
1831 setTargetDAGCombine(ISD::XOR);
1832 setTargetDAGCombine(ISD::MSCATTER);
1833 setTargetDAGCombine(ISD::MGATHER);
1835 computeRegisterProperties(Subtarget->getRegisterInfo());
1837 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1838 MaxStoresPerMemsetOptSize = 8;
1839 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1840 MaxStoresPerMemcpyOptSize = 4;
1841 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1842 MaxStoresPerMemmoveOptSize = 4;
1843 setPrefLoopAlignment(4); // 2^4 bytes.
1845 // A predictable cmov does not hurt on an in-order CPU.
1846 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1847 PredictableSelectIsExpensive = !Subtarget->isAtom();
1848 EnableExtLdPromotion = true;
1849 setPrefFunctionAlignment(4); // 2^4 bytes.
1851 verifyIntrinsicTables();
1854 // This has so far only been implemented for 64-bit MachO.
1855 bool X86TargetLowering::useLoadStackGuardNode() const {
1856 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1859 TargetLoweringBase::LegalizeTypeAction
1860 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1861 if (ExperimentalVectorWideningLegalization &&
1862 VT.getVectorNumElements() != 1 &&
1863 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1864 return TypeWidenVector;
1866 return TargetLoweringBase::getPreferredVectorAction(VT);
1869 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1872 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1874 if (VT.isSimple()) {
1875 MVT VVT = VT.getSimpleVT();
1876 const unsigned NumElts = VVT.getVectorNumElements();
1877 const MVT EltVT = VVT.getVectorElementType();
1878 if (VVT.is512BitVector()) {
1879 if (Subtarget->hasAVX512())
1880 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1881 EltVT == MVT::f32 || EltVT == MVT::f64)
1883 case 8: return MVT::v8i1;
1884 case 16: return MVT::v16i1;
1886 if (Subtarget->hasBWI())
1887 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1889 case 32: return MVT::v32i1;
1890 case 64: return MVT::v64i1;
1894 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1895 if (Subtarget->hasVLX())
1896 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1897 EltVT == MVT::f32 || EltVT == MVT::f64)
1899 case 2: return MVT::v2i1;
1900 case 4: return MVT::v4i1;
1901 case 8: return MVT::v8i1;
1903 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1904 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1906 case 8: return MVT::v8i1;
1907 case 16: return MVT::v16i1;
1908 case 32: return MVT::v32i1;
1913 return VT.changeVectorElementTypeToInteger();
1916 /// Helper for getByValTypeAlignment to determine
1917 /// the desired ByVal argument alignment.
1918 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1921 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1922 if (VTy->getBitWidth() == 128)
1924 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1925 unsigned EltAlign = 0;
1926 getMaxByValAlign(ATy->getElementType(), EltAlign);
1927 if (EltAlign > MaxAlign)
1928 MaxAlign = EltAlign;
1929 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1930 for (auto *EltTy : STy->elements()) {
1931 unsigned EltAlign = 0;
1932 getMaxByValAlign(EltTy, EltAlign);
1933 if (EltAlign > MaxAlign)
1934 MaxAlign = EltAlign;
1941 /// Return the desired alignment for ByVal aggregate
1942 /// function arguments in the caller parameter area. For X86, aggregates
1943 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1944 /// are at 4-byte boundaries.
1945 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1946 const DataLayout &DL) const {
1947 if (Subtarget->is64Bit()) {
1948 // Max of 8 and alignment of type.
1949 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1956 if (Subtarget->hasSSE1())
1957 getMaxByValAlign(Ty, Align);
1961 /// Returns the target specific optimal type for load
1962 /// and store operations as a result of memset, memcpy, and memmove
1963 /// lowering. If DstAlign is zero that means it's safe to destination
1964 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1965 /// means there isn't a need to check it against alignment requirement,
1966 /// probably because the source does not need to be loaded. If 'IsMemset' is
1967 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1968 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1969 /// source is constant so it does not need to be loaded.
1970 /// It returns EVT::Other if the type should be determined using generic
1971 /// target-independent logic.
1973 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1974 unsigned DstAlign, unsigned SrcAlign,
1975 bool IsMemset, bool ZeroMemset,
1977 MachineFunction &MF) const {
1978 const Function *F = MF.getFunction();
1979 if ((!IsMemset || ZeroMemset) &&
1980 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1982 (!Subtarget->isUnalignedMem16Slow() ||
1983 ((DstAlign == 0 || DstAlign >= 16) &&
1984 (SrcAlign == 0 || SrcAlign >= 16)))) {
1986 // FIXME: Check if unaligned 32-byte accesses are slow.
1987 if (Subtarget->hasInt256())
1989 if (Subtarget->hasFp256())
1992 if (Subtarget->hasSSE2())
1994 if (Subtarget->hasSSE1())
1996 } else if (!MemcpyStrSrc && Size >= 8 &&
1997 !Subtarget->is64Bit() &&
1998 Subtarget->hasSSE2()) {
1999 // Do not use f64 to lower memcpy if source is string constant. It's
2000 // better to use i32 to avoid the loads.
2004 // This is a compromise. If we reach here, unaligned accesses may be slow on
2005 // this target. However, creating smaller, aligned accesses could be even
2006 // slower and would certainly be a lot more code.
2007 if (Subtarget->is64Bit() && Size >= 8)
2012 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2014 return X86ScalarSSEf32;
2015 else if (VT == MVT::f64)
2016 return X86ScalarSSEf64;
2021 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2026 switch (VT.getSizeInBits()) {
2028 // 8-byte and under are always assumed to be fast.
2032 *Fast = !Subtarget->isUnalignedMem16Slow();
2035 *Fast = !Subtarget->isUnalignedMem32Slow();
2037 // TODO: What about AVX-512 (512-bit) accesses?
2040 // Misaligned accesses of any size are always allowed.
2044 /// Return the entry encoding for a jump table in the
2045 /// current function. The returned value is a member of the
2046 /// MachineJumpTableInfo::JTEntryKind enum.
2047 unsigned X86TargetLowering::getJumpTableEncoding() const {
2048 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2050 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2051 Subtarget->isPICStyleGOT())
2052 return MachineJumpTableInfo::EK_Custom32;
2054 // Otherwise, use the normal jump table encoding heuristics.
2055 return TargetLowering::getJumpTableEncoding();
2058 bool X86TargetLowering::useSoftFloat() const {
2059 return Subtarget->useSoftFloat();
2063 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2064 const MachineBasicBlock *MBB,
2065 unsigned uid,MCContext &Ctx) const{
2066 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2067 Subtarget->isPICStyleGOT());
2068 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2070 return MCSymbolRefExpr::create(MBB->getSymbol(),
2071 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2074 /// Returns relocation base for the given PIC jumptable.
2075 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2076 SelectionDAG &DAG) const {
2077 if (!Subtarget->is64Bit())
2078 // This doesn't have SDLoc associated with it, but is not really the
2079 // same as a Register.
2080 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2081 getPointerTy(DAG.getDataLayout()));
2085 /// This returns the relocation base for the given PIC jumptable,
2086 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2087 const MCExpr *X86TargetLowering::
2088 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2089 MCContext &Ctx) const {
2090 // X86-64 uses RIP relative addressing based on the jump table label.
2091 if (Subtarget->isPICStyleRIPRel())
2092 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2094 // Otherwise, the reference is relative to the PIC base.
2095 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2098 std::pair<const TargetRegisterClass *, uint8_t>
2099 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2101 const TargetRegisterClass *RRC = nullptr;
2103 switch (VT.SimpleTy) {
2105 return TargetLowering::findRepresentativeClass(TRI, VT);
2106 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2107 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2110 RRC = &X86::VR64RegClass;
2112 case MVT::f32: case MVT::f64:
2113 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2114 case MVT::v4f32: case MVT::v2f64:
2115 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2117 RRC = &X86::VR128RegClass;
2120 return std::make_pair(RRC, Cost);
2123 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2124 unsigned &Offset) const {
2125 if (!Subtarget->isTargetLinux())
2128 if (Subtarget->is64Bit()) {
2129 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2131 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2143 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2144 if (!Subtarget->isTargetAndroid())
2145 return TargetLowering::getSafeStackPointerLocation(IRB);
2147 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2148 // definition of TLS_SLOT_SAFESTACK in
2149 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2150 unsigned AddressSpace, Offset;
2151 if (Subtarget->is64Bit()) {
2152 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2154 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2164 return ConstantExpr::getIntToPtr(
2165 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2166 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2169 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2170 unsigned DestAS) const {
2171 assert(SrcAS != DestAS && "Expected different address spaces!");
2173 return SrcAS < 256 && DestAS < 256;
2176 //===----------------------------------------------------------------------===//
2177 // Return Value Calling Convention Implementation
2178 //===----------------------------------------------------------------------===//
2180 #include "X86GenCallingConv.inc"
2182 bool X86TargetLowering::CanLowerReturn(
2183 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2184 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2185 SmallVector<CCValAssign, 16> RVLocs;
2186 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2187 return CCInfo.CheckReturn(Outs, RetCC_X86);
2190 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2191 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2196 X86TargetLowering::LowerReturn(SDValue Chain,
2197 CallingConv::ID CallConv, bool isVarArg,
2198 const SmallVectorImpl<ISD::OutputArg> &Outs,
2199 const SmallVectorImpl<SDValue> &OutVals,
2200 SDLoc dl, SelectionDAG &DAG) const {
2201 MachineFunction &MF = DAG.getMachineFunction();
2202 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2204 if (CallConv == CallingConv::X86_INTR && !Outs.empty())
2205 report_fatal_error("X86 interrupts may not return any value");
2207 SmallVector<CCValAssign, 16> RVLocs;
2208 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2209 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2212 SmallVector<SDValue, 6> RetOps;
2213 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2214 // Operand #1 = Bytes To Pop
2215 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2218 // Copy the result values into the output registers.
2219 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2220 CCValAssign &VA = RVLocs[i];
2221 assert(VA.isRegLoc() && "Can only return in registers!");
2222 SDValue ValToCopy = OutVals[i];
2223 EVT ValVT = ValToCopy.getValueType();
2225 // Promote values to the appropriate types.
2226 if (VA.getLocInfo() == CCValAssign::SExt)
2227 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2228 else if (VA.getLocInfo() == CCValAssign::ZExt)
2229 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2230 else if (VA.getLocInfo() == CCValAssign::AExt) {
2231 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2232 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2234 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2236 else if (VA.getLocInfo() == CCValAssign::BCvt)
2237 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2239 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2240 "Unexpected FP-extend for return value.");
2242 // If this is x86-64, and we disabled SSE, we can't return FP values,
2243 // or SSE or MMX vectors.
2244 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2245 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2246 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2247 report_fatal_error("SSE register return with SSE disabled");
2249 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2250 // llvm-gcc has never done it right and no one has noticed, so this
2251 // should be OK for now.
2252 if (ValVT == MVT::f64 &&
2253 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2254 report_fatal_error("SSE2 register return with SSE2 disabled");
2256 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2257 // the RET instruction and handled by the FP Stackifier.
2258 if (VA.getLocReg() == X86::FP0 ||
2259 VA.getLocReg() == X86::FP1) {
2260 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2261 // change the value to the FP stack register class.
2262 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2263 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2264 RetOps.push_back(ValToCopy);
2265 // Don't emit a copytoreg.
2269 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2270 // which is returned in RAX / RDX.
2271 if (Subtarget->is64Bit()) {
2272 if (ValVT == MVT::x86mmx) {
2273 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2274 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2275 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2277 // If we don't have SSE2 available, convert to v4f32 so the generated
2278 // register is legal.
2279 if (!Subtarget->hasSSE2())
2280 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2285 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2286 Flag = Chain.getValue(1);
2287 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2290 // All x86 ABIs require that for returning structs by value we copy
2291 // the sret argument into %rax/%eax (depending on ABI) for the return.
2292 // We saved the argument into a virtual register in the entry block,
2293 // so now we copy the value out and into %rax/%eax.
2295 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2296 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2297 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2298 // either case FuncInfo->setSRetReturnReg() will have been called.
2299 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2300 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2301 getPointerTy(MF.getDataLayout()));
2304 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2305 X86::RAX : X86::EAX;
2306 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2307 Flag = Chain.getValue(1);
2309 // RAX/EAX now acts like a return value.
2311 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2314 RetOps[0] = Chain; // Update chain.
2316 // Add the flag if we have it.
2318 RetOps.push_back(Flag);
2320 X86ISD::NodeType opcode = X86ISD::RET_FLAG;
2321 if (CallConv == CallingConv::X86_INTR)
2322 opcode = X86ISD::IRET;
2323 return DAG.getNode(opcode, dl, MVT::Other, RetOps);
2326 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2327 if (N->getNumValues() != 1)
2329 if (!N->hasNUsesOfValue(1, 0))
2332 SDValue TCChain = Chain;
2333 SDNode *Copy = *N->use_begin();
2334 if (Copy->getOpcode() == ISD::CopyToReg) {
2335 // If the copy has a glue operand, we conservatively assume it isn't safe to
2336 // perform a tail call.
2337 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2339 TCChain = Copy->getOperand(0);
2340 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2343 bool HasRet = false;
2344 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2346 if (UI->getOpcode() != X86ISD::RET_FLAG)
2348 // If we are returning more than one value, we can definitely
2349 // not make a tail call see PR19530
2350 if (UI->getNumOperands() > 4)
2352 if (UI->getNumOperands() == 4 &&
2353 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2366 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2367 ISD::NodeType ExtendKind) const {
2369 // TODO: Is this also valid on 32-bit?
2370 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2371 ReturnMVT = MVT::i8;
2373 ReturnMVT = MVT::i32;
2375 EVT MinVT = getRegisterType(Context, ReturnMVT);
2376 return VT.bitsLT(MinVT) ? MinVT : VT;
2379 /// Lower the result values of a call into the
2380 /// appropriate copies out of appropriate physical registers.
2383 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2384 CallingConv::ID CallConv, bool isVarArg,
2385 const SmallVectorImpl<ISD::InputArg> &Ins,
2386 SDLoc dl, SelectionDAG &DAG,
2387 SmallVectorImpl<SDValue> &InVals) const {
2389 // Assign locations to each value returned by this call.
2390 SmallVector<CCValAssign, 16> RVLocs;
2391 bool Is64Bit = Subtarget->is64Bit();
2392 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2394 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2396 // Copy all of the result registers out of their specified physreg.
2397 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2398 CCValAssign &VA = RVLocs[i];
2399 EVT CopyVT = VA.getLocVT();
2401 // If this is x86-64, and we disabled SSE, we can't return FP values
2402 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) &&
2403 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2404 report_fatal_error("SSE register return with SSE disabled");
2407 // If we prefer to use the value in xmm registers, copy it out as f80 and
2408 // use a truncate to move it from fp stack reg to xmm reg.
2409 bool RoundAfterCopy = false;
2410 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2411 isScalarFPTypeInSSEReg(VA.getValVT())) {
2413 RoundAfterCopy = (CopyVT != VA.getLocVT());
2416 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2417 CopyVT, InFlag).getValue(1);
2418 SDValue Val = Chain.getValue(0);
2421 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2422 // This truncation won't change the value.
2423 DAG.getIntPtrConstant(1, dl));
2425 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2426 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2428 InFlag = Chain.getValue(2);
2429 InVals.push_back(Val);
2435 //===----------------------------------------------------------------------===//
2436 // C & StdCall & Fast Calling Convention implementation
2437 //===----------------------------------------------------------------------===//
2438 // StdCall calling convention seems to be standard for many Windows' API
2439 // routines and around. It differs from C calling convention just a little:
2440 // callee should clean up the stack, not caller. Symbols should be also
2441 // decorated in some fancy way :) It doesn't support any vector arguments.
2442 // For info on fast calling convention see Fast Calling Convention (tail call)
2443 // implementation LowerX86_32FastCCCallTo.
2445 /// CallIsStructReturn - Determines whether a call uses struct return
2447 enum StructReturnType {
2452 static StructReturnType
2453 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsMCU) {
2455 return NotStructReturn;
2457 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2458 if (!Flags.isSRet())
2459 return NotStructReturn;
2460 if (Flags.isInReg() || IsMCU)
2461 return RegStructReturn;
2462 return StackStructReturn;
2465 /// Determines whether a function uses struct return semantics.
2466 static StructReturnType
2467 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsMCU) {
2469 return NotStructReturn;
2471 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2472 if (!Flags.isSRet())
2473 return NotStructReturn;
2474 if (Flags.isInReg() || IsMCU)
2475 return RegStructReturn;
2476 return StackStructReturn;
2479 /// Make a copy of an aggregate at address specified by "Src" to address
2480 /// "Dst" with size and alignment information specified by the specific
2481 /// parameter attribute. The copy will be passed as a byval function parameter.
2483 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2484 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2486 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2488 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2489 /*isVolatile*/false, /*AlwaysInline=*/true,
2490 /*isTailCall*/false,
2491 MachinePointerInfo(), MachinePointerInfo());
2494 /// Return true if the calling convention is one that we can guarantee TCO for.
2495 static bool canGuaranteeTCO(CallingConv::ID CC) {
2496 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2497 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2500 /// Return true if we might ever do TCO for calls with this calling convention.
2501 static bool mayTailCallThisCC(CallingConv::ID CC) {
2503 // C calling conventions:
2504 case CallingConv::C:
2505 case CallingConv::X86_64_Win64:
2506 case CallingConv::X86_64_SysV:
2507 // Callee pop conventions:
2508 case CallingConv::X86_ThisCall:
2509 case CallingConv::X86_StdCall:
2510 case CallingConv::X86_VectorCall:
2511 case CallingConv::X86_FastCall:
2514 return canGuaranteeTCO(CC);
2518 /// Return true if the function is being made into a tailcall target by
2519 /// changing its ABI.
2520 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2521 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2524 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2526 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2527 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2531 CallingConv::ID CalleeCC = CS.getCallingConv();
2532 if (!mayTailCallThisCC(CalleeCC))
2539 X86TargetLowering::LowerMemArgument(SDValue Chain,
2540 CallingConv::ID CallConv,
2541 const SmallVectorImpl<ISD::InputArg> &Ins,
2542 SDLoc dl, SelectionDAG &DAG,
2543 const CCValAssign &VA,
2544 MachineFrameInfo *MFI,
2546 // Create the nodes corresponding to a load from this parameter slot.
2547 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2548 bool AlwaysUseMutable = shouldGuaranteeTCO(
2549 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2550 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2553 // If value is passed by pointer we have address passed instead of the value
2555 bool ExtendedInMem = VA.isExtInLoc() &&
2556 VA.getValVT().getScalarType() == MVT::i1;
2558 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2559 ValVT = VA.getLocVT();
2561 ValVT = VA.getValVT();
2563 // Calculate SP offset of interrupt parameter, re-arrange the slot normally
2564 // taken by a return address.
2566 if (CallConv == CallingConv::X86_INTR) {
2567 const X86Subtarget& Subtarget =
2568 static_cast<const X86Subtarget&>(DAG.getSubtarget());
2569 // X86 interrupts may take one or two arguments.
2570 // On the stack there will be no return address as in regular call.
2571 // Offset of last argument need to be set to -4/-8 bytes.
2572 // Where offset of the first argument out of two, should be set to 0 bytes.
2573 Offset = (Subtarget.is64Bit() ? 8 : 4) * ((i + 1) % Ins.size() - 1);
2576 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2577 // changed with more analysis.
2578 // In case of tail call optimization mark all arguments mutable. Since they
2579 // could be overwritten by lowering of arguments in case of a tail call.
2580 if (Flags.isByVal()) {
2581 unsigned Bytes = Flags.getByValSize();
2582 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2583 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2584 // Adjust SP offset of interrupt parameter.
2585 if (CallConv == CallingConv::X86_INTR) {
2586 MFI->setObjectOffset(FI, Offset);
2588 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2590 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2591 VA.getLocMemOffset(), isImmutable);
2592 // Adjust SP offset of interrupt parameter.
2593 if (CallConv == CallingConv::X86_INTR) {
2594 MFI->setObjectOffset(FI, Offset);
2597 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2598 SDValue Val = DAG.getLoad(
2599 ValVT, dl, Chain, FIN,
2600 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2602 return ExtendedInMem ?
2603 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2607 // FIXME: Get this from tablegen.
2608 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2609 const X86Subtarget *Subtarget) {
2610 assert(Subtarget->is64Bit());
2612 if (Subtarget->isCallingConvWin64(CallConv)) {
2613 static const MCPhysReg GPR64ArgRegsWin64[] = {
2614 X86::RCX, X86::RDX, X86::R8, X86::R9
2616 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2619 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2620 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2622 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2625 // FIXME: Get this from tablegen.
2626 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2627 CallingConv::ID CallConv,
2628 const X86Subtarget *Subtarget) {
2629 assert(Subtarget->is64Bit());
2630 if (Subtarget->isCallingConvWin64(CallConv)) {
2631 // The XMM registers which might contain var arg parameters are shadowed
2632 // in their paired GPR. So we only need to save the GPR to their home
2634 // TODO: __vectorcall will change this.
2638 const Function *Fn = MF.getFunction();
2639 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2640 bool isSoftFloat = Subtarget->useSoftFloat();
2641 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2642 "SSE register cannot be used when SSE is disabled!");
2643 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2644 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2648 static const MCPhysReg XMMArgRegs64Bit[] = {
2649 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2650 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2652 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2655 SDValue X86TargetLowering::LowerFormalArguments(
2656 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2657 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2658 SmallVectorImpl<SDValue> &InVals) const {
2659 MachineFunction &MF = DAG.getMachineFunction();
2660 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2661 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2663 const Function* Fn = MF.getFunction();
2664 if (Fn->hasExternalLinkage() &&
2665 Subtarget->isTargetCygMing() &&
2666 Fn->getName() == "main")
2667 FuncInfo->setForceFramePointer(true);
2669 MachineFrameInfo *MFI = MF.getFrameInfo();
2670 bool Is64Bit = Subtarget->is64Bit();
2671 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2673 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2674 "Var args not supported with calling convention fastcc, ghc or hipe");
2676 if (CallConv == CallingConv::X86_INTR) {
2677 bool isLegal = Ins.size() == 1 ||
2678 (Ins.size() == 2 && ((Is64Bit && Ins[1].VT == MVT::i64) ||
2679 (!Is64Bit && Ins[1].VT == MVT::i32)));
2681 report_fatal_error("X86 interrupts may take one or two arguments");
2684 // Assign locations to all of the incoming arguments.
2685 SmallVector<CCValAssign, 16> ArgLocs;
2686 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2688 // Allocate shadow area for Win64
2690 CCInfo.AllocateStack(32, 8);
2692 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2694 unsigned LastVal = ~0U;
2696 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2697 CCValAssign &VA = ArgLocs[i];
2698 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2700 assert(VA.getValNo() != LastVal &&
2701 "Don't support value assigned to multiple locs yet");
2703 LastVal = VA.getValNo();
2705 if (VA.isRegLoc()) {
2706 EVT RegVT = VA.getLocVT();
2707 const TargetRegisterClass *RC;
2708 if (RegVT == MVT::i32)
2709 RC = &X86::GR32RegClass;
2710 else if (Is64Bit && RegVT == MVT::i64)
2711 RC = &X86::GR64RegClass;
2712 else if (RegVT == MVT::f32)
2713 RC = &X86::FR32RegClass;
2714 else if (RegVT == MVT::f64)
2715 RC = &X86::FR64RegClass;
2716 else if (RegVT == MVT::f128)
2717 RC = &X86::FR128RegClass;
2718 else if (RegVT.is512BitVector())
2719 RC = &X86::VR512RegClass;
2720 else if (RegVT.is256BitVector())
2721 RC = &X86::VR256RegClass;
2722 else if (RegVT.is128BitVector())
2723 RC = &X86::VR128RegClass;
2724 else if (RegVT == MVT::x86mmx)
2725 RC = &X86::VR64RegClass;
2726 else if (RegVT == MVT::i1)
2727 RC = &X86::VK1RegClass;
2728 else if (RegVT == MVT::v8i1)
2729 RC = &X86::VK8RegClass;
2730 else if (RegVT == MVT::v16i1)
2731 RC = &X86::VK16RegClass;
2732 else if (RegVT == MVT::v32i1)
2733 RC = &X86::VK32RegClass;
2734 else if (RegVT == MVT::v64i1)
2735 RC = &X86::VK64RegClass;
2737 llvm_unreachable("Unknown argument type!");
2739 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2740 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2742 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2743 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2745 if (VA.getLocInfo() == CCValAssign::SExt)
2746 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2747 DAG.getValueType(VA.getValVT()));
2748 else if (VA.getLocInfo() == CCValAssign::ZExt)
2749 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2750 DAG.getValueType(VA.getValVT()));
2751 else if (VA.getLocInfo() == CCValAssign::BCvt)
2752 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2754 if (VA.isExtInLoc()) {
2755 // Handle MMX values passed in XMM regs.
2756 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2757 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2759 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2762 assert(VA.isMemLoc());
2763 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2766 // If value is passed via pointer - do a load.
2767 if (VA.getLocInfo() == CCValAssign::Indirect)
2768 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2769 MachinePointerInfo(), false, false, false, 0);
2771 InVals.push_back(ArgValue);
2774 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2775 // All x86 ABIs require that for returning structs by value we copy the
2776 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2777 // the argument into a virtual register so that we can access it from the
2779 if (Ins[i].Flags.isSRet()) {
2780 unsigned Reg = FuncInfo->getSRetReturnReg();
2782 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2783 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2784 FuncInfo->setSRetReturnReg(Reg);
2786 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2787 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2792 unsigned StackSize = CCInfo.getNextStackOffset();
2793 // Align stack specially for tail calls.
2794 if (shouldGuaranteeTCO(CallConv,
2795 MF.getTarget().Options.GuaranteedTailCallOpt))
2796 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2798 // If the function takes variable number of arguments, make a frame index for
2799 // the start of the first vararg value... for expansion of llvm.va_start. We
2800 // can skip this if there are no va_start calls.
2801 if (MFI->hasVAStart() &&
2802 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2803 CallConv != CallingConv::X86_ThisCall))) {
2804 FuncInfo->setVarArgsFrameIndex(
2805 MFI->CreateFixedObject(1, StackSize, true));
2808 // Figure out if XMM registers are in use.
2809 assert(!(Subtarget->useSoftFloat() &&
2810 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2811 "SSE register cannot be used when SSE is disabled!");
2813 // 64-bit calling conventions support varargs and register parameters, so we
2814 // have to do extra work to spill them in the prologue.
2815 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2816 // Find the first unallocated argument registers.
2817 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2818 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2819 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2820 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2821 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2822 "SSE register cannot be used when SSE is disabled!");
2824 // Gather all the live in physical registers.
2825 SmallVector<SDValue, 6> LiveGPRs;
2826 SmallVector<SDValue, 8> LiveXMMRegs;
2828 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2829 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2831 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2833 if (!ArgXMMs.empty()) {
2834 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2835 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2836 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2837 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2838 LiveXMMRegs.push_back(
2839 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2844 // Get to the caller-allocated home save location. Add 8 to account
2845 // for the return address.
2846 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2847 FuncInfo->setRegSaveFrameIndex(
2848 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2849 // Fixup to set vararg frame on shadow area (4 x i64).
2851 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2853 // For X86-64, if there are vararg parameters that are passed via
2854 // registers, then we must store them to their spots on the stack so
2855 // they may be loaded by deferencing the result of va_next.
2856 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2857 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2858 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2859 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2862 // Store the integer parameter registers.
2863 SmallVector<SDValue, 8> MemOps;
2864 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2865 getPointerTy(DAG.getDataLayout()));
2866 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2867 for (SDValue Val : LiveGPRs) {
2868 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2869 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2871 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2872 MachinePointerInfo::getFixedStack(
2873 DAG.getMachineFunction(),
2874 FuncInfo->getRegSaveFrameIndex(), Offset),
2876 MemOps.push_back(Store);
2880 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2881 // Now store the XMM (fp + vector) parameter registers.
2882 SmallVector<SDValue, 12> SaveXMMOps;
2883 SaveXMMOps.push_back(Chain);
2884 SaveXMMOps.push_back(ALVal);
2885 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2886 FuncInfo->getRegSaveFrameIndex(), dl));
2887 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2888 FuncInfo->getVarArgsFPOffset(), dl));
2889 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2891 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2892 MVT::Other, SaveXMMOps));
2895 if (!MemOps.empty())
2896 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2899 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2900 // Find the largest legal vector type.
2901 MVT VecVT = MVT::Other;
2902 // FIXME: Only some x86_32 calling conventions support AVX512.
2903 if (Subtarget->hasAVX512() &&
2904 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2905 CallConv == CallingConv::Intel_OCL_BI)))
2906 VecVT = MVT::v16f32;
2907 else if (Subtarget->hasAVX())
2909 else if (Subtarget->hasSSE2())
2912 // We forward some GPRs and some vector types.
2913 SmallVector<MVT, 2> RegParmTypes;
2914 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2915 RegParmTypes.push_back(IntVT);
2916 if (VecVT != MVT::Other)
2917 RegParmTypes.push_back(VecVT);
2919 // Compute the set of forwarded registers. The rest are scratch.
2920 SmallVectorImpl<ForwardedRegister> &Forwards =
2921 FuncInfo->getForwardedMustTailRegParms();
2922 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2924 // Conservatively forward AL on x86_64, since it might be used for varargs.
2925 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2926 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2927 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2930 // Copy all forwards from physical to virtual registers.
2931 for (ForwardedRegister &F : Forwards) {
2932 // FIXME: Can we use a less constrained schedule?
2933 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2934 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2935 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2939 // Some CCs need callee pop.
2940 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2941 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2942 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2943 } else if (CallConv == CallingConv::X86_INTR && Ins.size() == 2) {
2944 // X86 interrupts must pop the error code if present
2945 FuncInfo->setBytesToPopOnReturn(Is64Bit ? 8 : 4);
2947 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2948 // If this is an sret function, the return should pop the hidden pointer.
2949 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2950 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2951 argsAreStructReturn(Ins, Subtarget->isTargetMCU()) == StackStructReturn)
2952 FuncInfo->setBytesToPopOnReturn(4);
2956 // RegSaveFrameIndex is X86-64 only.
2957 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2958 if (CallConv == CallingConv::X86_FastCall ||
2959 CallConv == CallingConv::X86_ThisCall)
2960 // fastcc functions can't have varargs.
2961 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2964 FuncInfo->setArgumentStackSize(StackSize);
2966 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2967 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2968 if (Personality == EHPersonality::CoreCLR) {
2970 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2971 // that we'd prefer this slot be allocated towards the bottom of the frame
2972 // (i.e. near the stack pointer after allocating the frame). Every
2973 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2974 // offset from the bottom of this and each funclet's frame must be the
2975 // same, so the size of funclets' (mostly empty) frames is dictated by
2976 // how far this slot is from the bottom (since they allocate just enough
2977 // space to accomodate holding this slot at the correct offset).
2978 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2979 EHInfo->PSPSymFrameIdx = PSPSymFI;
2987 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2988 SDValue StackPtr, SDValue Arg,
2989 SDLoc dl, SelectionDAG &DAG,
2990 const CCValAssign &VA,
2991 ISD::ArgFlagsTy Flags) const {
2992 unsigned LocMemOffset = VA.getLocMemOffset();
2993 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2994 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2996 if (Flags.isByVal())
2997 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2999 return DAG.getStore(
3000 Chain, dl, Arg, PtrOff,
3001 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
3005 /// Emit a load of return address if tail call
3006 /// optimization is performed and it is required.
3008 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
3009 SDValue &OutRetAddr, SDValue Chain,
3010 bool IsTailCall, bool Is64Bit,
3011 int FPDiff, SDLoc dl) const {
3012 // Adjust the Return address stack slot.
3013 EVT VT = getPointerTy(DAG.getDataLayout());
3014 OutRetAddr = getReturnAddressFrameIndex(DAG);
3016 // Load the "old" Return address.
3017 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
3018 false, false, false, 0);
3019 return SDValue(OutRetAddr.getNode(), 1);
3022 /// Emit a store of the return address if tail call
3023 /// optimization is performed and it is required (FPDiff!=0).
3024 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
3025 SDValue Chain, SDValue RetAddrFrIdx,
3026 EVT PtrVT, unsigned SlotSize,
3027 int FPDiff, SDLoc dl) {
3028 // Store the return address to the appropriate stack slot.
3029 if (!FPDiff) return Chain;
3030 // Calculate the new stack slot for the return address.
3031 int NewReturnAddrFI =
3032 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
3034 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
3035 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
3036 MachinePointerInfo::getFixedStack(
3037 DAG.getMachineFunction(), NewReturnAddrFI),
3042 /// Returns a vector_shuffle mask for an movs{s|d}, movd
3043 /// operation of specified width.
3044 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
3046 unsigned NumElems = VT.getVectorNumElements();
3047 SmallVector<int, 8> Mask;
3048 Mask.push_back(NumElems);
3049 for (unsigned i = 1; i != NumElems; ++i)
3051 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3055 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3056 SmallVectorImpl<SDValue> &InVals) const {
3057 SelectionDAG &DAG = CLI.DAG;
3059 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3060 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3061 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3062 SDValue Chain = CLI.Chain;
3063 SDValue Callee = CLI.Callee;
3064 CallingConv::ID CallConv = CLI.CallConv;
3065 bool &isTailCall = CLI.IsTailCall;
3066 bool isVarArg = CLI.IsVarArg;
3068 MachineFunction &MF = DAG.getMachineFunction();
3069 bool Is64Bit = Subtarget->is64Bit();
3070 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3071 StructReturnType SR = callIsStructReturn(Outs, Subtarget->isTargetMCU());
3072 bool IsSibcall = false;
3073 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3074 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3076 if (CallConv == CallingConv::X86_INTR)
3077 report_fatal_error("X86 interrupts may not be called directly");
3079 if (Attr.getValueAsString() == "true")
3082 if (Subtarget->isPICStyleGOT() &&
3083 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3084 // If we are using a GOT, disable tail calls to external symbols with
3085 // default visibility. Tail calling such a symbol requires using a GOT
3086 // relocation, which forces early binding of the symbol. This breaks code
3087 // that require lazy function symbol resolution. Using musttail or
3088 // GuaranteedTailCallOpt will override this.
3089 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3090 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3091 G->getGlobal()->hasDefaultVisibility()))
3095 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3097 // Force this to be a tail call. The verifier rules are enough to ensure
3098 // that we can lower this successfully without moving the return address
3101 } else if (isTailCall) {
3102 // Check if it's really possible to do a tail call.
3103 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3104 isVarArg, SR != NotStructReturn,
3105 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3106 Outs, OutVals, Ins, DAG);
3108 // Sibcalls are automatically detected tailcalls which do not require
3110 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3117 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3118 "Var args not supported with calling convention fastcc, ghc or hipe");
3120 // Analyze operands of the call, assigning locations to each operand.
3121 SmallVector<CCValAssign, 16> ArgLocs;
3122 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3124 // Allocate shadow area for Win64
3126 CCInfo.AllocateStack(32, 8);
3128 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3130 // Get a count of how many bytes are to be pushed on the stack.
3131 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3133 // This is a sibcall. The memory operands are available in caller's
3134 // own caller's stack.
3136 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3137 canGuaranteeTCO(CallConv))
3138 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3141 if (isTailCall && !IsSibcall && !IsMustTail) {
3142 // Lower arguments at fp - stackoffset + fpdiff.
3143 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3145 FPDiff = NumBytesCallerPushed - NumBytes;
3147 // Set the delta of movement of the returnaddr stackslot.
3148 // But only set if delta is greater than previous delta.
3149 if (FPDiff < X86Info->getTCReturnAddrDelta())
3150 X86Info->setTCReturnAddrDelta(FPDiff);
3153 unsigned NumBytesToPush = NumBytes;
3154 unsigned NumBytesToPop = NumBytes;
3156 // If we have an inalloca argument, all stack space has already been allocated
3157 // for us and be right at the top of the stack. We don't support multiple
3158 // arguments passed in memory when using inalloca.
3159 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3161 if (!ArgLocs.back().isMemLoc())
3162 report_fatal_error("cannot use inalloca attribute on a register "
3164 if (ArgLocs.back().getLocMemOffset() != 0)
3165 report_fatal_error("any parameter with the inalloca attribute must be "
3166 "the only memory argument");
3170 Chain = DAG.getCALLSEQ_START(
3171 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3173 SDValue RetAddrFrIdx;
3174 // Load return address for tail calls.
3175 if (isTailCall && FPDiff)
3176 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3177 Is64Bit, FPDiff, dl);
3179 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3180 SmallVector<SDValue, 8> MemOpChains;
3183 // Walk the register/memloc assignments, inserting copies/loads. In the case
3184 // of tail call optimization arguments are handle later.
3185 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3186 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3187 // Skip inalloca arguments, they have already been written.
3188 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3189 if (Flags.isInAlloca())
3192 CCValAssign &VA = ArgLocs[i];
3193 EVT RegVT = VA.getLocVT();
3194 SDValue Arg = OutVals[i];
3195 bool isByVal = Flags.isByVal();
3197 // Promote the value if needed.
3198 switch (VA.getLocInfo()) {
3199 default: llvm_unreachable("Unknown loc info!");
3200 case CCValAssign::Full: break;
3201 case CCValAssign::SExt:
3202 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3204 case CCValAssign::ZExt:
3205 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3207 case CCValAssign::AExt:
3208 if (Arg.getValueType().isVector() &&
3209 Arg.getValueType().getVectorElementType() == MVT::i1)
3210 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3211 else if (RegVT.is128BitVector()) {
3212 // Special case: passing MMX values in XMM registers.
3213 Arg = DAG.getBitcast(MVT::i64, Arg);
3214 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3215 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3217 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3219 case CCValAssign::BCvt:
3220 Arg = DAG.getBitcast(RegVT, Arg);
3222 case CCValAssign::Indirect: {
3223 // Store the argument.
3224 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3225 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3226 Chain = DAG.getStore(
3227 Chain, dl, Arg, SpillSlot,
3228 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3235 if (VA.isRegLoc()) {
3236 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3237 if (isVarArg && IsWin64) {
3238 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3239 // shadow reg if callee is a varargs function.
3240 unsigned ShadowReg = 0;
3241 switch (VA.getLocReg()) {
3242 case X86::XMM0: ShadowReg = X86::RCX; break;
3243 case X86::XMM1: ShadowReg = X86::RDX; break;
3244 case X86::XMM2: ShadowReg = X86::R8; break;
3245 case X86::XMM3: ShadowReg = X86::R9; break;
3248 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3250 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3251 assert(VA.isMemLoc());
3252 if (!StackPtr.getNode())
3253 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3254 getPointerTy(DAG.getDataLayout()));
3255 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3256 dl, DAG, VA, Flags));
3260 if (!MemOpChains.empty())
3261 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3263 if (Subtarget->isPICStyleGOT()) {
3264 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3267 RegsToPass.push_back(std::make_pair(
3268 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3269 getPointerTy(DAG.getDataLayout()))));
3271 // If we are tail calling and generating PIC/GOT style code load the
3272 // address of the callee into ECX. The value in ecx is used as target of
3273 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3274 // for tail calls on PIC/GOT architectures. Normally we would just put the
3275 // address of GOT into ebx and then call target@PLT. But for tail calls
3276 // ebx would be restored (since ebx is callee saved) before jumping to the
3279 // Note: The actual moving to ECX is done further down.
3280 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3281 if (G && !G->getGlobal()->hasLocalLinkage() &&
3282 G->getGlobal()->hasDefaultVisibility())
3283 Callee = LowerGlobalAddress(Callee, DAG);
3284 else if (isa<ExternalSymbolSDNode>(Callee))
3285 Callee = LowerExternalSymbol(Callee, DAG);
3289 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3290 // From AMD64 ABI document:
3291 // For calls that may call functions that use varargs or stdargs
3292 // (prototype-less calls or calls to functions containing ellipsis (...) in
3293 // the declaration) %al is used as hidden argument to specify the number
3294 // of SSE registers used. The contents of %al do not need to match exactly
3295 // the number of registers, but must be an ubound on the number of SSE
3296 // registers used and is in the range 0 - 8 inclusive.
3298 // Count the number of XMM registers allocated.
3299 static const MCPhysReg XMMArgRegs[] = {
3300 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3301 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3303 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3304 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3305 && "SSE registers cannot be used when SSE is disabled");
3307 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3308 DAG.getConstant(NumXMMRegs, dl,
3312 if (isVarArg && IsMustTail) {
3313 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3314 for (const auto &F : Forwards) {
3315 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3316 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3320 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3321 // don't need this because the eligibility check rejects calls that require
3322 // shuffling arguments passed in memory.
3323 if (!IsSibcall && isTailCall) {
3324 // Force all the incoming stack arguments to be loaded from the stack
3325 // before any new outgoing arguments are stored to the stack, because the
3326 // outgoing stack slots may alias the incoming argument stack slots, and
3327 // the alias isn't otherwise explicit. This is slightly more conservative
3328 // than necessary, because it means that each store effectively depends
3329 // on every argument instead of just those arguments it would clobber.
3330 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3332 SmallVector<SDValue, 8> MemOpChains2;
3335 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3336 CCValAssign &VA = ArgLocs[i];
3339 assert(VA.isMemLoc());
3340 SDValue Arg = OutVals[i];
3341 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3342 // Skip inalloca arguments. They don't require any work.
3343 if (Flags.isInAlloca())
3345 // Create frame index.
3346 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3347 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3348 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3349 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3351 if (Flags.isByVal()) {
3352 // Copy relative to framepointer.
3353 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3354 if (!StackPtr.getNode())
3355 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3356 getPointerTy(DAG.getDataLayout()));
3357 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3360 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3364 // Store relative to framepointer.
3365 MemOpChains2.push_back(DAG.getStore(
3366 ArgChain, dl, Arg, FIN,
3367 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3372 if (!MemOpChains2.empty())
3373 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3375 // Store the return address to the appropriate stack slot.
3376 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3377 getPointerTy(DAG.getDataLayout()),
3378 RegInfo->getSlotSize(), FPDiff, dl);
3381 // Build a sequence of copy-to-reg nodes chained together with token chain
3382 // and flag operands which copy the outgoing args into registers.
3384 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3385 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3386 RegsToPass[i].second, InFlag);
3387 InFlag = Chain.getValue(1);
3390 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3391 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3392 // In the 64-bit large code model, we have to make all calls
3393 // through a register, since the call instruction's 32-bit
3394 // pc-relative offset may not be large enough to hold the whole
3396 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3397 // If the callee is a GlobalAddress node (quite common, every direct call
3398 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3400 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3402 // We should use extra load for direct calls to dllimported functions in
3404 const GlobalValue *GV = G->getGlobal();
3405 if (!GV->hasDLLImportStorageClass()) {
3406 unsigned char OpFlags = 0;
3407 bool ExtraLoad = false;
3408 unsigned WrapperKind = ISD::DELETED_NODE;
3410 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3411 // external symbols most go through the PLT in PIC mode. If the symbol
3412 // has hidden or protected visibility, or if it is static or local, then
3413 // we don't need to use the PLT - we can directly call it.
3414 if (Subtarget->isTargetELF() &&
3415 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3416 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3417 OpFlags = X86II::MO_PLT;
3418 } else if (Subtarget->isPICStyleStubAny() &&
3419 !GV->isStrongDefinitionForLinker() &&
3420 (!Subtarget->getTargetTriple().isMacOSX() ||
3421 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3422 // PC-relative references to external symbols should go through $stub,
3423 // unless we're building with the leopard linker or later, which
3424 // automatically synthesizes these stubs.
3425 OpFlags = X86II::MO_DARWIN_STUB;
3426 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3427 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3428 // If the function is marked as non-lazy, generate an indirect call
3429 // which loads from the GOT directly. This avoids runtime overhead
3430 // at the cost of eager binding (and one extra byte of encoding).
3431 OpFlags = X86II::MO_GOTPCREL;
3432 WrapperKind = X86ISD::WrapperRIP;
3436 Callee = DAG.getTargetGlobalAddress(
3437 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3439 // Add a wrapper if needed.
3440 if (WrapperKind != ISD::DELETED_NODE)
3441 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3442 getPointerTy(DAG.getDataLayout()), Callee);
3443 // Add extra indirection if needed.
3445 Callee = DAG.getLoad(
3446 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3447 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3450 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3451 unsigned char OpFlags = 0;
3453 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3454 // external symbols should go through the PLT.
3455 if (Subtarget->isTargetELF() &&
3456 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3457 OpFlags = X86II::MO_PLT;
3458 } else if (Subtarget->isPICStyleStubAny() &&
3459 (!Subtarget->getTargetTriple().isMacOSX() ||
3460 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3461 // PC-relative references to external symbols should go through $stub,
3462 // unless we're building with the leopard linker or later, which
3463 // automatically synthesizes these stubs.
3464 OpFlags = X86II::MO_DARWIN_STUB;
3467 Callee = DAG.getTargetExternalSymbol(
3468 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3469 } else if (Subtarget->isTarget64BitILP32() &&
3470 Callee->getValueType(0) == MVT::i32) {
3471 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3472 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3475 // Returns a chain & a flag for retval copy to use.
3476 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3477 SmallVector<SDValue, 8> Ops;
3479 if (!IsSibcall && isTailCall) {
3480 Chain = DAG.getCALLSEQ_END(Chain,
3481 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3482 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3483 InFlag = Chain.getValue(1);
3486 Ops.push_back(Chain);
3487 Ops.push_back(Callee);
3490 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3492 // Add argument registers to the end of the list so that they are known live
3494 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3495 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3496 RegsToPass[i].second.getValueType()));
3498 // Add a register mask operand representing the call-preserved registers.
3499 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3500 assert(Mask && "Missing call preserved mask for calling convention");
3502 // If this is an invoke in a 32-bit function using a funclet-based
3503 // personality, assume the function clobbers all registers. If an exception
3504 // is thrown, the runtime will not restore CSRs.
3505 // FIXME: Model this more precisely so that we can register allocate across
3506 // the normal edge and spill and fill across the exceptional edge.
3507 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3508 const Function *CallerFn = MF.getFunction();
3509 EHPersonality Pers =
3510 CallerFn->hasPersonalityFn()
3511 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3512 : EHPersonality::Unknown;
3513 if (isFuncletEHPersonality(Pers))
3514 Mask = RegInfo->getNoPreservedMask();
3517 Ops.push_back(DAG.getRegisterMask(Mask));
3519 if (InFlag.getNode())
3520 Ops.push_back(InFlag);
3524 //// If this is the first return lowered for this function, add the regs
3525 //// to the liveout set for the function.
3526 // This isn't right, although it's probably harmless on x86; liveouts
3527 // should be computed from returns not tail calls. Consider a void
3528 // function making a tail call to a function returning int.
3529 MF.getFrameInfo()->setHasTailCall();
3530 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3533 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3534 InFlag = Chain.getValue(1);
3536 // Create the CALLSEQ_END node.
3537 unsigned NumBytesForCalleeToPop;
3538 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3539 DAG.getTarget().Options.GuaranteedTailCallOpt))
3540 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3541 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3542 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3543 SR == StackStructReturn)
3544 // If this is a call to a struct-return function, the callee
3545 // pops the hidden struct pointer, so we have to push it back.
3546 // This is common for Darwin/X86, Linux & Mingw32 targets.
3547 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3548 NumBytesForCalleeToPop = 4;
3550 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3552 // Returns a flag for retval copy to use.
3554 Chain = DAG.getCALLSEQ_END(Chain,
3555 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3556 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3559 InFlag = Chain.getValue(1);
3562 // Handle result values, copying them out of physregs into vregs that we
3564 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3565 Ins, dl, DAG, InVals);
3568 //===----------------------------------------------------------------------===//
3569 // Fast Calling Convention (tail call) implementation
3570 //===----------------------------------------------------------------------===//
3572 // Like std call, callee cleans arguments, convention except that ECX is
3573 // reserved for storing the tail called function address. Only 2 registers are
3574 // free for argument passing (inreg). Tail call optimization is performed
3576 // * tailcallopt is enabled
3577 // * caller/callee are fastcc
3578 // On X86_64 architecture with GOT-style position independent code only local
3579 // (within module) calls are supported at the moment.
3580 // To keep the stack aligned according to platform abi the function
3581 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3582 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3583 // If a tail called function callee has more arguments than the caller the
3584 // caller needs to make sure that there is room to move the RETADDR to. This is
3585 // achieved by reserving an area the size of the argument delta right after the
3586 // original RETADDR, but before the saved framepointer or the spilled registers
3587 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3599 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3602 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3603 SelectionDAG& DAG) const {
3604 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3605 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3606 unsigned StackAlignment = TFI.getStackAlignment();
3607 uint64_t AlignMask = StackAlignment - 1;
3608 int64_t Offset = StackSize;
3609 unsigned SlotSize = RegInfo->getSlotSize();
3610 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3611 // Number smaller than 12 so just add the difference.
3612 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3614 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3615 Offset = ((~AlignMask) & Offset) + StackAlignment +
3616 (StackAlignment-SlotSize);
3621 /// Return true if the given stack call argument is already available in the
3622 /// same position (relatively) of the caller's incoming argument stack.
3624 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3625 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3626 const X86InstrInfo *TII) {
3627 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3629 if (Arg.getOpcode() == ISD::CopyFromReg) {
3630 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3631 if (!TargetRegisterInfo::isVirtualRegister(VR))
3633 MachineInstr *Def = MRI->getVRegDef(VR);
3636 if (!Flags.isByVal()) {
3637 if (!TII->isLoadFromStackSlot(Def, FI))
3640 unsigned Opcode = Def->getOpcode();
3641 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3642 Opcode == X86::LEA64_32r) &&
3643 Def->getOperand(1).isFI()) {
3644 FI = Def->getOperand(1).getIndex();
3645 Bytes = Flags.getByValSize();
3649 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3650 if (Flags.isByVal())
3651 // ByVal argument is passed in as a pointer but it's now being
3652 // dereferenced. e.g.
3653 // define @foo(%struct.X* %A) {
3654 // tail call @bar(%struct.X* byval %A)
3657 SDValue Ptr = Ld->getBasePtr();
3658 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3661 FI = FINode->getIndex();
3662 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3663 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3664 FI = FINode->getIndex();
3665 Bytes = Flags.getByValSize();
3669 assert(FI != INT_MAX);
3670 if (!MFI->isFixedObjectIndex(FI))
3672 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3675 /// Check whether the call is eligible for tail call optimization. Targets
3676 /// that want to do tail call optimization should implement this function.
3677 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3678 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3679 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3680 const SmallVectorImpl<ISD::OutputArg> &Outs,
3681 const SmallVectorImpl<SDValue> &OutVals,
3682 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3683 if (!mayTailCallThisCC(CalleeCC))
3686 // If -tailcallopt is specified, make fastcc functions tail-callable.
3687 MachineFunction &MF = DAG.getMachineFunction();
3688 const Function *CallerF = MF.getFunction();
3690 // If the function return type is x86_fp80 and the callee return type is not,
3691 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3692 // perform a tailcall optimization here.
3693 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3696 CallingConv::ID CallerCC = CallerF->getCallingConv();
3697 bool CCMatch = CallerCC == CalleeCC;
3698 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3699 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3701 // Win64 functions have extra shadow space for argument homing. Don't do the
3702 // sibcall if the caller and callee have mismatched expectations for this
3704 if (IsCalleeWin64 != IsCallerWin64)
3707 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3708 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3713 // Look for obvious safe cases to perform tail call optimization that do not
3714 // require ABI changes. This is what gcc calls sibcall.
3716 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3717 // emit a special epilogue.
3718 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3719 if (RegInfo->needsStackRealignment(MF))
3722 // Also avoid sibcall optimization if either caller or callee uses struct
3723 // return semantics.
3724 if (isCalleeStructRet || isCallerStructRet)
3727 // Do not sibcall optimize vararg calls unless all arguments are passed via
3729 if (isVarArg && !Outs.empty()) {
3730 // Optimizing for varargs on Win64 is unlikely to be safe without
3731 // additional testing.
3732 if (IsCalleeWin64 || IsCallerWin64)
3735 SmallVector<CCValAssign, 16> ArgLocs;
3736 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3739 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3740 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3741 if (!ArgLocs[i].isRegLoc())
3745 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3746 // stack. Therefore, if it's not used by the call it is not safe to optimize
3747 // this into a sibcall.
3748 bool Unused = false;
3749 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3756 SmallVector<CCValAssign, 16> RVLocs;
3757 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3759 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3760 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3761 CCValAssign &VA = RVLocs[i];
3762 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3767 // If the calling conventions do not match, then we'd better make sure the
3768 // results are returned in the same way as what the caller expects.
3770 SmallVector<CCValAssign, 16> RVLocs1;
3771 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3773 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3775 SmallVector<CCValAssign, 16> RVLocs2;
3776 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3778 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3780 if (RVLocs1.size() != RVLocs2.size())
3782 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3783 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3785 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3787 if (RVLocs1[i].isRegLoc()) {
3788 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3791 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3797 unsigned StackArgsSize = 0;
3799 // If the callee takes no arguments then go on to check the results of the
3801 if (!Outs.empty()) {
3802 // Check if stack adjustment is needed. For now, do not do this if any
3803 // argument is passed on the stack.
3804 SmallVector<CCValAssign, 16> ArgLocs;
3805 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3808 // Allocate shadow area for Win64
3810 CCInfo.AllocateStack(32, 8);
3812 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3813 StackArgsSize = CCInfo.getNextStackOffset();
3815 if (CCInfo.getNextStackOffset()) {
3816 // Check if the arguments are already laid out in the right way as
3817 // the caller's fixed stack objects.
3818 MachineFrameInfo *MFI = MF.getFrameInfo();
3819 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3820 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3821 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3822 CCValAssign &VA = ArgLocs[i];
3823 SDValue Arg = OutVals[i];
3824 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3825 if (VA.getLocInfo() == CCValAssign::Indirect)
3827 if (!VA.isRegLoc()) {
3828 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3835 // If the tailcall address may be in a register, then make sure it's
3836 // possible to register allocate for it. In 32-bit, the call address can
3837 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3838 // callee-saved registers are restored. These happen to be the same
3839 // registers used to pass 'inreg' arguments so watch out for those.
3840 if (!Subtarget->is64Bit() &&
3841 ((!isa<GlobalAddressSDNode>(Callee) &&
3842 !isa<ExternalSymbolSDNode>(Callee)) ||
3843 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3844 unsigned NumInRegs = 0;
3845 // In PIC we need an extra register to formulate the address computation
3847 unsigned MaxInRegs =
3848 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3850 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3851 CCValAssign &VA = ArgLocs[i];
3854 unsigned Reg = VA.getLocReg();
3857 case X86::EAX: case X86::EDX: case X86::ECX:
3858 if (++NumInRegs == MaxInRegs)
3866 bool CalleeWillPop =
3867 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3868 MF.getTarget().Options.GuaranteedTailCallOpt);
3870 if (unsigned BytesToPop =
3871 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3872 // If we have bytes to pop, the callee must pop them.
3873 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3874 if (!CalleePopMatches)
3876 } else if (CalleeWillPop && StackArgsSize > 0) {
3877 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3885 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3886 const TargetLibraryInfo *libInfo) const {
3887 return X86::createFastISel(funcInfo, libInfo);
3890 //===----------------------------------------------------------------------===//
3891 // Other Lowering Hooks
3892 //===----------------------------------------------------------------------===//
3894 static bool MayFoldLoad(SDValue Op) {
3895 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3898 static bool MayFoldIntoStore(SDValue Op) {
3899 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3902 static bool isTargetShuffle(unsigned Opcode) {
3904 default: return false;
3905 case X86ISD::BLENDI:
3906 case X86ISD::PSHUFB:
3907 case X86ISD::PSHUFD:
3908 case X86ISD::PSHUFHW:
3909 case X86ISD::PSHUFLW:
3911 case X86ISD::INSERTPS:
3912 case X86ISD::PALIGNR:
3913 case X86ISD::MOVLHPS:
3914 case X86ISD::MOVLHPD:
3915 case X86ISD::MOVHLPS:
3916 case X86ISD::MOVLPS:
3917 case X86ISD::MOVLPD:
3918 case X86ISD::MOVSHDUP:
3919 case X86ISD::MOVSLDUP:
3920 case X86ISD::MOVDDUP:
3923 case X86ISD::UNPCKL:
3924 case X86ISD::UNPCKH:
3925 case X86ISD::VPERMILPI:
3926 case X86ISD::VPERM2X128:
3927 case X86ISD::VPERMI:
3928 case X86ISD::VPERMV:
3929 case X86ISD::VPERMV3:
3934 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3935 SDValue V1, unsigned TargetMask,
3936 SelectionDAG &DAG) {
3938 default: llvm_unreachable("Unknown x86 shuffle node");
3939 case X86ISD::PSHUFD:
3940 case X86ISD::PSHUFHW:
3941 case X86ISD::PSHUFLW:
3942 case X86ISD::VPERMILPI:
3943 case X86ISD::VPERMI:
3944 return DAG.getNode(Opc, dl, VT, V1,
3945 DAG.getConstant(TargetMask, dl, MVT::i8));
3949 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3950 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3952 default: llvm_unreachable("Unknown x86 shuffle node");
3953 case X86ISD::MOVLHPS:
3954 case X86ISD::MOVLHPD:
3955 case X86ISD::MOVHLPS:
3956 case X86ISD::MOVLPS:
3957 case X86ISD::MOVLPD:
3960 case X86ISD::UNPCKL:
3961 case X86ISD::UNPCKH:
3962 return DAG.getNode(Opc, dl, VT, V1, V2);
3966 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3967 MachineFunction &MF = DAG.getMachineFunction();
3968 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3969 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3970 int ReturnAddrIndex = FuncInfo->getRAIndex();
3972 if (ReturnAddrIndex == 0) {
3973 // Set up a frame object for the return address.
3974 unsigned SlotSize = RegInfo->getSlotSize();
3975 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3978 FuncInfo->setRAIndex(ReturnAddrIndex);
3981 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3984 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3985 bool hasSymbolicDisplacement) {
3986 // Offset should fit into 32 bit immediate field.
3987 if (!isInt<32>(Offset))
3990 // If we don't have a symbolic displacement - we don't have any extra
3992 if (!hasSymbolicDisplacement)
3995 // FIXME: Some tweaks might be needed for medium code model.
3996 if (M != CodeModel::Small && M != CodeModel::Kernel)
3999 // For small code model we assume that latest object is 16MB before end of 31
4000 // bits boundary. We may also accept pretty large negative constants knowing
4001 // that all objects are in the positive half of address space.
4002 if (M == CodeModel::Small && Offset < 16*1024*1024)
4005 // For kernel code model we know that all object resist in the negative half
4006 // of 32bits address space. We may not accept negative offsets, since they may
4007 // be just off and we may accept pretty large positive ones.
4008 if (M == CodeModel::Kernel && Offset >= 0)
4014 /// Determines whether the callee is required to pop its own arguments.
4015 /// Callee pop is necessary to support tail calls.
4016 bool X86::isCalleePop(CallingConv::ID CallingConv,
4017 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
4018 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
4019 // can guarantee TCO.
4020 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
4023 switch (CallingConv) {
4026 case CallingConv::X86_StdCall:
4027 case CallingConv::X86_FastCall:
4028 case CallingConv::X86_ThisCall:
4029 case CallingConv::X86_VectorCall:
4034 /// \brief Return true if the condition is an unsigned comparison operation.
4035 static bool isX86CCUnsigned(unsigned X86CC) {
4037 default: llvm_unreachable("Invalid integer condition!");
4038 case X86::COND_E: return true;
4039 case X86::COND_G: return false;
4040 case X86::COND_GE: return false;
4041 case X86::COND_L: return false;
4042 case X86::COND_LE: return false;
4043 case X86::COND_NE: return true;
4044 case X86::COND_B: return true;
4045 case X86::COND_A: return true;
4046 case X86::COND_BE: return true;
4047 case X86::COND_AE: return true;
4051 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
4052 switch (SetCCOpcode) {
4053 default: llvm_unreachable("Invalid integer condition!");
4054 case ISD::SETEQ: return X86::COND_E;
4055 case ISD::SETGT: return X86::COND_G;
4056 case ISD::SETGE: return X86::COND_GE;
4057 case ISD::SETLT: return X86::COND_L;
4058 case ISD::SETLE: return X86::COND_LE;
4059 case ISD::SETNE: return X86::COND_NE;
4060 case ISD::SETULT: return X86::COND_B;
4061 case ISD::SETUGT: return X86::COND_A;
4062 case ISD::SETULE: return X86::COND_BE;
4063 case ISD::SETUGE: return X86::COND_AE;
4067 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
4068 /// condition code, returning the condition code and the LHS/RHS of the
4069 /// comparison to make.
4070 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4071 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4073 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4074 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4075 // X > -1 -> X == 0, jump !sign.
4076 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4077 return X86::COND_NS;
4079 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4080 // X < 0 -> X == 0, jump on sign.
4083 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4085 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4086 return X86::COND_LE;
4090 return TranslateIntegerX86CC(SetCCOpcode);
4093 // First determine if it is required or is profitable to flip the operands.
4095 // If LHS is a foldable load, but RHS is not, flip the condition.
4096 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4097 !ISD::isNON_EXTLoad(RHS.getNode())) {
4098 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4099 std::swap(LHS, RHS);
4102 switch (SetCCOpcode) {
4108 std::swap(LHS, RHS);
4112 // On a floating point condition, the flags are set as follows:
4114 // 0 | 0 | 0 | X > Y
4115 // 0 | 0 | 1 | X < Y
4116 // 1 | 0 | 0 | X == Y
4117 // 1 | 1 | 1 | unordered
4118 switch (SetCCOpcode) {
4119 default: llvm_unreachable("Condcode should be pre-legalized away");
4121 case ISD::SETEQ: return X86::COND_E;
4122 case ISD::SETOLT: // flipped
4124 case ISD::SETGT: return X86::COND_A;
4125 case ISD::SETOLE: // flipped
4127 case ISD::SETGE: return X86::COND_AE;
4128 case ISD::SETUGT: // flipped
4130 case ISD::SETLT: return X86::COND_B;
4131 case ISD::SETUGE: // flipped
4133 case ISD::SETLE: return X86::COND_BE;
4135 case ISD::SETNE: return X86::COND_NE;
4136 case ISD::SETUO: return X86::COND_P;
4137 case ISD::SETO: return X86::COND_NP;
4139 case ISD::SETUNE: return X86::COND_INVALID;
4143 /// Is there a floating point cmov for the specific X86 condition code?
4144 /// Current x86 isa includes the following FP cmov instructions:
4145 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4146 static bool hasFPCMov(unsigned X86CC) {
4162 /// Returns true if the target can instruction select the
4163 /// specified FP immediate natively. If false, the legalizer will
4164 /// materialize the FP immediate as a load from a constant pool.
4165 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4166 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4167 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4173 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4174 ISD::LoadExtType ExtTy,
4176 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4177 // relocation target a movq or addq instruction: don't let the load shrink.
4178 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4179 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4180 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4181 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4185 /// \brief Returns true if it is beneficial to convert a load of a constant
4186 /// to just the constant itself.
4187 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4189 assert(Ty->isIntegerTy());
4191 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4192 if (BitSize == 0 || BitSize > 64)
4197 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4198 unsigned Index) const {
4199 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4202 return (Index == 0 || Index == ResVT.getVectorNumElements());
4205 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4206 // Speculate cttz only if we can directly use TZCNT.
4207 return Subtarget->hasBMI();
4210 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4211 // Speculate ctlz only if we can directly use LZCNT.
4212 return Subtarget->hasLZCNT();
4215 /// Return true if every element in Mask, beginning
4216 /// from position Pos and ending in Pos+Size is undef.
4217 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4218 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4224 /// Return true if Val is undef or if its value falls within the
4225 /// specified range (L, H].
4226 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4227 return (Val < 0) || (Val >= Low && Val < Hi);
4230 /// Val is either less than zero (undef) or equal to the specified value.
4231 static bool isUndefOrEqual(int Val, int CmpVal) {
4232 return (Val < 0 || Val == CmpVal);
4235 /// Return true if every element in Mask, beginning
4236 /// from position Pos and ending in Pos+Size, falls within the specified
4237 /// sequential range (Low, Low+Size]. or is undef.
4238 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4239 unsigned Pos, unsigned Size, int Low) {
4240 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4241 if (!isUndefOrEqual(Mask[i], Low))
4246 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4247 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4248 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4249 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4250 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4253 // The index should be aligned on a vecWidth-bit boundary.
4255 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4257 MVT VT = N->getSimpleValueType(0);
4258 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4259 bool Result = (Index * ElSize) % vecWidth == 0;
4264 /// Return true if the specified INSERT_SUBVECTOR
4265 /// operand specifies a subvector insert that is suitable for input to
4266 /// insertion of 128 or 256-bit subvectors
4267 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4268 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4269 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4271 // The index should be aligned on a vecWidth-bit boundary.
4273 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4275 MVT VT = N->getSimpleValueType(0);
4276 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4277 bool Result = (Index * ElSize) % vecWidth == 0;
4282 bool X86::isVINSERT128Index(SDNode *N) {
4283 return isVINSERTIndex(N, 128);
4286 bool X86::isVINSERT256Index(SDNode *N) {
4287 return isVINSERTIndex(N, 256);
4290 bool X86::isVEXTRACT128Index(SDNode *N) {
4291 return isVEXTRACTIndex(N, 128);
4294 bool X86::isVEXTRACT256Index(SDNode *N) {
4295 return isVEXTRACTIndex(N, 256);
4298 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4299 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4300 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4301 "Illegal extract subvector for VEXTRACT");
4304 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4306 MVT VecVT = N->getOperand(0).getSimpleValueType();
4307 MVT ElVT = VecVT.getVectorElementType();
4309 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4310 return Index / NumElemsPerChunk;
4313 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4314 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4315 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4316 "Illegal insert subvector for VINSERT");
4319 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4321 MVT VecVT = N->getSimpleValueType(0);
4322 MVT ElVT = VecVT.getVectorElementType();
4324 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4325 return Index / NumElemsPerChunk;
4328 /// Return the appropriate immediate to extract the specified
4329 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4330 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4331 return getExtractVEXTRACTImmediate(N, 128);
4334 /// Return the appropriate immediate to extract the specified
4335 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4336 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4337 return getExtractVEXTRACTImmediate(N, 256);
4340 /// Return the appropriate immediate to insert at the specified
4341 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4342 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4343 return getInsertVINSERTImmediate(N, 128);
4346 /// Return the appropriate immediate to insert at the specified
4347 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4348 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4349 return getInsertVINSERTImmediate(N, 256);
4352 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4353 bool X86::isZeroNode(SDValue Elt) {
4354 return isNullConstant(Elt) || isNullFPConstant(Elt);
4357 // Build a vector of constants
4358 // Use an UNDEF node if MaskElt == -1.
4359 // Spilt 64-bit constants in the 32-bit mode.
4360 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4362 SDLoc dl, bool IsMask = false) {
4364 SmallVector<SDValue, 32> Ops;
4367 MVT ConstVecVT = VT;
4368 unsigned NumElts = VT.getVectorNumElements();
4369 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4370 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4371 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4375 MVT EltVT = ConstVecVT.getVectorElementType();
4376 for (unsigned i = 0; i < NumElts; ++i) {
4377 bool IsUndef = Values[i] < 0 && IsMask;
4378 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4379 DAG.getConstant(Values[i], dl, EltVT);
4380 Ops.push_back(OpNode);
4382 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4383 DAG.getConstant(0, dl, EltVT));
4385 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4387 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4391 /// Returns a vector of specified type with all zero elements.
4392 static SDValue getZeroVector(MVT VT, const X86Subtarget *Subtarget,
4393 SelectionDAG &DAG, SDLoc dl) {
4394 assert(VT.isVector() && "Expected a vector type");
4396 // Always build SSE zero vectors as <4 x i32> bitcasted
4397 // to their dest type. This ensures they get CSE'd.
4399 if (VT.is128BitVector()) { // SSE
4400 if (Subtarget->hasSSE2()) { // SSE2
4401 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4402 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4404 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4405 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4407 } else if (VT.is256BitVector()) { // AVX
4408 if (Subtarget->hasInt256()) { // AVX2
4409 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4410 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4411 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4413 // 256-bit logic and arithmetic instructions in AVX are all
4414 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4415 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4416 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4417 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4419 } else if (VT.is512BitVector()) { // AVX-512
4420 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4421 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4422 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4423 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4424 } else if (VT.getVectorElementType() == MVT::i1) {
4426 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4427 && "Unexpected vector type");
4428 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4429 && "Unexpected vector type");
4430 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4431 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4432 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4434 llvm_unreachable("Unexpected vector type");
4436 return DAG.getBitcast(VT, Vec);
4439 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4440 SelectionDAG &DAG, SDLoc dl,
4441 unsigned vectorWidth) {
4442 assert((vectorWidth == 128 || vectorWidth == 256) &&
4443 "Unsupported vector width");
4444 EVT VT = Vec.getValueType();
4445 EVT ElVT = VT.getVectorElementType();
4446 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4447 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4448 VT.getVectorNumElements()/Factor);
4450 // Extract from UNDEF is UNDEF.
4451 if (Vec.getOpcode() == ISD::UNDEF)
4452 return DAG.getUNDEF(ResultVT);
4454 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4455 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4456 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4458 // This is the index of the first element of the vectorWidth-bit chunk
4459 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4460 IdxVal &= ~(ElemsPerChunk - 1);
4462 // If the input is a buildvector just emit a smaller one.
4463 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4464 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4465 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4467 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4468 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4471 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4472 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4473 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4474 /// instructions or a simple subregister reference. Idx is an index in the
4475 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4476 /// lowering EXTRACT_VECTOR_ELT operations easier.
4477 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4478 SelectionDAG &DAG, SDLoc dl) {
4479 assert((Vec.getValueType().is256BitVector() ||
4480 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4481 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4484 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4485 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4486 SelectionDAG &DAG, SDLoc dl) {
4487 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4488 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4491 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4492 unsigned IdxVal, SelectionDAG &DAG,
4493 SDLoc dl, unsigned vectorWidth) {
4494 assert((vectorWidth == 128 || vectorWidth == 256) &&
4495 "Unsupported vector width");
4496 // Inserting UNDEF is Result
4497 if (Vec.getOpcode() == ISD::UNDEF)
4499 EVT VT = Vec.getValueType();
4500 EVT ElVT = VT.getVectorElementType();
4501 EVT ResultVT = Result.getValueType();
4503 // Insert the relevant vectorWidth bits.
4504 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4505 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4507 // This is the index of the first element of the vectorWidth-bit chunk
4508 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4509 IdxVal &= ~(ElemsPerChunk - 1);
4511 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4512 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4515 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4516 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4517 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4518 /// simple superregister reference. Idx is an index in the 128 bits
4519 /// we want. It need not be aligned to a 128-bit boundary. That makes
4520 /// lowering INSERT_VECTOR_ELT operations easier.
4521 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4522 SelectionDAG &DAG, SDLoc dl) {
4523 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4525 // For insertion into the zero index (low half) of a 256-bit vector, it is
4526 // more efficient to generate a blend with immediate instead of an insert*128.
4527 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4528 // extend the subvector to the size of the result vector. Make sure that
4529 // we are not recursing on that node by checking for undef here.
4530 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4531 Result.getOpcode() != ISD::UNDEF) {
4532 EVT ResultVT = Result.getValueType();
4533 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4534 SDValue Undef = DAG.getUNDEF(ResultVT);
4535 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4538 // The blend instruction, and therefore its mask, depend on the data type.
4539 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4540 if (ScalarType.isFloatingPoint()) {
4541 // Choose either vblendps (float) or vblendpd (double).
4542 unsigned ScalarSize = ScalarType.getSizeInBits();
4543 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4544 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4545 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4546 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4549 const X86Subtarget &Subtarget =
4550 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4552 // AVX2 is needed for 256-bit integer blend support.
4553 // Integers must be cast to 32-bit because there is only vpblendd;
4554 // vpblendw can't be used for this because it has a handicapped mask.
4556 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4557 // is still more efficient than using the wrong domain vinsertf128 that
4558 // will be created by InsertSubVector().
4559 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4561 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4562 Result = DAG.getBitcast(CastVT, Result);
4563 Vec256 = DAG.getBitcast(CastVT, Vec256);
4564 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4565 return DAG.getBitcast(ResultVT, Vec256);
4568 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4571 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4572 SelectionDAG &DAG, SDLoc dl) {
4573 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4574 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4577 /// Insert i1-subvector to i1-vector.
4578 static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
4581 SDValue Vec = Op.getOperand(0);
4582 SDValue SubVec = Op.getOperand(1);
4583 SDValue Idx = Op.getOperand(2);
4585 if (!isa<ConstantSDNode>(Idx))
4588 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
4589 if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
4592 MVT OpVT = Op.getSimpleValueType();
4593 MVT SubVecVT = SubVec.getSimpleValueType();
4594 unsigned NumElems = OpVT.getVectorNumElements();
4595 unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
4597 assert(IdxVal + SubVecNumElems <= NumElems &&
4598 IdxVal % SubVecVT.getSizeInBits() == 0 &&
4599 "Unexpected index value in INSERT_SUBVECTOR");
4601 // There are 3 possible cases:
4602 // 1. Subvector should be inserted in the lower part (IdxVal == 0)
4603 // 2. Subvector should be inserted in the upper part
4604 // (IdxVal + SubVecNumElems == NumElems)
4605 // 3. Subvector should be inserted in the middle (for example v2i1
4606 // to v16i1, index 2)
4608 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
4609 SDValue Undef = DAG.getUNDEF(OpVT);
4610 SDValue WideSubVec =
4611 DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
4613 return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4614 DAG.getConstant(IdxVal, dl, MVT::i8));
4616 if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
4617 unsigned ShiftLeft = NumElems - SubVecNumElems;
4618 unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
4619 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4620 DAG.getConstant(ShiftLeft, dl, MVT::i8));
4621 return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
4622 DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
4626 // Zero lower bits of the Vec
4627 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4628 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4629 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4630 // Merge them together
4631 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4634 // Simple case when we put subvector in the upper part
4635 if (IdxVal + SubVecNumElems == NumElems) {
4636 // Zero upper bits of the Vec
4637 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
4638 DAG.getConstant(IdxVal, dl, MVT::i8));
4639 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4640 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4641 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4642 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4644 // Subvector should be inserted in the middle - use shuffle
4645 SmallVector<int, 64> Mask;
4646 for (unsigned i = 0; i < NumElems; ++i)
4647 Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
4649 return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
4652 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4653 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4654 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4655 /// large BUILD_VECTORS.
4656 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4657 unsigned NumElems, SelectionDAG &DAG,
4659 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4660 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4663 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4664 unsigned NumElems, SelectionDAG &DAG,
4666 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4667 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4670 /// Returns a vector of specified type with all bits set.
4671 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4672 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4673 /// Then bitcast to their original type, ensuring they get CSE'd.
4674 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4675 SelectionDAG &DAG, SDLoc dl) {
4676 assert(VT.isVector() && "Expected a vector type");
4678 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4680 if (VT.is512BitVector()) {
4681 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4682 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4683 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4684 } else if (VT.is256BitVector()) {
4685 if (Subtarget->hasInt256()) { // AVX2
4686 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4687 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4689 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4690 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4692 } else if (VT.is128BitVector()) {
4693 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4695 llvm_unreachable("Unexpected vector type");
4697 return DAG.getBitcast(VT, Vec);
4700 /// Returns a vector_shuffle node for an unpackl operation.
4701 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4703 unsigned NumElems = VT.getVectorNumElements();
4704 SmallVector<int, 8> Mask;
4705 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4707 Mask.push_back(i + NumElems);
4709 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4712 /// Returns a vector_shuffle node for an unpackh operation.
4713 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4715 unsigned NumElems = VT.getVectorNumElements();
4716 SmallVector<int, 8> Mask;
4717 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4718 Mask.push_back(i + Half);
4719 Mask.push_back(i + NumElems + Half);
4721 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4724 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4725 /// This produces a shuffle where the low element of V2 is swizzled into the
4726 /// zero/undef vector, landing at element Idx.
4727 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4728 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4730 const X86Subtarget *Subtarget,
4731 SelectionDAG &DAG) {
4732 MVT VT = V2.getSimpleValueType();
4734 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4735 unsigned NumElems = VT.getVectorNumElements();
4736 SmallVector<int, 16> MaskVec;
4737 for (unsigned i = 0; i != NumElems; ++i)
4738 // If this is the insertion idx, put the low elt of V2 here.
4739 MaskVec.push_back(i == Idx ? NumElems : i);
4740 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4743 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4744 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4745 /// uses one source. Note that this will set IsUnary for shuffles which use a
4746 /// single input multiple times, and in those cases it will
4747 /// adjust the mask to only have indices within that single input.
4748 static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
4749 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4750 unsigned NumElems = VT.getVectorNumElements();
4754 bool IsFakeUnary = false;
4755 switch(N->getOpcode()) {
4756 case X86ISD::BLENDI:
4757 ImmN = N->getOperand(N->getNumOperands()-1);
4758 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4761 ImmN = N->getOperand(N->getNumOperands()-1);
4762 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4763 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4765 case X86ISD::INSERTPS:
4766 ImmN = N->getOperand(N->getNumOperands()-1);
4767 DecodeINSERTPSMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4768 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4770 case X86ISD::UNPCKH:
4771 DecodeUNPCKHMask(VT, Mask);
4772 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4774 case X86ISD::UNPCKL:
4775 DecodeUNPCKLMask(VT, Mask);
4776 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4778 case X86ISD::MOVHLPS:
4779 DecodeMOVHLPSMask(NumElems, Mask);
4780 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4782 case X86ISD::MOVLHPS:
4783 DecodeMOVLHPSMask(NumElems, Mask);
4784 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4786 case X86ISD::PALIGNR:
4787 ImmN = N->getOperand(N->getNumOperands()-1);
4788 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4790 case X86ISD::PSHUFD:
4791 case X86ISD::VPERMILPI:
4792 ImmN = N->getOperand(N->getNumOperands()-1);
4793 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4796 case X86ISD::PSHUFHW:
4797 ImmN = N->getOperand(N->getNumOperands()-1);
4798 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4801 case X86ISD::PSHUFLW:
4802 ImmN = N->getOperand(N->getNumOperands()-1);
4803 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4806 case X86ISD::PSHUFB: {
4808 SDValue MaskNode = N->getOperand(1);
4809 while (MaskNode->getOpcode() == ISD::BITCAST)
4810 MaskNode = MaskNode->getOperand(0);
4812 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4813 // If we have a build-vector, then things are easy.
4814 MVT VT = MaskNode.getSimpleValueType();
4815 assert(VT.isVector() &&
4816 "Can't produce a non-vector with a build_vector!");
4817 if (!VT.isInteger())
4820 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4822 SmallVector<uint64_t, 32> RawMask;
4823 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4824 SDValue Op = MaskNode->getOperand(i);
4825 if (Op->getOpcode() == ISD::UNDEF) {
4826 RawMask.push_back((uint64_t)SM_SentinelUndef);
4829 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4832 APInt MaskElement = CN->getAPIntValue();
4834 // We now have to decode the element which could be any integer size and
4835 // extract each byte of it.
4836 for (int j = 0; j < NumBytesPerElement; ++j) {
4837 // Note that this is x86 and so always little endian: the low byte is
4838 // the first byte of the mask.
4839 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4840 MaskElement = MaskElement.lshr(8);
4843 DecodePSHUFBMask(RawMask, Mask);
4847 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4851 SDValue Ptr = MaskLoad->getBasePtr();
4852 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4853 Ptr->getOpcode() == X86ISD::WrapperRIP)
4854 Ptr = Ptr->getOperand(0);
4856 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4857 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4860 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4861 DecodePSHUFBMask(C, Mask);
4867 case X86ISD::VPERMI:
4868 ImmN = N->getOperand(N->getNumOperands()-1);
4869 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4874 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4876 case X86ISD::VPERM2X128:
4877 ImmN = N->getOperand(N->getNumOperands()-1);
4878 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4879 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4881 case X86ISD::MOVSLDUP:
4882 DecodeMOVSLDUPMask(VT, Mask);
4885 case X86ISD::MOVSHDUP:
4886 DecodeMOVSHDUPMask(VT, Mask);
4889 case X86ISD::MOVDDUP:
4890 DecodeMOVDDUPMask(VT, Mask);
4893 case X86ISD::MOVLHPD:
4894 case X86ISD::MOVLPD:
4895 case X86ISD::MOVLPS:
4896 // Not yet implemented
4898 case X86ISD::VPERMV: {
4900 SDValue MaskNode = N->getOperand(0);
4901 while (MaskNode->getOpcode() == ISD::BITCAST)
4902 MaskNode = MaskNode->getOperand(0);
4904 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4905 SmallVector<uint64_t, 32> RawMask;
4906 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4907 // If we have a build-vector, then things are easy.
4908 assert(MaskNode.getSimpleValueType().isInteger() &&
4909 MaskNode.getSimpleValueType().getVectorNumElements() ==
4910 VT.getVectorNumElements());
4912 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4913 SDValue Op = MaskNode->getOperand(i);
4914 if (Op->getOpcode() == ISD::UNDEF)
4915 RawMask.push_back((uint64_t)SM_SentinelUndef);
4916 else if (isa<ConstantSDNode>(Op)) {
4917 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4918 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4922 DecodeVPERMVMask(RawMask, Mask);
4925 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4926 unsigned NumEltsInMask = MaskNode->getNumOperands();
4927 MaskNode = MaskNode->getOperand(0);
4928 if (auto *CN = dyn_cast<ConstantSDNode>(MaskNode)) {
4929 APInt MaskEltValue = CN->getAPIntValue();
4930 for (unsigned i = 0; i < NumEltsInMask; ++i)
4931 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4932 DecodeVPERMVMask(RawMask, Mask);
4935 // It may be a scalar load
4938 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4942 SDValue Ptr = MaskLoad->getBasePtr();
4943 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4944 Ptr->getOpcode() == X86ISD::WrapperRIP)
4945 Ptr = Ptr->getOperand(0);
4947 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4948 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4951 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4952 DecodeVPERMVMask(C, VT, Mask);
4957 case X86ISD::VPERMV3: {
4959 SDValue MaskNode = N->getOperand(1);
4960 while (MaskNode->getOpcode() == ISD::BITCAST)
4961 MaskNode = MaskNode->getOperand(1);
4963 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4964 // If we have a build-vector, then things are easy.
4965 assert(MaskNode.getSimpleValueType().isInteger() &&
4966 MaskNode.getSimpleValueType().getVectorNumElements() ==
4967 VT.getVectorNumElements());
4969 SmallVector<uint64_t, 32> RawMask;
4970 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4972 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4973 SDValue Op = MaskNode->getOperand(i);
4974 if (Op->getOpcode() == ISD::UNDEF)
4975 RawMask.push_back((uint64_t)SM_SentinelUndef);
4977 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4980 APInt MaskElement = CN->getAPIntValue();
4981 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4984 DecodeVPERMV3Mask(RawMask, Mask);
4988 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4992 SDValue Ptr = MaskLoad->getBasePtr();
4993 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4994 Ptr->getOpcode() == X86ISD::WrapperRIP)
4995 Ptr = Ptr->getOperand(0);
4997 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4998 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5001 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5002 DecodeVPERMV3Mask(C, VT, Mask);
5007 default: llvm_unreachable("unknown target shuffle node");
5010 // Empty mask indicates the decode failed.
5014 // Check if we're getting a shuffle mask with zero'd elements.
5015 if (!AllowSentinelZero)
5016 if (std::any_of(Mask.begin(), Mask.end(),
5017 [](int M){ return M == SM_SentinelZero; }))
5020 // If we have a fake unary shuffle, the shuffle mask is spread across two
5021 // inputs that are actually the same node. Re-map the mask to always point
5022 // into the first input.
5025 if (M >= (int)Mask.size())
5031 /// Returns the scalar element that will make up the ith
5032 /// element of the result of the vector shuffle.
5033 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5036 return SDValue(); // Limit search depth.
5038 SDValue V = SDValue(N, 0);
5039 EVT VT = V.getValueType();
5040 unsigned Opcode = V.getOpcode();
5042 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5043 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5044 int Elt = SV->getMaskElt(Index);
5047 return DAG.getUNDEF(VT.getVectorElementType());
5049 unsigned NumElems = VT.getVectorNumElements();
5050 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5051 : SV->getOperand(1);
5052 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5055 // Recurse into target specific vector shuffles to find scalars.
5056 if (isTargetShuffle(Opcode)) {
5057 MVT ShufVT = V.getSimpleValueType();
5058 int NumElems = (int)ShufVT.getVectorNumElements();
5059 SmallVector<int, 16> ShuffleMask;
5062 if (!getTargetShuffleMask(N, ShufVT, false, ShuffleMask, IsUnary))
5065 int Elt = ShuffleMask[Index];
5066 if (Elt == SM_SentinelUndef)
5067 return DAG.getUNDEF(ShufVT.getVectorElementType());
5069 assert(0 <= Elt && Elt < (2*NumElems) && "Shuffle index out of range");
5070 SDValue NewV = (Elt < NumElems) ? N->getOperand(0) : N->getOperand(1);
5071 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5075 // Actual nodes that may contain scalar elements
5076 if (Opcode == ISD::BITCAST) {
5077 V = V.getOperand(0);
5078 EVT SrcVT = V.getValueType();
5079 unsigned NumElems = VT.getVectorNumElements();
5081 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5085 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5086 return (Index == 0) ? V.getOperand(0)
5087 : DAG.getUNDEF(VT.getVectorElementType());
5089 if (V.getOpcode() == ISD::BUILD_VECTOR)
5090 return V.getOperand(Index);
5095 /// Custom lower build_vector of v16i8.
5096 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5097 unsigned NumNonZero, unsigned NumZero,
5099 const X86Subtarget* Subtarget,
5100 const TargetLowering &TLI) {
5108 // SSE4.1 - use PINSRB to insert each byte directly.
5109 if (Subtarget->hasSSE41()) {
5110 for (unsigned i = 0; i < 16; ++i) {
5111 bool isNonZero = (NonZeros & (1 << i)) != 0;
5115 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
5117 V = DAG.getUNDEF(MVT::v16i8);
5120 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5121 MVT::v16i8, V, Op.getOperand(i),
5122 DAG.getIntPtrConstant(i, dl));
5129 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5130 for (unsigned i = 0; i < 16; ++i) {
5131 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5132 if (ThisIsNonZero && First) {
5134 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5136 V = DAG.getUNDEF(MVT::v8i16);
5141 SDValue ThisElt, LastElt;
5142 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5143 if (LastIsNonZero) {
5144 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5145 MVT::i16, Op.getOperand(i-1));
5147 if (ThisIsNonZero) {
5148 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5149 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5150 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5152 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5156 if (ThisElt.getNode())
5157 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5158 DAG.getIntPtrConstant(i/2, dl));
5162 return DAG.getBitcast(MVT::v16i8, V);
5165 /// Custom lower build_vector of v8i16.
5166 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5167 unsigned NumNonZero, unsigned NumZero,
5169 const X86Subtarget* Subtarget,
5170 const TargetLowering &TLI) {
5177 for (unsigned i = 0; i < 8; ++i) {
5178 bool isNonZero = (NonZeros & (1 << i)) != 0;
5182 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5184 V = DAG.getUNDEF(MVT::v8i16);
5187 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5188 MVT::v8i16, V, Op.getOperand(i),
5189 DAG.getIntPtrConstant(i, dl));
5196 /// Custom lower build_vector of v4i32 or v4f32.
5197 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5198 const X86Subtarget *Subtarget,
5199 const TargetLowering &TLI) {
5200 // Find all zeroable elements.
5201 std::bitset<4> Zeroable;
5202 for (int i=0; i < 4; ++i) {
5203 SDValue Elt = Op->getOperand(i);
5204 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5206 assert(Zeroable.size() - Zeroable.count() > 1 &&
5207 "We expect at least two non-zero elements!");
5209 // We only know how to deal with build_vector nodes where elements are either
5210 // zeroable or extract_vector_elt with constant index.
5211 SDValue FirstNonZero;
5212 unsigned FirstNonZeroIdx;
5213 for (unsigned i=0; i < 4; ++i) {
5216 SDValue Elt = Op->getOperand(i);
5217 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5218 !isa<ConstantSDNode>(Elt.getOperand(1)))
5220 // Make sure that this node is extracting from a 128-bit vector.
5221 MVT VT = Elt.getOperand(0).getSimpleValueType();
5222 if (!VT.is128BitVector())
5224 if (!FirstNonZero.getNode()) {
5226 FirstNonZeroIdx = i;
5230 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5231 SDValue V1 = FirstNonZero.getOperand(0);
5232 MVT VT = V1.getSimpleValueType();
5234 // See if this build_vector can be lowered as a blend with zero.
5236 unsigned EltMaskIdx, EltIdx;
5238 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5239 if (Zeroable[EltIdx]) {
5240 // The zero vector will be on the right hand side.
5241 Mask[EltIdx] = EltIdx+4;
5245 Elt = Op->getOperand(EltIdx);
5246 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5247 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5248 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5250 Mask[EltIdx] = EltIdx;
5254 // Let the shuffle legalizer deal with blend operations.
5255 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5256 if (V1.getSimpleValueType() != VT)
5257 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5258 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5261 // See if we can lower this build_vector to a INSERTPS.
5262 if (!Subtarget->hasSSE41())
5265 SDValue V2 = Elt.getOperand(0);
5266 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5269 bool CanFold = true;
5270 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5274 SDValue Current = Op->getOperand(i);
5275 SDValue SrcVector = Current->getOperand(0);
5278 CanFold = SrcVector == V1 &&
5279 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5285 assert(V1.getNode() && "Expected at least two non-zero elements!");
5286 if (V1.getSimpleValueType() != MVT::v4f32)
5287 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5288 if (V2.getSimpleValueType() != MVT::v4f32)
5289 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5291 // Ok, we can emit an INSERTPS instruction.
5292 unsigned ZMask = Zeroable.to_ulong();
5294 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5295 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5297 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5298 DAG.getIntPtrConstant(InsertPSMask, DL));
5299 return DAG.getBitcast(VT, Result);
5302 /// Return a vector logical shift node.
5303 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5304 unsigned NumBits, SelectionDAG &DAG,
5305 const TargetLowering &TLI, SDLoc dl) {
5306 assert(VT.is128BitVector() && "Unknown type for VShift");
5307 MVT ShVT = MVT::v2i64;
5308 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5309 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5310 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5311 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5312 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5313 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5317 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5319 // Check if the scalar load can be widened into a vector load. And if
5320 // the address is "base + cst" see if the cst can be "absorbed" into
5321 // the shuffle mask.
5322 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5323 SDValue Ptr = LD->getBasePtr();
5324 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5326 EVT PVT = LD->getValueType(0);
5327 if (PVT != MVT::i32 && PVT != MVT::f32)
5332 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5333 FI = FINode->getIndex();
5335 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5336 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5337 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5338 Offset = Ptr.getConstantOperandVal(1);
5339 Ptr = Ptr.getOperand(0);
5344 // FIXME: 256-bit vector instructions don't require a strict alignment,
5345 // improve this code to support it better.
5346 unsigned RequiredAlign = VT.getSizeInBits()/8;
5347 SDValue Chain = LD->getChain();
5348 // Make sure the stack object alignment is at least 16 or 32.
5349 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5350 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5351 if (MFI->isFixedObjectIndex(FI)) {
5352 // Can't change the alignment. FIXME: It's possible to compute
5353 // the exact stack offset and reference FI + adjust offset instead.
5354 // If someone *really* cares about this. That's the way to implement it.
5357 MFI->setObjectAlignment(FI, RequiredAlign);
5361 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5362 // Ptr + (Offset & ~15).
5365 if ((Offset % RequiredAlign) & 3)
5367 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5370 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5371 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5374 int EltNo = (Offset - StartOffset) >> 2;
5375 unsigned NumElems = VT.getVectorNumElements();
5377 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5378 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5379 LD->getPointerInfo().getWithOffset(StartOffset),
5380 false, false, false, 0);
5382 SmallVector<int, 8> Mask(NumElems, EltNo);
5384 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5390 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5391 /// elements can be replaced by a single large load which has the same value as
5392 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5394 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5396 /// FIXME: we'd also like to handle the case where the last elements are zero
5397 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5398 /// There's even a handy isZeroNode for that purpose.
5399 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5400 SDLoc &DL, SelectionDAG &DAG,
5401 bool isAfterLegalize) {
5402 unsigned NumElems = Elts.size();
5404 LoadSDNode *LDBase = nullptr;
5405 unsigned LastLoadedElt = -1U;
5407 // For each element in the initializer, see if we've found a load or an undef.
5408 // If we don't find an initial load element, or later load elements are
5409 // non-consecutive, bail out.
5410 for (unsigned i = 0; i < NumElems; ++i) {
5411 SDValue Elt = Elts[i];
5412 // Look through a bitcast.
5413 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5414 Elt = Elt.getOperand(0);
5415 if (!Elt.getNode() ||
5416 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5419 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5421 LDBase = cast<LoadSDNode>(Elt.getNode());
5425 if (Elt.getOpcode() == ISD::UNDEF)
5428 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5429 EVT LdVT = Elt.getValueType();
5430 // Each loaded element must be the correct fractional portion of the
5431 // requested vector load.
5432 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5434 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5439 // If we have found an entire vector of loads and undefs, then return a large
5440 // load of the entire vector width starting at the base pointer. If we found
5441 // consecutive loads for the low half, generate a vzext_load node.
5442 if (LastLoadedElt == NumElems - 1) {
5443 assert(LDBase && "Did not find base load for merging consecutive loads");
5444 EVT EltVT = LDBase->getValueType(0);
5445 // Ensure that the input vector size for the merged loads matches the
5446 // cumulative size of the input elements.
5447 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5450 if (isAfterLegalize &&
5451 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5454 SDValue NewLd = SDValue();
5456 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5457 LDBase->getPointerInfo(), LDBase->isVolatile(),
5458 LDBase->isNonTemporal(), LDBase->isInvariant(),
5459 LDBase->getAlignment());
5461 if (LDBase->hasAnyUseOfValue(1)) {
5462 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5464 SDValue(NewLd.getNode(), 1));
5465 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5466 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5467 SDValue(NewLd.getNode(), 1));
5473 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5474 //of a v4i32 / v4f32. It's probably worth generalizing.
5475 EVT EltVT = VT.getVectorElementType();
5476 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5477 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5478 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5479 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5481 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5482 LDBase->getPointerInfo(),
5483 LDBase->getAlignment(),
5484 false/*isVolatile*/, true/*ReadMem*/,
5487 // Make sure the newly-created LOAD is in the same position as LDBase in
5488 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5489 // update uses of LDBase's output chain to use the TokenFactor.
5490 if (LDBase->hasAnyUseOfValue(1)) {
5491 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5492 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5493 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5494 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5495 SDValue(ResNode.getNode(), 1));
5498 return DAG.getBitcast(VT, ResNode);
5503 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5504 /// to generate a splat value for the following cases:
5505 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5506 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5507 /// a scalar load, or a constant.
5508 /// The VBROADCAST node is returned when a pattern is found,
5509 /// or SDValue() otherwise.
5510 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5511 SelectionDAG &DAG) {
5512 // VBROADCAST requires AVX.
5513 // TODO: Splats could be generated for non-AVX CPUs using SSE
5514 // instructions, but there's less potential gain for only 128-bit vectors.
5515 if (!Subtarget->hasAVX())
5518 MVT VT = Op.getSimpleValueType();
5521 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5522 "Unsupported vector type for broadcast.");
5527 switch (Op.getOpcode()) {
5529 // Unknown pattern found.
5532 case ISD::BUILD_VECTOR: {
5533 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5534 BitVector UndefElements;
5535 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5537 // We need a splat of a single value to use broadcast, and it doesn't
5538 // make any sense if the value is only in one element of the vector.
5539 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5543 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5544 Ld.getOpcode() == ISD::ConstantFP);
5546 // Make sure that all of the users of a non-constant load are from the
5547 // BUILD_VECTOR node.
5548 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5553 case ISD::VECTOR_SHUFFLE: {
5554 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5556 // Shuffles must have a splat mask where the first element is
5558 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5561 SDValue Sc = Op.getOperand(0);
5562 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5563 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5565 if (!Subtarget->hasInt256())
5568 // Use the register form of the broadcast instruction available on AVX2.
5569 if (VT.getSizeInBits() >= 256)
5570 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5571 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5574 Ld = Sc.getOperand(0);
5575 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5576 Ld.getOpcode() == ISD::ConstantFP);
5578 // The scalar_to_vector node and the suspected
5579 // load node must have exactly one user.
5580 // Constants may have multiple users.
5582 // AVX-512 has register version of the broadcast
5583 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5584 Ld.getValueType().getSizeInBits() >= 32;
5585 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5592 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5593 bool IsGE256 = (VT.getSizeInBits() >= 256);
5595 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5596 // instruction to save 8 or more bytes of constant pool data.
5597 // TODO: If multiple splats are generated to load the same constant,
5598 // it may be detrimental to overall size. There needs to be a way to detect
5599 // that condition to know if this is truly a size win.
5600 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5602 // Handle broadcasting a single constant scalar from the constant pool
5604 // On Sandybridge (no AVX2), it is still better to load a constant vector
5605 // from the constant pool and not to broadcast it from a scalar.
5606 // But override that restriction when optimizing for size.
5607 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5608 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5609 EVT CVT = Ld.getValueType();
5610 assert(!CVT.isVector() && "Must not broadcast a vector type");
5612 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5613 // For size optimization, also splat v2f64 and v2i64, and for size opt
5614 // with AVX2, also splat i8 and i16.
5615 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5616 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5617 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5618 const Constant *C = nullptr;
5619 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5620 C = CI->getConstantIntValue();
5621 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5622 C = CF->getConstantFPValue();
5624 assert(C && "Invalid constant type");
5626 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5628 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5629 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5631 CVT, dl, DAG.getEntryNode(), CP,
5632 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5633 false, false, Alignment);
5635 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5639 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5641 // Handle AVX2 in-register broadcasts.
5642 if (!IsLoad && Subtarget->hasInt256() &&
5643 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5644 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5646 // The scalar source must be a normal load.
5650 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5651 (Subtarget->hasVLX() && ScalarSize == 64))
5652 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5654 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5655 // double since there is no vbroadcastsd xmm
5656 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5657 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5658 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5661 // Unsupported broadcast.
5665 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5666 /// underlying vector and index.
5668 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5670 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5672 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5673 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5676 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5678 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5680 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5681 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5684 // In this case the vector is the extract_subvector expression and the index
5685 // is 2, as specified by the shuffle.
5686 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5687 SDValue ShuffleVec = SVOp->getOperand(0);
5688 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5689 assert(ShuffleVecVT.getVectorElementType() ==
5690 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5692 int ShuffleIdx = SVOp->getMaskElt(Idx);
5693 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5694 ExtractedFromVec = ShuffleVec;
5700 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5701 MVT VT = Op.getSimpleValueType();
5703 // Skip if insert_vec_elt is not supported.
5704 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5705 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5709 unsigned NumElems = Op.getNumOperands();
5713 SmallVector<unsigned, 4> InsertIndices;
5714 SmallVector<int, 8> Mask(NumElems, -1);
5716 for (unsigned i = 0; i != NumElems; ++i) {
5717 unsigned Opc = Op.getOperand(i).getOpcode();
5719 if (Opc == ISD::UNDEF)
5722 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5723 // Quit if more than 1 elements need inserting.
5724 if (InsertIndices.size() > 1)
5727 InsertIndices.push_back(i);
5731 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5732 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5733 // Quit if non-constant index.
5734 if (!isa<ConstantSDNode>(ExtIdx))
5736 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5738 // Quit if extracted from vector of different type.
5739 if (ExtractedFromVec.getValueType() != VT)
5742 if (!VecIn1.getNode())
5743 VecIn1 = ExtractedFromVec;
5744 else if (VecIn1 != ExtractedFromVec) {
5745 if (!VecIn2.getNode())
5746 VecIn2 = ExtractedFromVec;
5747 else if (VecIn2 != ExtractedFromVec)
5748 // Quit if more than 2 vectors to shuffle
5752 if (ExtractedFromVec == VecIn1)
5754 else if (ExtractedFromVec == VecIn2)
5755 Mask[i] = Idx + NumElems;
5758 if (!VecIn1.getNode())
5761 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5762 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5763 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5764 unsigned Idx = InsertIndices[i];
5765 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5766 DAG.getIntPtrConstant(Idx, DL));
5772 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5773 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5774 Op.getScalarValueSizeInBits() == 1 &&
5775 "Can not convert non-constant vector");
5776 uint64_t Immediate = 0;
5777 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5778 SDValue In = Op.getOperand(idx);
5779 if (In.getOpcode() != ISD::UNDEF)
5780 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5784 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5785 return DAG.getConstant(Immediate, dl, VT);
5787 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5789 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5791 MVT VT = Op.getSimpleValueType();
5792 assert((VT.getVectorElementType() == MVT::i1) &&
5793 "Unexpected type in LowerBUILD_VECTORvXi1!");
5796 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5797 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5798 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5799 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5802 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5803 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5804 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5805 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5808 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5809 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5810 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5811 return DAG.getBitcast(VT, Imm);
5812 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5813 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5814 DAG.getIntPtrConstant(0, dl));
5817 // Vector has one or more non-const elements
5818 uint64_t Immediate = 0;
5819 SmallVector<unsigned, 16> NonConstIdx;
5820 bool IsSplat = true;
5821 bool HasConstElts = false;
5823 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5824 SDValue In = Op.getOperand(idx);
5825 if (In.getOpcode() == ISD::UNDEF)
5827 if (!isa<ConstantSDNode>(In))
5828 NonConstIdx.push_back(idx);
5830 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5831 HasConstElts = true;
5835 else if (In != Op.getOperand(SplatIdx))
5839 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5841 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5842 DAG.getConstant(1, dl, VT),
5843 DAG.getConstant(0, dl, VT));
5845 // insert elements one by one
5849 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5850 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5852 else if (HasConstElts)
5853 Imm = DAG.getConstant(0, dl, VT);
5855 Imm = DAG.getUNDEF(VT);
5856 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5857 DstVec = DAG.getBitcast(VT, Imm);
5859 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5860 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5861 DAG.getIntPtrConstant(0, dl));
5864 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5865 unsigned InsertIdx = NonConstIdx[i];
5866 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5867 Op.getOperand(InsertIdx),
5868 DAG.getIntPtrConstant(InsertIdx, dl));
5873 /// \brief Return true if \p N implements a horizontal binop and return the
5874 /// operands for the horizontal binop into V0 and V1.
5876 /// This is a helper function of LowerToHorizontalOp().
5877 /// This function checks that the build_vector \p N in input implements a
5878 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5879 /// operation to match.
5880 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5881 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5882 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5885 /// This function only analyzes elements of \p N whose indices are
5886 /// in range [BaseIdx, LastIdx).
5887 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5889 unsigned BaseIdx, unsigned LastIdx,
5890 SDValue &V0, SDValue &V1) {
5891 EVT VT = N->getValueType(0);
5893 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5894 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5895 "Invalid Vector in input!");
5897 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5898 bool CanFold = true;
5899 unsigned ExpectedVExtractIdx = BaseIdx;
5900 unsigned NumElts = LastIdx - BaseIdx;
5901 V0 = DAG.getUNDEF(VT);
5902 V1 = DAG.getUNDEF(VT);
5904 // Check if N implements a horizontal binop.
5905 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5906 SDValue Op = N->getOperand(i + BaseIdx);
5909 if (Op->getOpcode() == ISD::UNDEF) {
5910 // Update the expected vector extract index.
5911 if (i * 2 == NumElts)
5912 ExpectedVExtractIdx = BaseIdx;
5913 ExpectedVExtractIdx += 2;
5917 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5922 SDValue Op0 = Op.getOperand(0);
5923 SDValue Op1 = Op.getOperand(1);
5925 // Try to match the following pattern:
5926 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5927 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5928 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5929 Op0.getOperand(0) == Op1.getOperand(0) &&
5930 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5931 isa<ConstantSDNode>(Op1.getOperand(1)));
5935 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5936 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5938 if (i * 2 < NumElts) {
5939 if (V0.getOpcode() == ISD::UNDEF) {
5940 V0 = Op0.getOperand(0);
5941 if (V0.getValueType() != VT)
5945 if (V1.getOpcode() == ISD::UNDEF) {
5946 V1 = Op0.getOperand(0);
5947 if (V1.getValueType() != VT)
5950 if (i * 2 == NumElts)
5951 ExpectedVExtractIdx = BaseIdx;
5954 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5955 if (I0 == ExpectedVExtractIdx)
5956 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5957 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5958 // Try to match the following dag sequence:
5959 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5960 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5964 ExpectedVExtractIdx += 2;
5970 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5971 /// a concat_vector.
5973 /// This is a helper function of LowerToHorizontalOp().
5974 /// This function expects two 256-bit vectors called V0 and V1.
5975 /// At first, each vector is split into two separate 128-bit vectors.
5976 /// Then, the resulting 128-bit vectors are used to implement two
5977 /// horizontal binary operations.
5979 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5981 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5982 /// the two new horizontal binop.
5983 /// When Mode is set, the first horizontal binop dag node would take as input
5984 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5985 /// horizontal binop dag node would take as input the lower 128-bit of V1
5986 /// and the upper 128-bit of V1.
5988 /// HADD V0_LO, V0_HI
5989 /// HADD V1_LO, V1_HI
5991 /// Otherwise, the first horizontal binop dag node takes as input the lower
5992 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5993 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5995 /// HADD V0_LO, V1_LO
5996 /// HADD V0_HI, V1_HI
5998 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5999 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6000 /// the upper 128-bits of the result.
6001 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6002 SDLoc DL, SelectionDAG &DAG,
6003 unsigned X86Opcode, bool Mode,
6004 bool isUndefLO, bool isUndefHI) {
6005 EVT VT = V0.getValueType();
6006 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6007 "Invalid nodes in input!");
6009 unsigned NumElts = VT.getVectorNumElements();
6010 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6011 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6012 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6013 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6014 EVT NewVT = V0_LO.getValueType();
6016 SDValue LO = DAG.getUNDEF(NewVT);
6017 SDValue HI = DAG.getUNDEF(NewVT);
6020 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6021 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6022 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6023 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6024 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6026 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6027 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6028 V1_LO->getOpcode() != ISD::UNDEF))
6029 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6031 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6032 V1_HI->getOpcode() != ISD::UNDEF))
6033 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6036 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6039 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
6041 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
6042 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6043 MVT VT = BV->getSimpleValueType(0);
6044 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
6045 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
6049 unsigned NumElts = VT.getVectorNumElements();
6050 SDValue InVec0 = DAG.getUNDEF(VT);
6051 SDValue InVec1 = DAG.getUNDEF(VT);
6053 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6054 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6056 // Odd-numbered elements in the input build vector are obtained from
6057 // adding two integer/float elements.
6058 // Even-numbered elements in the input build vector are obtained from
6059 // subtracting two integer/float elements.
6060 unsigned ExpectedOpcode = ISD::FSUB;
6061 unsigned NextExpectedOpcode = ISD::FADD;
6062 bool AddFound = false;
6063 bool SubFound = false;
6065 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6066 SDValue Op = BV->getOperand(i);
6068 // Skip 'undef' values.
6069 unsigned Opcode = Op.getOpcode();
6070 if (Opcode == ISD::UNDEF) {
6071 std::swap(ExpectedOpcode, NextExpectedOpcode);
6075 // Early exit if we found an unexpected opcode.
6076 if (Opcode != ExpectedOpcode)
6079 SDValue Op0 = Op.getOperand(0);
6080 SDValue Op1 = Op.getOperand(1);
6082 // Try to match the following pattern:
6083 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6084 // Early exit if we cannot match that sequence.
6085 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6086 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6087 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6088 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6089 Op0.getOperand(1) != Op1.getOperand(1))
6092 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6096 // We found a valid add/sub node. Update the information accordingly.
6102 // Update InVec0 and InVec1.
6103 if (InVec0.getOpcode() == ISD::UNDEF) {
6104 InVec0 = Op0.getOperand(0);
6105 if (InVec0.getSimpleValueType() != VT)
6108 if (InVec1.getOpcode() == ISD::UNDEF) {
6109 InVec1 = Op1.getOperand(0);
6110 if (InVec1.getSimpleValueType() != VT)
6114 // Make sure that operands in input to each add/sub node always
6115 // come from a same pair of vectors.
6116 if (InVec0 != Op0.getOperand(0)) {
6117 if (ExpectedOpcode == ISD::FSUB)
6120 // FADD is commutable. Try to commute the operands
6121 // and then test again.
6122 std::swap(Op0, Op1);
6123 if (InVec0 != Op0.getOperand(0))
6127 if (InVec1 != Op1.getOperand(0))
6130 // Update the pair of expected opcodes.
6131 std::swap(ExpectedOpcode, NextExpectedOpcode);
6134 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6135 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6136 InVec1.getOpcode() != ISD::UNDEF)
6137 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6142 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6143 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6144 const X86Subtarget *Subtarget,
6145 SelectionDAG &DAG) {
6146 MVT VT = BV->getSimpleValueType(0);
6147 unsigned NumElts = VT.getVectorNumElements();
6148 unsigned NumUndefsLO = 0;
6149 unsigned NumUndefsHI = 0;
6150 unsigned Half = NumElts/2;
6152 // Count the number of UNDEF operands in the build_vector in input.
6153 for (unsigned i = 0, e = Half; i != e; ++i)
6154 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6157 for (unsigned i = Half, e = NumElts; i != e; ++i)
6158 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6161 // Early exit if this is either a build_vector of all UNDEFs or all the
6162 // operands but one are UNDEF.
6163 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6167 SDValue InVec0, InVec1;
6168 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6169 // Try to match an SSE3 float HADD/HSUB.
6170 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6171 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6173 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6174 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6175 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6176 // Try to match an SSSE3 integer HADD/HSUB.
6177 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6178 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6180 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6181 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6184 if (!Subtarget->hasAVX())
6187 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6188 // Try to match an AVX horizontal add/sub of packed single/double
6189 // precision floating point values from 256-bit vectors.
6190 SDValue InVec2, InVec3;
6191 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6192 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6193 ((InVec0.getOpcode() == ISD::UNDEF ||
6194 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6195 ((InVec1.getOpcode() == ISD::UNDEF ||
6196 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6197 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6199 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6200 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6201 ((InVec0.getOpcode() == ISD::UNDEF ||
6202 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6203 ((InVec1.getOpcode() == ISD::UNDEF ||
6204 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6205 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6206 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6207 // Try to match an AVX2 horizontal add/sub of signed integers.
6208 SDValue InVec2, InVec3;
6210 bool CanFold = true;
6212 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6213 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6214 ((InVec0.getOpcode() == ISD::UNDEF ||
6215 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6216 ((InVec1.getOpcode() == ISD::UNDEF ||
6217 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6218 X86Opcode = X86ISD::HADD;
6219 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6220 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6221 ((InVec0.getOpcode() == ISD::UNDEF ||
6222 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6223 ((InVec1.getOpcode() == ISD::UNDEF ||
6224 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6225 X86Opcode = X86ISD::HSUB;
6230 // Fold this build_vector into a single horizontal add/sub.
6231 // Do this only if the target has AVX2.
6232 if (Subtarget->hasAVX2())
6233 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6235 // Do not try to expand this build_vector into a pair of horizontal
6236 // add/sub if we can emit a pair of scalar add/sub.
6237 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6240 // Convert this build_vector into a pair of horizontal binop followed by
6242 bool isUndefLO = NumUndefsLO == Half;
6243 bool isUndefHI = NumUndefsHI == Half;
6244 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6245 isUndefLO, isUndefHI);
6249 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6250 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6252 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6253 X86Opcode = X86ISD::HADD;
6254 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6255 X86Opcode = X86ISD::HSUB;
6256 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6257 X86Opcode = X86ISD::FHADD;
6258 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6259 X86Opcode = X86ISD::FHSUB;
6263 // Don't try to expand this build_vector into a pair of horizontal add/sub
6264 // if we can simply emit a pair of scalar add/sub.
6265 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6268 // Convert this build_vector into two horizontal add/sub followed by
6270 bool isUndefLO = NumUndefsLO == Half;
6271 bool isUndefHI = NumUndefsHI == Half;
6272 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6273 isUndefLO, isUndefHI);
6280 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6283 MVT VT = Op.getSimpleValueType();
6284 MVT ExtVT = VT.getVectorElementType();
6285 unsigned NumElems = Op.getNumOperands();
6287 // Generate vectors for predicate vectors.
6288 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6289 return LowerBUILD_VECTORvXi1(Op, DAG);
6291 // Vectors containing all zeros can be matched by pxor and xorps later
6292 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6293 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6294 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6295 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6298 return getZeroVector(VT, Subtarget, DAG, dl);
6301 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6302 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6303 // vpcmpeqd on 256-bit vectors.
6304 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6305 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6308 if (!VT.is512BitVector())
6309 return getOnesVector(VT, Subtarget, DAG, dl);
6312 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6313 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6315 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6316 return HorizontalOp;
6317 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6320 unsigned EVTBits = ExtVT.getSizeInBits();
6322 unsigned NumZero = 0;
6323 unsigned NumNonZero = 0;
6324 uint64_t NonZeros = 0;
6325 bool IsAllConstants = true;
6326 SmallSet<SDValue, 8> Values;
6327 for (unsigned i = 0; i < NumElems; ++i) {
6328 SDValue Elt = Op.getOperand(i);
6329 if (Elt.getOpcode() == ISD::UNDEF)
6332 if (Elt.getOpcode() != ISD::Constant &&
6333 Elt.getOpcode() != ISD::ConstantFP)
6334 IsAllConstants = false;
6335 if (X86::isZeroNode(Elt))
6338 assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
6339 NonZeros |= ((uint64_t)1 << i);
6344 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6345 if (NumNonZero == 0)
6346 return DAG.getUNDEF(VT);
6348 // Special case for single non-zero, non-undef, element.
6349 if (NumNonZero == 1) {
6350 unsigned Idx = countTrailingZeros(NonZeros);
6351 SDValue Item = Op.getOperand(Idx);
6353 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6354 // the value are obviously zero, truncate the value to i32 and do the
6355 // insertion that way. Only do this if the value is non-constant or if the
6356 // value is a constant being inserted into element 0. It is cheaper to do
6357 // a constant pool load than it is to do a movd + shuffle.
6358 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6359 (!IsAllConstants || Idx == 0)) {
6360 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6362 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6363 MVT VecVT = MVT::v4i32;
6365 // Truncate the value (which may itself be a constant) to i32, and
6366 // convert it to a vector with movd (S2V+shuffle to zero extend).
6367 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6368 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6369 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6370 Item, Idx * 2, true, Subtarget, DAG));
6374 // If we have a constant or non-constant insertion into the low element of
6375 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6376 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6377 // depending on what the source datatype is.
6380 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6382 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6383 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6384 if (VT.is512BitVector()) {
6385 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6386 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6387 Item, DAG.getIntPtrConstant(0, dl));
6389 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6390 "Expected an SSE value type!");
6391 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6392 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6393 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6396 // We can't directly insert an i8 or i16 into a vector, so zero extend
6398 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6399 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6400 if (VT.is256BitVector()) {
6401 if (Subtarget->hasAVX()) {
6402 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6403 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6405 // Without AVX, we need to extend to a 128-bit vector and then
6406 // insert into the 256-bit vector.
6407 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6408 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6409 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6412 assert(VT.is128BitVector() && "Expected an SSE value type!");
6413 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6414 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6416 return DAG.getBitcast(VT, Item);
6420 // Is it a vector logical left shift?
6421 if (NumElems == 2 && Idx == 1 &&
6422 X86::isZeroNode(Op.getOperand(0)) &&
6423 !X86::isZeroNode(Op.getOperand(1))) {
6424 unsigned NumBits = VT.getSizeInBits();
6425 return getVShift(true, VT,
6426 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6427 VT, Op.getOperand(1)),
6428 NumBits/2, DAG, *this, dl);
6431 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6434 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6435 // is a non-constant being inserted into an element other than the low one,
6436 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6437 // movd/movss) to move this into the low element, then shuffle it into
6439 if (EVTBits == 32) {
6440 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6441 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6445 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6446 if (Values.size() == 1) {
6447 if (EVTBits == 32) {
6448 // Instead of a shuffle like this:
6449 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6450 // Check if it's possible to issue this instead.
6451 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6452 unsigned Idx = countTrailingZeros(NonZeros);
6453 SDValue Item = Op.getOperand(Idx);
6454 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6455 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6460 // A vector full of immediates; various special cases are already
6461 // handled, so this is best done with a single constant-pool load.
6465 // For AVX-length vectors, see if we can use a vector load to get all of the
6466 // elements, otherwise build the individual 128-bit pieces and use
6467 // shuffles to put them in place.
6468 if (VT.is256BitVector() || VT.is512BitVector()) {
6469 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6471 // Check for a build vector of consecutive loads.
6472 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6475 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6477 // Build both the lower and upper subvector.
6478 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6479 makeArrayRef(&V[0], NumElems/2));
6480 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6481 makeArrayRef(&V[NumElems / 2], NumElems/2));
6483 // Recreate the wider vector with the lower and upper part.
6484 if (VT.is256BitVector())
6485 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6486 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6489 // Let legalizer expand 2-wide build_vectors.
6490 if (EVTBits == 64) {
6491 if (NumNonZero == 1) {
6492 // One half is zero or undef.
6493 unsigned Idx = countTrailingZeros(NonZeros);
6494 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6495 Op.getOperand(Idx));
6496 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6501 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6502 if (EVTBits == 8 && NumElems == 16)
6503 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
6504 DAG, Subtarget, *this))
6507 if (EVTBits == 16 && NumElems == 8)
6508 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
6509 DAG, Subtarget, *this))
6512 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6513 if (EVTBits == 32 && NumElems == 4)
6514 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6517 // If element VT is == 32 bits, turn it into a number of shuffles.
6518 SmallVector<SDValue, 8> V(NumElems);
6519 if (NumElems == 4 && NumZero > 0) {
6520 for (unsigned i = 0; i < 4; ++i) {
6521 bool isZero = !(NonZeros & (1ULL << i));
6523 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6525 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6528 for (unsigned i = 0; i < 2; ++i) {
6529 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6532 V[i] = V[i*2]; // Must be a zero vector.
6535 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6538 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6541 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6546 bool Reverse1 = (NonZeros & 0x3) == 2;
6547 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6551 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6552 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6554 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6557 if (Values.size() > 1 && VT.is128BitVector()) {
6558 // Check for a build vector of consecutive loads.
6559 for (unsigned i = 0; i < NumElems; ++i)
6560 V[i] = Op.getOperand(i);
6562 // Check for elements which are consecutive loads.
6563 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6566 // Check for a build vector from mostly shuffle plus few inserting.
6567 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6570 // For SSE 4.1, use insertps to put the high elements into the low element.
6571 if (Subtarget->hasSSE41()) {
6573 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6574 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6576 Result = DAG.getUNDEF(VT);
6578 for (unsigned i = 1; i < NumElems; ++i) {
6579 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6580 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6581 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6586 // Otherwise, expand into a number of unpckl*, start by extending each of
6587 // our (non-undef) elements to the full vector width with the element in the
6588 // bottom slot of the vector (which generates no code for SSE).
6589 for (unsigned i = 0; i < NumElems; ++i) {
6590 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6591 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6593 V[i] = DAG.getUNDEF(VT);
6596 // Next, we iteratively mix elements, e.g. for v4f32:
6597 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6598 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6599 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6600 unsigned EltStride = NumElems >> 1;
6601 while (EltStride != 0) {
6602 for (unsigned i = 0; i < EltStride; ++i) {
6603 // If V[i+EltStride] is undef and this is the first round of mixing,
6604 // then it is safe to just drop this shuffle: V[i] is already in the
6605 // right place, the one element (since it's the first round) being
6606 // inserted as undef can be dropped. This isn't safe for successive
6607 // rounds because they will permute elements within both vectors.
6608 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6609 EltStride == NumElems/2)
6612 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6621 // 256-bit AVX can use the vinsertf128 instruction
6622 // to create 256-bit vectors from two other 128-bit ones.
6623 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6625 MVT ResVT = Op.getSimpleValueType();
6627 assert((ResVT.is256BitVector() ||
6628 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6630 SDValue V1 = Op.getOperand(0);
6631 SDValue V2 = Op.getOperand(1);
6632 unsigned NumElems = ResVT.getVectorNumElements();
6633 if (ResVT.is256BitVector())
6634 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6636 if (Op.getNumOperands() == 4) {
6637 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6638 ResVT.getVectorNumElements()/2);
6639 SDValue V3 = Op.getOperand(2);
6640 SDValue V4 = Op.getOperand(3);
6641 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6642 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6644 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6647 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6648 const X86Subtarget *Subtarget,
6649 SelectionDAG & DAG) {
6651 MVT ResVT = Op.getSimpleValueType();
6652 unsigned NumOfOperands = Op.getNumOperands();
6654 assert(isPowerOf2_32(NumOfOperands) &&
6655 "Unexpected number of operands in CONCAT_VECTORS");
6657 SDValue Undef = DAG.getUNDEF(ResVT);
6658 if (NumOfOperands > 2) {
6659 // Specialize the cases when all, or all but one, of the operands are undef.
6660 unsigned NumOfDefinedOps = 0;
6662 for (unsigned i = 0; i < NumOfOperands; i++)
6663 if (!Op.getOperand(i).isUndef()) {
6667 if (NumOfDefinedOps == 0)
6669 if (NumOfDefinedOps == 1) {
6670 unsigned SubVecNumElts =
6671 Op.getOperand(OpIdx).getValueType().getVectorNumElements();
6672 SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
6673 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
6674 Op.getOperand(OpIdx), IdxVal);
6677 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6678 ResVT.getVectorNumElements()/2);
6679 SmallVector<SDValue, 2> Ops;
6680 for (unsigned i = 0; i < NumOfOperands/2; i++)
6681 Ops.push_back(Op.getOperand(i));
6682 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6684 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6685 Ops.push_back(Op.getOperand(i));
6686 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6687 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6691 SDValue V1 = Op.getOperand(0);
6692 SDValue V2 = Op.getOperand(1);
6693 unsigned NumElems = ResVT.getVectorNumElements();
6694 assert(V1.getValueType() == V2.getValueType() &&
6695 V1.getValueType().getVectorNumElements() == NumElems/2 &&
6696 "Unexpected operands in CONCAT_VECTORS");
6698 if (ResVT.getSizeInBits() >= 16)
6699 return Op; // The operation is legal with KUNPCK
6701 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6702 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6703 SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
6704 if (IsZeroV1 && IsZeroV2)
6707 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6709 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6711 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
6713 SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
6715 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
6718 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
6720 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6721 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
6724 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6725 const X86Subtarget *Subtarget,
6726 SelectionDAG &DAG) {
6727 MVT VT = Op.getSimpleValueType();
6728 if (VT.getVectorElementType() == MVT::i1)
6729 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6731 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6732 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6733 Op.getNumOperands() == 4)));
6735 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6736 // from two other 128-bit ones.
6738 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6739 return LowerAVXCONCAT_VECTORS(Op, DAG);
6742 //===----------------------------------------------------------------------===//
6743 // Vector shuffle lowering
6745 // This is an experimental code path for lowering vector shuffles on x86. It is
6746 // designed to handle arbitrary vector shuffles and blends, gracefully
6747 // degrading performance as necessary. It works hard to recognize idiomatic
6748 // shuffles and lower them to optimal instruction patterns without leaving
6749 // a framework that allows reasonably efficient handling of all vector shuffle
6751 //===----------------------------------------------------------------------===//
6753 /// \brief Tiny helper function to identify a no-op mask.
6755 /// This is a somewhat boring predicate function. It checks whether the mask
6756 /// array input, which is assumed to be a single-input shuffle mask of the kind
6757 /// used by the X86 shuffle instructions (not a fully general
6758 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6759 /// in-place shuffle are 'no-op's.
6760 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6761 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6762 if (Mask[i] != -1 && Mask[i] != i)
6767 /// \brief Helper function to classify a mask as a single-input mask.
6769 /// This isn't a generic single-input test because in the vector shuffle
6770 /// lowering we canonicalize single inputs to be the first input operand. This
6771 /// means we can more quickly test for a single input by only checking whether
6772 /// an input from the second operand exists. We also assume that the size of
6773 /// mask corresponds to the size of the input vectors which isn't true in the
6774 /// fully general case.
6775 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6777 if (M >= (int)Mask.size())
6782 /// \brief Test whether there are elements crossing 128-bit lanes in this
6785 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6786 /// and we routinely test for these.
6787 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6788 int LaneSize = 128 / VT.getScalarSizeInBits();
6789 int Size = Mask.size();
6790 for (int i = 0; i < Size; ++i)
6791 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6796 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6798 /// This checks a shuffle mask to see if it is performing the same
6799 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6800 /// that it is also not lane-crossing. It may however involve a blend from the
6801 /// same lane of a second vector.
6803 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6804 /// non-trivial to compute in the face of undef lanes. The representation is
6805 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6806 /// entries from both V1 and V2 inputs to the wider mask.
6808 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6809 SmallVectorImpl<int> &RepeatedMask) {
6810 int LaneSize = 128 / VT.getScalarSizeInBits();
6811 RepeatedMask.resize(LaneSize, -1);
6812 int Size = Mask.size();
6813 for (int i = 0; i < Size; ++i) {
6816 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6817 // This entry crosses lanes, so there is no way to model this shuffle.
6820 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6821 if (RepeatedMask[i % LaneSize] == -1)
6822 // This is the first non-undef entry in this slot of a 128-bit lane.
6823 RepeatedMask[i % LaneSize] =
6824 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6825 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6826 // Found a mismatch with the repeated mask.
6832 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6835 /// This is a fast way to test a shuffle mask against a fixed pattern:
6837 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6839 /// It returns true if the mask is exactly as wide as the argument list, and
6840 /// each element of the mask is either -1 (signifying undef) or the value given
6841 /// in the argument.
6842 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6843 ArrayRef<int> ExpectedMask) {
6844 if (Mask.size() != ExpectedMask.size())
6847 int Size = Mask.size();
6849 // If the values are build vectors, we can look through them to find
6850 // equivalent inputs that make the shuffles equivalent.
6851 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6852 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6854 for (int i = 0; i < Size; ++i)
6855 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6856 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6857 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6858 if (!MaskBV || !ExpectedBV ||
6859 MaskBV->getOperand(Mask[i] % Size) !=
6860 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6867 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6869 /// This helper function produces an 8-bit shuffle immediate corresponding to
6870 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6871 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6874 /// NB: We rely heavily on "undef" masks preserving the input lane.
6875 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6876 SelectionDAG &DAG) {
6877 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6878 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6879 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6880 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6881 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6884 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6885 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6886 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6887 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6888 return DAG.getConstant(Imm, DL, MVT::i8);
6891 /// \brief Compute whether each element of a shuffle is zeroable.
6893 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6894 /// Either it is an undef element in the shuffle mask, the element of the input
6895 /// referenced is undef, or the element of the input referenced is known to be
6896 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6897 /// as many lanes with this technique as possible to simplify the remaining
6899 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6900 SDValue V1, SDValue V2) {
6901 SmallBitVector Zeroable(Mask.size(), false);
6903 while (V1.getOpcode() == ISD::BITCAST)
6904 V1 = V1->getOperand(0);
6905 while (V2.getOpcode() == ISD::BITCAST)
6906 V2 = V2->getOperand(0);
6908 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6909 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6911 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6913 // Handle the easy cases.
6914 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6919 // If this is an index into a build_vector node (which has the same number
6920 // of elements), dig out the input value and use it.
6921 SDValue V = M < Size ? V1 : V2;
6922 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6925 SDValue Input = V.getOperand(M % Size);
6926 // The UNDEF opcode check really should be dead code here, but not quite
6927 // worth asserting on (it isn't invalid, just unexpected).
6928 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6935 // X86 has dedicated unpack instructions that can handle specific blend
6936 // operations: UNPCKH and UNPCKL.
6937 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6938 SDValue V1, SDValue V2,
6939 SelectionDAG &DAG) {
6940 int NumElts = VT.getVectorNumElements();
6941 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6942 SmallVector<int, 8> Unpckl;
6943 SmallVector<int, 8> Unpckh;
6945 for (int i = 0; i < NumElts; ++i) {
6946 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6947 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6948 int HiPos = LoPos + NumEltsInLane / 2;
6949 Unpckl.push_back(LoPos);
6950 Unpckh.push_back(HiPos);
6953 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6954 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6955 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6956 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6958 // Commute and try again.
6959 ShuffleVectorSDNode::commuteMask(Unpckl);
6960 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6961 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6963 ShuffleVectorSDNode::commuteMask(Unpckh);
6964 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6965 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6970 /// \brief Try to emit a bitmask instruction for a shuffle.
6972 /// This handles cases where we can model a blend exactly as a bitmask due to
6973 /// one of the inputs being zeroable.
6974 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6975 SDValue V2, ArrayRef<int> Mask,
6976 SelectionDAG &DAG) {
6977 MVT EltVT = VT.getVectorElementType();
6978 int NumEltBits = EltVT.getSizeInBits();
6979 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6980 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6981 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6983 if (EltVT.isFloatingPoint()) {
6984 Zero = DAG.getBitcast(EltVT, Zero);
6985 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6987 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6988 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6990 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6993 if (Mask[i] % Size != i)
6994 return SDValue(); // Not a blend.
6996 V = Mask[i] < Size ? V1 : V2;
6997 else if (V != (Mask[i] < Size ? V1 : V2))
6998 return SDValue(); // Can only let one input through the mask.
7000 VMaskOps[i] = AllOnes;
7003 return SDValue(); // No non-zeroable elements!
7005 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
7006 V = DAG.getNode(VT.isFloatingPoint()
7007 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
7012 /// \brief Try to emit a blend instruction for a shuffle using bit math.
7014 /// This is used as a fallback approach when first class blend instructions are
7015 /// unavailable. Currently it is only suitable for integer vectors, but could
7016 /// be generalized for floating point vectors if desirable.
7017 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
7018 SDValue V2, ArrayRef<int> Mask,
7019 SelectionDAG &DAG) {
7020 assert(VT.isInteger() && "Only supports integer vector types!");
7021 MVT EltVT = VT.getVectorElementType();
7022 int NumEltBits = EltVT.getSizeInBits();
7023 SDValue Zero = DAG.getConstant(0, DL, EltVT);
7024 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
7026 SmallVector<SDValue, 16> MaskOps;
7027 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7028 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
7029 return SDValue(); // Shuffled input!
7030 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
7033 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
7034 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
7035 // We have to cast V2 around.
7036 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
7037 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
7038 DAG.getBitcast(MaskVT, V1Mask),
7039 DAG.getBitcast(MaskVT, V2)));
7040 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
7043 /// \brief Try to emit a blend instruction for a shuffle.
7045 /// This doesn't do any checks for the availability of instructions for blending
7046 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7047 /// be matched in the backend with the type given. What it does check for is
7048 /// that the shuffle mask is a blend, or convertible into a blend with zero.
7049 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7050 SDValue V2, ArrayRef<int> Original,
7051 const X86Subtarget *Subtarget,
7052 SelectionDAG &DAG) {
7053 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7054 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7055 SmallVector<int, 8> Mask(Original.begin(), Original.end());
7056 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7057 bool ForceV1Zero = false, ForceV2Zero = false;
7059 // Attempt to generate the binary blend mask. If an input is zero then
7060 // we can use any lane.
7061 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
7062 unsigned BlendMask = 0;
7063 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7069 if (M == i + Size) {
7070 BlendMask |= 1u << i;
7081 BlendMask |= 1u << i;
7086 return SDValue(); // Shuffled input!
7089 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
7091 V1 = getZeroVector(VT, Subtarget, DAG, DL);
7093 V2 = getZeroVector(VT, Subtarget, DAG, DL);
7095 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
7096 unsigned ScaledMask = 0;
7097 for (int i = 0; i != Size; ++i)
7098 if (BlendMask & (1u << i))
7099 for (int j = 0; j != Scale; ++j)
7100 ScaledMask |= 1u << (i * Scale + j);
7104 switch (VT.SimpleTy) {
7109 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7110 DAG.getConstant(BlendMask, DL, MVT::i8));
7114 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7118 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7119 // that instruction.
7120 if (Subtarget->hasAVX2()) {
7121 // Scale the blend by the number of 32-bit dwords per element.
7122 int Scale = VT.getScalarSizeInBits() / 32;
7123 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7124 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7125 V1 = DAG.getBitcast(BlendVT, V1);
7126 V2 = DAG.getBitcast(BlendVT, V2);
7127 return DAG.getBitcast(
7128 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7129 DAG.getConstant(BlendMask, DL, MVT::i8)));
7133 // For integer shuffles we need to expand the mask and cast the inputs to
7134 // v8i16s prior to blending.
7135 int Scale = 8 / VT.getVectorNumElements();
7136 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7137 V1 = DAG.getBitcast(MVT::v8i16, V1);
7138 V2 = DAG.getBitcast(MVT::v8i16, V2);
7139 return DAG.getBitcast(VT,
7140 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7141 DAG.getConstant(BlendMask, DL, MVT::i8)));
7145 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7146 SmallVector<int, 8> RepeatedMask;
7147 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7148 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7149 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7151 for (int i = 0; i < 8; ++i)
7152 if (RepeatedMask[i] >= 16)
7153 BlendMask |= 1u << i;
7154 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7155 DAG.getConstant(BlendMask, DL, MVT::i8));
7161 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
7162 "256-bit byte-blends require AVX2 support!");
7164 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
7165 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
7168 // Scale the blend by the number of bytes per element.
7169 int Scale = VT.getScalarSizeInBits() / 8;
7171 // This form of blend is always done on bytes. Compute the byte vector
7173 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7175 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7176 // mix of LLVM's code generator and the x86 backend. We tell the code
7177 // generator that boolean values in the elements of an x86 vector register
7178 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7179 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7180 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7181 // of the element (the remaining are ignored) and 0 in that high bit would
7182 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7183 // the LLVM model for boolean values in vector elements gets the relevant
7184 // bit set, it is set backwards and over constrained relative to x86's
7186 SmallVector<SDValue, 32> VSELECTMask;
7187 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7188 for (int j = 0; j < Scale; ++j)
7189 VSELECTMask.push_back(
7190 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7191 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7194 V1 = DAG.getBitcast(BlendVT, V1);
7195 V2 = DAG.getBitcast(BlendVT, V2);
7196 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7197 DAG.getNode(ISD::BUILD_VECTOR, DL,
7198 BlendVT, VSELECTMask),
7203 llvm_unreachable("Not a supported integer vector type!");
7207 /// \brief Try to lower as a blend of elements from two inputs followed by
7208 /// a single-input permutation.
7210 /// This matches the pattern where we can blend elements from two inputs and
7211 /// then reduce the shuffle to a single-input permutation.
7212 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7215 SelectionDAG &DAG) {
7216 // We build up the blend mask while checking whether a blend is a viable way
7217 // to reduce the shuffle.
7218 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7219 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7221 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7225 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7227 if (BlendMask[Mask[i] % Size] == -1)
7228 BlendMask[Mask[i] % Size] = Mask[i];
7229 else if (BlendMask[Mask[i] % Size] != Mask[i])
7230 return SDValue(); // Can't blend in the needed input!
7232 PermuteMask[i] = Mask[i] % Size;
7235 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7236 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7239 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7240 /// blends and permutes.
7242 /// This matches the extremely common pattern for handling combined
7243 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7244 /// operations. It will try to pick the best arrangement of shuffles and
7246 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7250 SelectionDAG &DAG) {
7251 // Shuffle the input elements into the desired positions in V1 and V2 and
7252 // blend them together.
7253 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7254 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7255 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7256 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7257 if (Mask[i] >= 0 && Mask[i] < Size) {
7258 V1Mask[i] = Mask[i];
7260 } else if (Mask[i] >= Size) {
7261 V2Mask[i] = Mask[i] - Size;
7262 BlendMask[i] = i + Size;
7265 // Try to lower with the simpler initial blend strategy unless one of the
7266 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7267 // shuffle may be able to fold with a load or other benefit. However, when
7268 // we'll have to do 2x as many shuffles in order to achieve this, blending
7269 // first is a better strategy.
7270 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7271 if (SDValue BlendPerm =
7272 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7275 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7276 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7277 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7280 /// \brief Try to lower a vector shuffle as a byte rotation.
7282 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7283 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7284 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7285 /// try to generically lower a vector shuffle through such an pattern. It
7286 /// does not check for the profitability of lowering either as PALIGNR or
7287 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7288 /// This matches shuffle vectors that look like:
7290 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7292 /// Essentially it concatenates V1 and V2, shifts right by some number of
7293 /// elements, and takes the low elements as the result. Note that while this is
7294 /// specified as a *right shift* because x86 is little-endian, it is a *left
7295 /// rotate* of the vector lanes.
7296 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7299 const X86Subtarget *Subtarget,
7300 SelectionDAG &DAG) {
7301 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7303 int NumElts = Mask.size();
7304 int NumLanes = VT.getSizeInBits() / 128;
7305 int NumLaneElts = NumElts / NumLanes;
7307 // We need to detect various ways of spelling a rotation:
7308 // [11, 12, 13, 14, 15, 0, 1, 2]
7309 // [-1, 12, 13, 14, -1, -1, 1, -1]
7310 // [-1, -1, -1, -1, -1, -1, 1, 2]
7311 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7312 // [-1, 4, 5, 6, -1, -1, 9, -1]
7313 // [-1, 4, 5, 6, -1, -1, -1, -1]
7316 for (int l = 0; l < NumElts; l += NumLaneElts) {
7317 for (int i = 0; i < NumLaneElts; ++i) {
7318 if (Mask[l + i] == -1)
7320 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7322 // Get the mod-Size index and lane correct it.
7323 int LaneIdx = (Mask[l + i] % NumElts) - l;
7324 // Make sure it was in this lane.
7325 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7328 // Determine where a rotated vector would have started.
7329 int StartIdx = i - LaneIdx;
7331 // The identity rotation isn't interesting, stop.
7334 // If we found the tail of a vector the rotation must be the missing
7335 // front. If we found the head of a vector, it must be how much of the
7337 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7340 Rotation = CandidateRotation;
7341 else if (Rotation != CandidateRotation)
7342 // The rotations don't match, so we can't match this mask.
7345 // Compute which value this mask is pointing at.
7346 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7348 // Compute which of the two target values this index should be assigned
7349 // to. This reflects whether the high elements are remaining or the low
7350 // elements are remaining.
7351 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7353 // Either set up this value if we've not encountered it before, or check
7354 // that it remains consistent.
7357 else if (TargetV != MaskV)
7358 // This may be a rotation, but it pulls from the inputs in some
7359 // unsupported interleaving.
7364 // Check that we successfully analyzed the mask, and normalize the results.
7365 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7366 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7372 // The actual rotate instruction rotates bytes, so we need to scale the
7373 // rotation based on how many bytes are in the vector lane.
7374 int Scale = 16 / NumLaneElts;
7376 // SSSE3 targets can use the palignr instruction.
7377 if (Subtarget->hasSSSE3()) {
7378 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7379 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7380 Lo = DAG.getBitcast(AlignVT, Lo);
7381 Hi = DAG.getBitcast(AlignVT, Hi);
7383 return DAG.getBitcast(
7384 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7385 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7388 assert(VT.is128BitVector() &&
7389 "Rotate-based lowering only supports 128-bit lowering!");
7390 assert(Mask.size() <= 16 &&
7391 "Can shuffle at most 16 bytes in a 128-bit vector!");
7393 // Default SSE2 implementation
7394 int LoByteShift = 16 - Rotation * Scale;
7395 int HiByteShift = Rotation * Scale;
7397 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7398 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7399 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7401 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7402 DAG.getConstant(LoByteShift, DL, MVT::i8));
7403 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7404 DAG.getConstant(HiByteShift, DL, MVT::i8));
7405 return DAG.getBitcast(VT,
7406 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7409 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7411 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7412 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7413 /// matches elements from one of the input vectors shuffled to the left or
7414 /// right with zeroable elements 'shifted in'. It handles both the strictly
7415 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7418 /// PSHL : (little-endian) left bit shift.
7419 /// [ zz, 0, zz, 2 ]
7420 /// [ -1, 4, zz, -1 ]
7421 /// PSRL : (little-endian) right bit shift.
7423 /// [ -1, -1, 7, zz]
7424 /// PSLLDQ : (little-endian) left byte shift
7425 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7426 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7427 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7428 /// PSRLDQ : (little-endian) right byte shift
7429 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7430 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7431 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7432 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7433 SDValue V2, ArrayRef<int> Mask,
7434 SelectionDAG &DAG) {
7435 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7437 int Size = Mask.size();
7438 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7440 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7441 for (int i = 0; i < Size; i += Scale)
7442 for (int j = 0; j < Shift; ++j)
7443 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7449 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7450 for (int i = 0; i != Size; i += Scale) {
7451 unsigned Pos = Left ? i + Shift : i;
7452 unsigned Low = Left ? i : i + Shift;
7453 unsigned Len = Scale - Shift;
7454 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7455 Low + (V == V1 ? 0 : Size)))
7459 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7460 bool ByteShift = ShiftEltBits > 64;
7461 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7462 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7463 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7465 // Normalize the scale for byte shifts to still produce an i64 element
7467 Scale = ByteShift ? Scale / 2 : Scale;
7469 // We need to round trip through the appropriate type for the shift.
7470 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7471 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7472 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7473 "Illegal integer vector type");
7474 V = DAG.getBitcast(ShiftVT, V);
7476 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7477 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7478 return DAG.getBitcast(VT, V);
7481 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7482 // keep doubling the size of the integer elements up to that. We can
7483 // then shift the elements of the integer vector by whole multiples of
7484 // their width within the elements of the larger integer vector. Test each
7485 // multiple to see if we can find a match with the moved element indices
7486 // and that the shifted in elements are all zeroable.
7487 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7488 for (int Shift = 1; Shift != Scale; ++Shift)
7489 for (bool Left : {true, false})
7490 if (CheckZeros(Shift, Scale, Left))
7491 for (SDValue V : {V1, V2})
7492 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7499 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7500 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7501 SDValue V2, ArrayRef<int> Mask,
7502 SelectionDAG &DAG) {
7503 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7504 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7506 int Size = Mask.size();
7507 int HalfSize = Size / 2;
7508 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7510 // Upper half must be undefined.
7511 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7514 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7515 // Remainder of lower half result is zero and upper half is all undef.
7516 auto LowerAsEXTRQ = [&]() {
7517 // Determine the extraction length from the part of the
7518 // lower half that isn't zeroable.
7520 for (; Len > 0; --Len)
7521 if (!Zeroable[Len - 1])
7523 assert(Len > 0 && "Zeroable shuffle mask");
7525 // Attempt to match first Len sequential elements from the lower half.
7528 for (int i = 0; i != Len; ++i) {
7532 SDValue &V = (M < Size ? V1 : V2);
7535 // The extracted elements must start at a valid index and all mask
7536 // elements must be in the lower half.
7537 if (i > M || M >= HalfSize)
7540 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7551 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7552 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7553 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7554 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7555 DAG.getConstant(BitLen, DL, MVT::i8),
7556 DAG.getConstant(BitIdx, DL, MVT::i8));
7559 if (SDValue ExtrQ = LowerAsEXTRQ())
7562 // INSERTQ: Extract lowest Len elements from lower half of second source and
7563 // insert over first source, starting at Idx.
7564 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7565 auto LowerAsInsertQ = [&]() {
7566 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7569 // Attempt to match first source from mask before insertion point.
7570 if (isUndefInRange(Mask, 0, Idx)) {
7572 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7574 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7580 // Extend the extraction length looking to match both the insertion of
7581 // the second source and the remaining elements of the first.
7582 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7587 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7589 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7595 // Match the remaining elements of the lower half.
7596 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7598 } else if ((!Base || (Base == V1)) &&
7599 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7601 } else if ((!Base || (Base == V2)) &&
7602 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7609 // We may not have a base (first source) - this can safely be undefined.
7611 Base = DAG.getUNDEF(VT);
7613 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7614 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7615 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7616 DAG.getConstant(BitLen, DL, MVT::i8),
7617 DAG.getConstant(BitIdx, DL, MVT::i8));
7624 if (SDValue InsertQ = LowerAsInsertQ())
7630 /// \brief Lower a vector shuffle as a zero or any extension.
7632 /// Given a specific number of elements, element bit width, and extension
7633 /// stride, produce either a zero or any extension based on the available
7634 /// features of the subtarget. The extended elements are consecutive and
7635 /// begin and can start from an offseted element index in the input; to
7636 /// avoid excess shuffling the offset must either being in the bottom lane
7637 /// or at the start of a higher lane. All extended elements must be from
7639 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7640 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7641 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7642 assert(Scale > 1 && "Need a scale to extend.");
7643 int EltBits = VT.getScalarSizeInBits();
7644 int NumElements = VT.getVectorNumElements();
7645 int NumEltsPerLane = 128 / EltBits;
7646 int OffsetLane = Offset / NumEltsPerLane;
7647 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7648 "Only 8, 16, and 32 bit elements can be extended.");
7649 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7650 assert(0 <= Offset && "Extension offset must be positive.");
7651 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7652 "Extension offset must be in the first lane or start an upper lane.");
7654 // Check that an index is in same lane as the base offset.
7655 auto SafeOffset = [&](int Idx) {
7656 return OffsetLane == (Idx / NumEltsPerLane);
7659 // Shift along an input so that the offset base moves to the first element.
7660 auto ShuffleOffset = [&](SDValue V) {
7664 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7665 for (int i = 0; i * Scale < NumElements; ++i) {
7666 int SrcIdx = i + Offset;
7667 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7669 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7672 // Found a valid zext mask! Try various lowering strategies based on the
7673 // input type and available ISA extensions.
7674 if (Subtarget->hasSSE41()) {
7675 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7676 // PUNPCK will catch this in a later shuffle match.
7677 if (Offset && Scale == 2 && VT.is128BitVector())
7679 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7680 NumElements / Scale);
7681 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7682 return DAG.getBitcast(VT, InputV);
7685 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7687 // For any extends we can cheat for larger element sizes and use shuffle
7688 // instructions that can fold with a load and/or copy.
7689 if (AnyExt && EltBits == 32) {
7690 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7692 return DAG.getBitcast(
7693 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7694 DAG.getBitcast(MVT::v4i32, InputV),
7695 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7697 if (AnyExt && EltBits == 16 && Scale > 2) {
7698 int PSHUFDMask[4] = {Offset / 2, -1,
7699 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7700 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7701 DAG.getBitcast(MVT::v4i32, InputV),
7702 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7703 int PSHUFWMask[4] = {1, -1, -1, -1};
7704 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7705 return DAG.getBitcast(
7706 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7707 DAG.getBitcast(MVT::v8i16, InputV),
7708 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7711 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7713 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7714 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7715 assert(VT.is128BitVector() && "Unexpected vector width!");
7717 int LoIdx = Offset * EltBits;
7718 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7719 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7720 DAG.getConstant(EltBits, DL, MVT::i8),
7721 DAG.getConstant(LoIdx, DL, MVT::i8)));
7723 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7724 !SafeOffset(Offset + 1))
7725 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7727 int HiIdx = (Offset + 1) * EltBits;
7728 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7729 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7730 DAG.getConstant(EltBits, DL, MVT::i8),
7731 DAG.getConstant(HiIdx, DL, MVT::i8)));
7732 return DAG.getNode(ISD::BITCAST, DL, VT,
7733 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7736 // If this would require more than 2 unpack instructions to expand, use
7737 // pshufb when available. We can only use more than 2 unpack instructions
7738 // when zero extending i8 elements which also makes it easier to use pshufb.
7739 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7740 assert(NumElements == 16 && "Unexpected byte vector width!");
7741 SDValue PSHUFBMask[16];
7742 for (int i = 0; i < 16; ++i) {
7743 int Idx = Offset + (i / Scale);
7744 PSHUFBMask[i] = DAG.getConstant(
7745 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7747 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7748 return DAG.getBitcast(VT,
7749 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7750 DAG.getNode(ISD::BUILD_VECTOR, DL,
7751 MVT::v16i8, PSHUFBMask)));
7754 // If we are extending from an offset, ensure we start on a boundary that
7755 // we can unpack from.
7756 int AlignToUnpack = Offset % (NumElements / Scale);
7757 if (AlignToUnpack) {
7758 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7759 for (int i = AlignToUnpack; i < NumElements; ++i)
7760 ShMask[i - AlignToUnpack] = i;
7761 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7762 Offset -= AlignToUnpack;
7765 // Otherwise emit a sequence of unpacks.
7767 unsigned UnpackLoHi = X86ISD::UNPCKL;
7768 if (Offset >= (NumElements / 2)) {
7769 UnpackLoHi = X86ISD::UNPCKH;
7770 Offset -= (NumElements / 2);
7773 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7774 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7775 : getZeroVector(InputVT, Subtarget, DAG, DL);
7776 InputV = DAG.getBitcast(InputVT, InputV);
7777 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7781 } while (Scale > 1);
7782 return DAG.getBitcast(VT, InputV);
7785 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7787 /// This routine will try to do everything in its power to cleverly lower
7788 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7789 /// check for the profitability of this lowering, it tries to aggressively
7790 /// match this pattern. It will use all of the micro-architectural details it
7791 /// can to emit an efficient lowering. It handles both blends with all-zero
7792 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7793 /// masking out later).
7795 /// The reason we have dedicated lowering for zext-style shuffles is that they
7796 /// are both incredibly common and often quite performance sensitive.
7797 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7798 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7799 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7800 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7802 int Bits = VT.getSizeInBits();
7803 int NumLanes = Bits / 128;
7804 int NumElements = VT.getVectorNumElements();
7805 int NumEltsPerLane = NumElements / NumLanes;
7806 assert(VT.getScalarSizeInBits() <= 32 &&
7807 "Exceeds 32-bit integer zero extension limit");
7808 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7810 // Define a helper function to check a particular ext-scale and lower to it if
7812 auto Lower = [&](int Scale) -> SDValue {
7817 for (int i = 0; i < NumElements; ++i) {
7820 continue; // Valid anywhere but doesn't tell us anything.
7821 if (i % Scale != 0) {
7822 // Each of the extended elements need to be zeroable.
7826 // We no longer are in the anyext case.
7831 // Each of the base elements needs to be consecutive indices into the
7832 // same input vector.
7833 SDValue V = M < NumElements ? V1 : V2;
7834 M = M % NumElements;
7837 Offset = M - (i / Scale);
7838 } else if (InputV != V)
7839 return SDValue(); // Flip-flopping inputs.
7841 // Offset must start in the lowest 128-bit lane or at the start of an
7843 // FIXME: Is it ever worth allowing a negative base offset?
7844 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7845 (Offset % NumEltsPerLane) == 0))
7848 // If we are offsetting, all referenced entries must come from the same
7850 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7853 if ((M % NumElements) != (Offset + (i / Scale)))
7854 return SDValue(); // Non-consecutive strided elements.
7858 // If we fail to find an input, we have a zero-shuffle which should always
7859 // have already been handled.
7860 // FIXME: Maybe handle this here in case during blending we end up with one?
7864 // If we are offsetting, don't extend if we only match a single input, we
7865 // can always do better by using a basic PSHUF or PUNPCK.
7866 if (Offset != 0 && Matches < 2)
7869 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7870 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7873 // The widest scale possible for extending is to a 64-bit integer.
7874 assert(Bits % 64 == 0 &&
7875 "The number of bits in a vector must be divisible by 64 on x86!");
7876 int NumExtElements = Bits / 64;
7878 // Each iteration, try extending the elements half as much, but into twice as
7880 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7881 assert(NumElements % NumExtElements == 0 &&
7882 "The input vector size must be divisible by the extended size.");
7883 if (SDValue V = Lower(NumElements / NumExtElements))
7887 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7891 // Returns one of the source operands if the shuffle can be reduced to a
7892 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7893 auto CanZExtLowHalf = [&]() {
7894 for (int i = NumElements / 2; i != NumElements; ++i)
7897 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7899 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7904 if (SDValue V = CanZExtLowHalf()) {
7905 V = DAG.getBitcast(MVT::v2i64, V);
7906 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7907 return DAG.getBitcast(VT, V);
7910 // No viable ext lowering found.
7914 /// \brief Try to get a scalar value for a specific element of a vector.
7916 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7917 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7918 SelectionDAG &DAG) {
7919 MVT VT = V.getSimpleValueType();
7920 MVT EltVT = VT.getVectorElementType();
7921 while (V.getOpcode() == ISD::BITCAST)
7922 V = V.getOperand(0);
7923 // If the bitcasts shift the element size, we can't extract an equivalent
7925 MVT NewVT = V.getSimpleValueType();
7926 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7929 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7930 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7931 // Ensure the scalar operand is the same size as the destination.
7932 // FIXME: Add support for scalar truncation where possible.
7933 SDValue S = V.getOperand(Idx);
7934 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7935 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7941 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7943 /// This is particularly important because the set of instructions varies
7944 /// significantly based on whether the operand is a load or not.
7945 static bool isShuffleFoldableLoad(SDValue V) {
7946 while (V.getOpcode() == ISD::BITCAST)
7947 V = V.getOperand(0);
7949 return ISD::isNON_EXTLoad(V.getNode());
7952 /// \brief Try to lower insertion of a single element into a zero vector.
7954 /// This is a common pattern that we have especially efficient patterns to lower
7955 /// across all subtarget feature sets.
7956 static SDValue lowerVectorShuffleAsElementInsertion(
7957 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7958 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7959 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7961 MVT EltVT = VT.getVectorElementType();
7963 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7964 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7966 bool IsV1Zeroable = true;
7967 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7968 if (i != V2Index && !Zeroable[i]) {
7969 IsV1Zeroable = false;
7973 // Check for a single input from a SCALAR_TO_VECTOR node.
7974 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7975 // all the smarts here sunk into that routine. However, the current
7976 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7977 // vector shuffle lowering is dead.
7978 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7980 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7981 // We need to zext the scalar if it is smaller than an i32.
7982 V2S = DAG.getBitcast(EltVT, V2S);
7983 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7984 // Using zext to expand a narrow element won't work for non-zero
7989 // Zero-extend directly to i32.
7991 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7993 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7994 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7995 EltVT == MVT::i16) {
7996 // Either not inserting from the low element of the input or the input
7997 // element size is too small to use VZEXT_MOVL to clear the high bits.
8001 if (!IsV1Zeroable) {
8002 // If V1 can't be treated as a zero vector we have fewer options to lower
8003 // this. We can't support integer vectors or non-zero targets cheaply, and
8004 // the V1 elements can't be permuted in any way.
8005 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8006 if (!VT.isFloatingPoint() || V2Index != 0)
8008 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8009 V1Mask[V2Index] = -1;
8010 if (!isNoopShuffleMask(V1Mask))
8012 // This is essentially a special case blend operation, but if we have
8013 // general purpose blend operations, they are always faster. Bail and let
8014 // the rest of the lowering handle these as blends.
8015 if (Subtarget->hasSSE41())
8018 // Otherwise, use MOVSD or MOVSS.
8019 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8020 "Only two types of floating point element types to handle!");
8021 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8025 // This lowering only works for the low element with floating point vectors.
8026 if (VT.isFloatingPoint() && V2Index != 0)
8029 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8031 V2 = DAG.getBitcast(VT, V2);
8034 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8035 // the desired position. Otherwise it is more efficient to do a vector
8036 // shift left. We know that we can do a vector shift left because all
8037 // the inputs are zero.
8038 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8039 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8040 V2Shuffle[V2Index] = 0;
8041 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8043 V2 = DAG.getBitcast(MVT::v2i64, V2);
8045 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8046 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
8047 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
8048 DAG.getDataLayout(), VT)));
8049 V2 = DAG.getBitcast(VT, V2);
8055 /// \brief Try to lower broadcast of a single - truncated - integer element,
8056 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
8058 /// This assumes we have AVX2.
8059 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
8061 const X86Subtarget *Subtarget,
8062 SelectionDAG &DAG) {
8063 assert(Subtarget->hasAVX2() &&
8064 "We can only lower integer broadcasts with AVX2!");
8066 EVT EltVT = VT.getVectorElementType();
8067 EVT V0VT = V0.getValueType();
8069 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
8070 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
8072 EVT V0EltVT = V0VT.getVectorElementType();
8073 if (!V0EltVT.isInteger())
8076 const unsigned EltSize = EltVT.getSizeInBits();
8077 const unsigned V0EltSize = V0EltVT.getSizeInBits();
8079 // This is only a truncation if the original element type is larger.
8080 if (V0EltSize <= EltSize)
8083 assert(((V0EltSize % EltSize) == 0) &&
8084 "Scalar type sizes must all be powers of 2 on x86!");
8086 const unsigned V0Opc = V0.getOpcode();
8087 const unsigned Scale = V0EltSize / EltSize;
8088 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
8090 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
8091 V0Opc != ISD::BUILD_VECTOR)
8094 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
8096 // If we're extracting non-least-significant bits, shift so we can truncate.
8097 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
8098 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
8099 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
8100 if (const int OffsetIdx = BroadcastIdx % Scale)
8101 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
8102 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
8104 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
8105 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
8108 /// \brief Try to lower broadcast of a single element.
8110 /// For convenience, this code also bundles all of the subtarget feature set
8111 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8112 /// a convenient way to factor it out.
8113 /// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
8114 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
8116 const X86Subtarget *Subtarget,
8117 SelectionDAG &DAG) {
8118 if (!Subtarget->hasAVX())
8120 if (VT.isInteger() && !Subtarget->hasAVX2())
8123 // Check that the mask is a broadcast.
8124 int BroadcastIdx = -1;
8126 if (M >= 0 && BroadcastIdx == -1)
8128 else if (M >= 0 && M != BroadcastIdx)
8131 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8132 "a sorted mask where the broadcast "
8135 // Go up the chain of (vector) values to find a scalar load that we can
8136 // combine with the broadcast.
8138 switch (V.getOpcode()) {
8139 case ISD::CONCAT_VECTORS: {
8140 int OperandSize = Mask.size() / V.getNumOperands();
8141 V = V.getOperand(BroadcastIdx / OperandSize);
8142 BroadcastIdx %= OperandSize;
8146 case ISD::INSERT_SUBVECTOR: {
8147 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8148 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8152 int BeginIdx = (int)ConstantIdx->getZExtValue();
8154 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
8155 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8156 BroadcastIdx -= BeginIdx;
8167 // Check if this is a broadcast of a scalar. We special case lowering
8168 // for scalars so that we can more effectively fold with loads.
8169 // First, look through bitcast: if the original value has a larger element
8170 // type than the shuffle, the broadcast element is in essence truncated.
8171 // Make that explicit to ease folding.
8172 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
8173 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
8174 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
8175 return TruncBroadcast;
8177 MVT BroadcastVT = VT;
8179 // Peek through any bitcast (only useful for loads).
8181 while (BC.getOpcode() == ISD::BITCAST)
8182 BC = BC.getOperand(0);
8184 // Also check the simpler case, where we can directly reuse the scalar.
8185 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8186 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8187 V = V.getOperand(BroadcastIdx);
8189 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8190 // Only AVX2 has register broadcasts.
8191 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8193 } else if (MayFoldLoad(BC) && !cast<LoadSDNode>(BC)->isVolatile()) {
8194 // 32-bit targets need to load i64 as a f64 and then bitcast the result.
8195 if (!Subtarget->is64Bit() && VT.getScalarType() == MVT::i64)
8196 BroadcastVT = MVT::getVectorVT(MVT::f64, VT.getVectorNumElements());
8198 // If we are broadcasting a load that is only used by the shuffle
8199 // then we can reduce the vector load to the broadcasted scalar load.
8200 LoadSDNode *Ld = cast<LoadSDNode>(BC);
8201 SDValue BaseAddr = Ld->getOperand(1);
8202 EVT AddrVT = BaseAddr.getValueType();
8203 EVT SVT = BroadcastVT.getScalarType();
8204 unsigned Offset = BroadcastIdx * SVT.getStoreSize();
8205 SDValue NewAddr = DAG.getNode(
8206 ISD::ADD, DL, AddrVT, BaseAddr,
8207 DAG.getConstant(Offset, DL, AddrVT));
8208 V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
8209 DAG.getMachineFunction().getMachineMemOperand(
8210 Ld->getMemOperand(), Offset, SVT.getStoreSize()));
8211 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8212 // We can't broadcast from a vector register without AVX2, and we can only
8213 // broadcast from the zero-element of a vector register.
8217 V = DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, V);
8218 return DAG.getBitcast(VT, V);
8221 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8222 // INSERTPS when the V1 elements are already in the correct locations
8223 // because otherwise we can just always use two SHUFPS instructions which
8224 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8225 // perform INSERTPS if a single V1 element is out of place and all V2
8226 // elements are zeroable.
8227 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8229 SelectionDAG &DAG) {
8230 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8231 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8232 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8233 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8235 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8238 int V1DstIndex = -1;
8239 int V2DstIndex = -1;
8240 bool V1UsedInPlace = false;
8242 for (int i = 0; i < 4; ++i) {
8243 // Synthesize a zero mask from the zeroable elements (includes undefs).
8249 // Flag if we use any V1 inputs in place.
8251 V1UsedInPlace = true;
8255 // We can only insert a single non-zeroable element.
8256 if (V1DstIndex != -1 || V2DstIndex != -1)
8260 // V1 input out of place for insertion.
8263 // V2 input for insertion.
8268 // Don't bother if we have no (non-zeroable) element for insertion.
8269 if (V1DstIndex == -1 && V2DstIndex == -1)
8272 // Determine element insertion src/dst indices. The src index is from the
8273 // start of the inserted vector, not the start of the concatenated vector.
8274 unsigned V2SrcIndex = 0;
8275 if (V1DstIndex != -1) {
8276 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8277 // and don't use the original V2 at all.
8278 V2SrcIndex = Mask[V1DstIndex];
8279 V2DstIndex = V1DstIndex;
8282 V2SrcIndex = Mask[V2DstIndex] - 4;
8285 // If no V1 inputs are used in place, then the result is created only from
8286 // the zero mask and the V2 insertion - so remove V1 dependency.
8288 V1 = DAG.getUNDEF(MVT::v4f32);
8290 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8291 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8293 // Insert the V2 element into the desired position.
8295 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8296 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8299 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8300 /// UNPCK instruction.
8302 /// This specifically targets cases where we end up with alternating between
8303 /// the two inputs, and so can permute them into something that feeds a single
8304 /// UNPCK instruction. Note that this routine only targets integer vectors
8305 /// because for floating point vectors we have a generalized SHUFPS lowering
8306 /// strategy that handles everything that doesn't *exactly* match an unpack,
8307 /// making this clever lowering unnecessary.
8308 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8309 SDValue V1, SDValue V2,
8311 SelectionDAG &DAG) {
8312 assert(!VT.isFloatingPoint() &&
8313 "This routine only supports integer vectors.");
8314 assert(!isSingleInputShuffleMask(Mask) &&
8315 "This routine should only be used when blending two inputs.");
8316 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8318 int Size = Mask.size();
8320 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8321 return M >= 0 && M % Size < Size / 2;
8323 int NumHiInputs = std::count_if(
8324 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8326 bool UnpackLo = NumLoInputs >= NumHiInputs;
8328 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8329 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8330 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8332 for (int i = 0; i < Size; ++i) {
8336 // Each element of the unpack contains Scale elements from this mask.
8337 int UnpackIdx = i / Scale;
8339 // We only handle the case where V1 feeds the first slots of the unpack.
8340 // We rely on canonicalization to ensure this is the case.
8341 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8344 // Setup the mask for this input. The indexing is tricky as we have to
8345 // handle the unpack stride.
8346 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8347 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8351 // If we will have to shuffle both inputs to use the unpack, check whether
8352 // we can just unpack first and shuffle the result. If so, skip this unpack.
8353 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8354 !isNoopShuffleMask(V2Mask))
8357 // Shuffle the inputs into place.
8358 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8359 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8361 // Cast the inputs to the type we will use to unpack them.
8362 V1 = DAG.getBitcast(UnpackVT, V1);
8363 V2 = DAG.getBitcast(UnpackVT, V2);
8365 // Unpack the inputs and cast the result back to the desired type.
8366 return DAG.getBitcast(
8367 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8371 // We try each unpack from the largest to the smallest to try and find one
8372 // that fits this mask.
8373 int OrigNumElements = VT.getVectorNumElements();
8374 int OrigScalarSize = VT.getScalarSizeInBits();
8375 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8376 int Scale = ScalarSize / OrigScalarSize;
8377 int NumElements = OrigNumElements / Scale;
8378 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8379 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8383 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8385 if (NumLoInputs == 0 || NumHiInputs == 0) {
8386 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8387 "We have to have *some* inputs!");
8388 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8390 // FIXME: We could consider the total complexity of the permute of each
8391 // possible unpacking. Or at the least we should consider how many
8392 // half-crossings are created.
8393 // FIXME: We could consider commuting the unpacks.
8395 SmallVector<int, 32> PermMask;
8396 PermMask.assign(Size, -1);
8397 for (int i = 0; i < Size; ++i) {
8401 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8404 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8406 return DAG.getVectorShuffle(
8407 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8409 DAG.getUNDEF(VT), PermMask);
8415 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8417 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8418 /// support for floating point shuffles but not integer shuffles. These
8419 /// instructions will incur a domain crossing penalty on some chips though so
8420 /// it is better to avoid lowering through this for integer vectors where
8422 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8423 const X86Subtarget *Subtarget,
8424 SelectionDAG &DAG) {
8426 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8427 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8428 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8429 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8430 ArrayRef<int> Mask = SVOp->getMask();
8431 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8433 if (isSingleInputShuffleMask(Mask)) {
8434 // Use low duplicate instructions for masks that match their pattern.
8435 if (Subtarget->hasSSE3())
8436 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8437 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8439 // Straight shuffle of a single input vector. Simulate this by using the
8440 // single input as both of the "inputs" to this instruction..
8441 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8443 if (Subtarget->hasAVX()) {
8444 // If we have AVX, we can use VPERMILPS which will allow folding a load
8445 // into the shuffle.
8446 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8447 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8450 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8451 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8453 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8454 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8456 // If we have a single input, insert that into V1 if we can do so cheaply.
8457 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8458 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8459 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8461 // Try inverting the insertion since for v2 masks it is easy to do and we
8462 // can't reliably sort the mask one way or the other.
8463 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8464 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8465 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8466 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8470 // Try to use one of the special instruction patterns to handle two common
8471 // blend patterns if a zero-blend above didn't work.
8472 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8473 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8474 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8475 // We can either use a special instruction to load over the low double or
8476 // to move just the low double.
8478 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8480 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8482 if (Subtarget->hasSSE41())
8483 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8487 // Use dedicated unpack instructions for masks that match their pattern.
8489 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8492 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8493 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8494 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8497 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8499 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8500 /// the integer unit to minimize domain crossing penalties. However, for blends
8501 /// it falls back to the floating point shuffle operation with appropriate bit
8503 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8504 const X86Subtarget *Subtarget,
8505 SelectionDAG &DAG) {
8507 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8508 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8509 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8510 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8511 ArrayRef<int> Mask = SVOp->getMask();
8512 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8514 if (isSingleInputShuffleMask(Mask)) {
8515 // Check for being able to broadcast a single element.
8516 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8517 Mask, Subtarget, DAG))
8520 // Straight shuffle of a single input vector. For everything from SSE2
8521 // onward this has a single fast instruction with no scary immediates.
8522 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8523 V1 = DAG.getBitcast(MVT::v4i32, V1);
8524 int WidenedMask[4] = {
8525 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8526 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8527 return DAG.getBitcast(
8529 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8530 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8532 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8533 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8534 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8535 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8537 // If we have a blend of two PACKUS operations an the blend aligns with the
8538 // low and half halves, we can just merge the PACKUS operations. This is
8539 // particularly important as it lets us merge shuffles that this routine itself
8541 auto GetPackNode = [](SDValue V) {
8542 while (V.getOpcode() == ISD::BITCAST)
8543 V = V.getOperand(0);
8545 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8547 if (SDValue V1Pack = GetPackNode(V1))
8548 if (SDValue V2Pack = GetPackNode(V2))
8549 return DAG.getBitcast(MVT::v2i64,
8550 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8551 Mask[0] == 0 ? V1Pack.getOperand(0)
8552 : V1Pack.getOperand(1),
8553 Mask[1] == 2 ? V2Pack.getOperand(0)
8554 : V2Pack.getOperand(1)));
8556 // Try to use shift instructions.
8558 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8561 // When loading a scalar and then shuffling it into a vector we can often do
8562 // the insertion cheaply.
8563 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8564 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8566 // Try inverting the insertion since for v2 masks it is easy to do and we
8567 // can't reliably sort the mask one way or the other.
8568 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8569 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8570 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8573 // We have different paths for blend lowering, but they all must use the
8574 // *exact* same predicate.
8575 bool IsBlendSupported = Subtarget->hasSSE41();
8576 if (IsBlendSupported)
8577 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8581 // Use dedicated unpack instructions for masks that match their pattern.
8583 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8586 // Try to use byte rotation instructions.
8587 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8588 if (Subtarget->hasSSSE3())
8589 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8590 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8593 // If we have direct support for blends, we should lower by decomposing into
8594 // a permute. That will be faster than the domain cross.
8595 if (IsBlendSupported)
8596 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8599 // We implement this with SHUFPD which is pretty lame because it will likely
8600 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8601 // However, all the alternatives are still more cycles and newer chips don't
8602 // have this problem. It would be really nice if x86 had better shuffles here.
8603 V1 = DAG.getBitcast(MVT::v2f64, V1);
8604 V2 = DAG.getBitcast(MVT::v2f64, V2);
8605 return DAG.getBitcast(MVT::v2i64,
8606 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8609 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8611 /// This is used to disable more specialized lowerings when the shufps lowering
8612 /// will happen to be efficient.
8613 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8614 // This routine only handles 128-bit shufps.
8615 assert(Mask.size() == 4 && "Unsupported mask size!");
8617 // To lower with a single SHUFPS we need to have the low half and high half
8618 // each requiring a single input.
8619 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8621 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8627 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8629 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8630 /// It makes no assumptions about whether this is the *best* lowering, it simply
8632 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8633 ArrayRef<int> Mask, SDValue V1,
8634 SDValue V2, SelectionDAG &DAG) {
8635 SDValue LowV = V1, HighV = V2;
8636 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8639 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8641 if (NumV2Elements == 1) {
8643 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8646 // Compute the index adjacent to V2Index and in the same half by toggling
8648 int V2AdjIndex = V2Index ^ 1;
8650 if (Mask[V2AdjIndex] == -1) {
8651 // Handles all the cases where we have a single V2 element and an undef.
8652 // This will only ever happen in the high lanes because we commute the
8653 // vector otherwise.
8655 std::swap(LowV, HighV);
8656 NewMask[V2Index] -= 4;
8658 // Handle the case where the V2 element ends up adjacent to a V1 element.
8659 // To make this work, blend them together as the first step.
8660 int V1Index = V2AdjIndex;
8661 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8662 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8663 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8665 // Now proceed to reconstruct the final blend as we have the necessary
8666 // high or low half formed.
8673 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8674 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8676 } else if (NumV2Elements == 2) {
8677 if (Mask[0] < 4 && Mask[1] < 4) {
8678 // Handle the easy case where we have V1 in the low lanes and V2 in the
8682 } else if (Mask[2] < 4 && Mask[3] < 4) {
8683 // We also handle the reversed case because this utility may get called
8684 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8685 // arrange things in the right direction.
8691 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8692 // trying to place elements directly, just blend them and set up the final
8693 // shuffle to place them.
8695 // The first two blend mask elements are for V1, the second two are for
8697 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8698 Mask[2] < 4 ? Mask[2] : Mask[3],
8699 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8700 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8701 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8702 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8704 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8707 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8708 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8709 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8710 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8713 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8714 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8717 /// \brief Lower 4-lane 32-bit floating point shuffles.
8719 /// Uses instructions exclusively from the floating point unit to minimize
8720 /// domain crossing penalties, as these are sufficient to implement all v4f32
8722 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8723 const X86Subtarget *Subtarget,
8724 SelectionDAG &DAG) {
8726 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8727 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8728 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8729 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8730 ArrayRef<int> Mask = SVOp->getMask();
8731 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8734 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8736 if (NumV2Elements == 0) {
8737 // Check for being able to broadcast a single element.
8738 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8739 Mask, Subtarget, DAG))
8742 // Use even/odd duplicate instructions for masks that match their pattern.
8743 if (Subtarget->hasSSE3()) {
8744 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8745 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8746 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8747 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8750 if (Subtarget->hasAVX()) {
8751 // If we have AVX, we can use VPERMILPS which will allow folding a load
8752 // into the shuffle.
8753 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8754 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8757 // Otherwise, use a straight shuffle of a single input vector. We pass the
8758 // input vector to both operands to simulate this with a SHUFPS.
8759 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8760 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8763 // There are special ways we can lower some single-element blends. However, we
8764 // have custom ways we can lower more complex single-element blends below that
8765 // we defer to if both this and BLENDPS fail to match, so restrict this to
8766 // when the V2 input is targeting element 0 of the mask -- that is the fast
8768 if (NumV2Elements == 1 && Mask[0] >= 4)
8769 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8770 Mask, Subtarget, DAG))
8773 if (Subtarget->hasSSE41()) {
8774 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8778 // Use INSERTPS if we can complete the shuffle efficiently.
8779 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8782 if (!isSingleSHUFPSMask(Mask))
8783 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8784 DL, MVT::v4f32, V1, V2, Mask, DAG))
8788 // Use dedicated unpack instructions for masks that match their pattern.
8790 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8793 // Otherwise fall back to a SHUFPS lowering strategy.
8794 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8797 /// \brief Lower 4-lane i32 vector shuffles.
8799 /// We try to handle these with integer-domain shuffles where we can, but for
8800 /// blends we use the floating point domain blend instructions.
8801 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8802 const X86Subtarget *Subtarget,
8803 SelectionDAG &DAG) {
8805 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8806 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8807 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8808 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8809 ArrayRef<int> Mask = SVOp->getMask();
8810 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8812 // Whenever we can lower this as a zext, that instruction is strictly faster
8813 // than any alternative. It also allows us to fold memory operands into the
8814 // shuffle in many cases.
8815 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8816 Mask, Subtarget, DAG))
8820 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8822 if (NumV2Elements == 0) {
8823 // Check for being able to broadcast a single element.
8824 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8825 Mask, Subtarget, DAG))
8828 // Straight shuffle of a single input vector. For everything from SSE2
8829 // onward this has a single fast instruction with no scary immediates.
8830 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8831 // but we aren't actually going to use the UNPCK instruction because doing
8832 // so prevents folding a load into this instruction or making a copy.
8833 const int UnpackLoMask[] = {0, 0, 1, 1};
8834 const int UnpackHiMask[] = {2, 2, 3, 3};
8835 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8836 Mask = UnpackLoMask;
8837 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8838 Mask = UnpackHiMask;
8840 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8841 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8844 // Try to use shift instructions.
8846 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8849 // There are special ways we can lower some single-element blends.
8850 if (NumV2Elements == 1)
8851 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8852 Mask, Subtarget, DAG))
8855 // We have different paths for blend lowering, but they all must use the
8856 // *exact* same predicate.
8857 bool IsBlendSupported = Subtarget->hasSSE41();
8858 if (IsBlendSupported)
8859 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8863 if (SDValue Masked =
8864 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8867 // Use dedicated unpack instructions for masks that match their pattern.
8869 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8872 // Try to use byte rotation instructions.
8873 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8874 if (Subtarget->hasSSSE3())
8875 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8876 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8879 // If we have direct support for blends, we should lower by decomposing into
8880 // a permute. That will be faster than the domain cross.
8881 if (IsBlendSupported)
8882 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8885 // Try to lower by permuting the inputs into an unpack instruction.
8886 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8890 // We implement this with SHUFPS because it can blend from two vectors.
8891 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8892 // up the inputs, bypassing domain shift penalties that we would encur if we
8893 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8895 return DAG.getBitcast(
8897 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8898 DAG.getBitcast(MVT::v4f32, V2), Mask));
8901 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8902 /// shuffle lowering, and the most complex part.
8904 /// The lowering strategy is to try to form pairs of input lanes which are
8905 /// targeted at the same half of the final vector, and then use a dword shuffle
8906 /// to place them onto the right half, and finally unpack the paired lanes into
8907 /// their final position.
8909 /// The exact breakdown of how to form these dword pairs and align them on the
8910 /// correct sides is really tricky. See the comments within the function for
8911 /// more of the details.
8913 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8914 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8915 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8916 /// vector, form the analogous 128-bit 8-element Mask.
8917 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8918 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8919 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8920 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8921 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8923 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8924 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8925 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8927 SmallVector<int, 4> LoInputs;
8928 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8929 [](int M) { return M >= 0; });
8930 std::sort(LoInputs.begin(), LoInputs.end());
8931 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8932 SmallVector<int, 4> HiInputs;
8933 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8934 [](int M) { return M >= 0; });
8935 std::sort(HiInputs.begin(), HiInputs.end());
8936 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8938 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8939 int NumHToL = LoInputs.size() - NumLToL;
8941 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8942 int NumHToH = HiInputs.size() - NumLToH;
8943 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8944 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8945 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8946 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8948 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8949 // such inputs we can swap two of the dwords across the half mark and end up
8950 // with <=2 inputs to each half in each half. Once there, we can fall through
8951 // to the generic code below. For example:
8953 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8954 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8956 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8957 // and an existing 2-into-2 on the other half. In this case we may have to
8958 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8959 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8960 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8961 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8962 // half than the one we target for fixing) will be fixed when we re-enter this
8963 // path. We will also combine away any sequence of PSHUFD instructions that
8964 // result into a single instruction. Here is an example of the tricky case:
8966 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8967 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8969 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8971 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8972 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8974 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8975 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8977 // The result is fine to be handled by the generic logic.
8978 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8979 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8980 int AOffset, int BOffset) {
8981 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8982 "Must call this with A having 3 or 1 inputs from the A half.");
8983 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8984 "Must call this with B having 1 or 3 inputs from the B half.");
8985 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8986 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8988 bool ThreeAInputs = AToAInputs.size() == 3;
8990 // Compute the index of dword with only one word among the three inputs in
8991 // a half by taking the sum of the half with three inputs and subtracting
8992 // the sum of the actual three inputs. The difference is the remaining
8995 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8996 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8997 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8998 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8999 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
9000 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
9001 int TripleNonInputIdx =
9002 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
9003 TripleDWord = TripleNonInputIdx / 2;
9005 // We use xor with one to compute the adjacent DWord to whichever one the
9007 OneInputDWord = (OneInput / 2) ^ 1;
9009 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
9010 // and BToA inputs. If there is also such a problem with the BToB and AToB
9011 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
9012 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
9013 // is essential that we don't *create* a 3<-1 as then we might oscillate.
9014 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
9015 // Compute how many inputs will be flipped by swapping these DWords. We
9017 // to balance this to ensure we don't form a 3-1 shuffle in the other
9019 int NumFlippedAToBInputs =
9020 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
9021 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
9022 int NumFlippedBToBInputs =
9023 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
9024 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
9025 if ((NumFlippedAToBInputs == 1 &&
9026 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
9027 (NumFlippedBToBInputs == 1 &&
9028 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
9029 // We choose whether to fix the A half or B half based on whether that
9030 // half has zero flipped inputs. At zero, we may not be able to fix it
9031 // with that half. We also bias towards fixing the B half because that
9032 // will more commonly be the high half, and we have to bias one way.
9033 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
9034 ArrayRef<int> Inputs) {
9035 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
9036 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
9037 PinnedIdx ^ 1) != Inputs.end();
9038 // Determine whether the free index is in the flipped dword or the
9039 // unflipped dword based on where the pinned index is. We use this bit
9040 // in an xor to conditionally select the adjacent dword.
9041 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
9042 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9043 FixFreeIdx) != Inputs.end();
9044 if (IsFixIdxInput == IsFixFreeIdxInput)
9046 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9047 FixFreeIdx) != Inputs.end();
9048 assert(IsFixIdxInput != IsFixFreeIdxInput &&
9049 "We need to be changing the number of flipped inputs!");
9050 int PSHUFHalfMask[] = {0, 1, 2, 3};
9051 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
9052 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
9054 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
9057 if (M != -1 && M == FixIdx)
9059 else if (M != -1 && M == FixFreeIdx)
9062 if (NumFlippedBToBInputs != 0) {
9064 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9065 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
9067 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
9068 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
9069 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9074 int PSHUFDMask[] = {0, 1, 2, 3};
9075 PSHUFDMask[ADWord] = BDWord;
9076 PSHUFDMask[BDWord] = ADWord;
9079 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9080 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9082 // Adjust the mask to match the new locations of A and B.
9084 if (M != -1 && M/2 == ADWord)
9085 M = 2 * BDWord + M % 2;
9086 else if (M != -1 && M/2 == BDWord)
9087 M = 2 * ADWord + M % 2;
9089 // Recurse back into this routine to re-compute state now that this isn't
9090 // a 3 and 1 problem.
9091 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
9094 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9095 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9096 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9097 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9099 // At this point there are at most two inputs to the low and high halves from
9100 // each half. That means the inputs can always be grouped into dwords and
9101 // those dwords can then be moved to the correct half with a dword shuffle.
9102 // We use at most one low and one high word shuffle to collect these paired
9103 // inputs into dwords, and finally a dword shuffle to place them.
9104 int PSHUFLMask[4] = {-1, -1, -1, -1};
9105 int PSHUFHMask[4] = {-1, -1, -1, -1};
9106 int PSHUFDMask[4] = {-1, -1, -1, -1};
9108 // First fix the masks for all the inputs that are staying in their
9109 // original halves. This will then dictate the targets of the cross-half
9111 auto fixInPlaceInputs =
9112 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9113 MutableArrayRef<int> SourceHalfMask,
9114 MutableArrayRef<int> HalfMask, int HalfOffset) {
9115 if (InPlaceInputs.empty())
9117 if (InPlaceInputs.size() == 1) {
9118 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9119 InPlaceInputs[0] - HalfOffset;
9120 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9123 if (IncomingInputs.empty()) {
9124 // Just fix all of the in place inputs.
9125 for (int Input : InPlaceInputs) {
9126 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9127 PSHUFDMask[Input / 2] = Input / 2;
9132 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9133 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9134 InPlaceInputs[0] - HalfOffset;
9135 // Put the second input next to the first so that they are packed into
9136 // a dword. We find the adjacent index by toggling the low bit.
9137 int AdjIndex = InPlaceInputs[0] ^ 1;
9138 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9139 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9140 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9142 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9143 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9145 // Now gather the cross-half inputs and place them into a free dword of
9146 // their target half.
9147 // FIXME: This operation could almost certainly be simplified dramatically to
9148 // look more like the 3-1 fixing operation.
9149 auto moveInputsToRightHalf = [&PSHUFDMask](
9150 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9151 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9152 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9154 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9155 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9157 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9159 int LowWord = Word & ~1;
9160 int HighWord = Word | 1;
9161 return isWordClobbered(SourceHalfMask, LowWord) ||
9162 isWordClobbered(SourceHalfMask, HighWord);
9165 if (IncomingInputs.empty())
9168 if (ExistingInputs.empty()) {
9169 // Map any dwords with inputs from them into the right half.
9170 for (int Input : IncomingInputs) {
9171 // If the source half mask maps over the inputs, turn those into
9172 // swaps and use the swapped lane.
9173 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9174 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9175 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9176 Input - SourceOffset;
9177 // We have to swap the uses in our half mask in one sweep.
9178 for (int &M : HalfMask)
9179 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9181 else if (M == Input)
9182 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9184 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9185 Input - SourceOffset &&
9186 "Previous placement doesn't match!");
9188 // Note that this correctly re-maps both when we do a swap and when
9189 // we observe the other side of the swap above. We rely on that to
9190 // avoid swapping the members of the input list directly.
9191 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9194 // Map the input's dword into the correct half.
9195 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9196 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9198 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9200 "Previous placement doesn't match!");
9203 // And just directly shift any other-half mask elements to be same-half
9204 // as we will have mirrored the dword containing the element into the
9205 // same position within that half.
9206 for (int &M : HalfMask)
9207 if (M >= SourceOffset && M < SourceOffset + 4) {
9208 M = M - SourceOffset + DestOffset;
9209 assert(M >= 0 && "This should never wrap below zero!");
9214 // Ensure we have the input in a viable dword of its current half. This
9215 // is particularly tricky because the original position may be clobbered
9216 // by inputs being moved and *staying* in that half.
9217 if (IncomingInputs.size() == 1) {
9218 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9219 int InputFixed = std::find(std::begin(SourceHalfMask),
9220 std::end(SourceHalfMask), -1) -
9221 std::begin(SourceHalfMask) + SourceOffset;
9222 SourceHalfMask[InputFixed - SourceOffset] =
9223 IncomingInputs[0] - SourceOffset;
9224 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9226 IncomingInputs[0] = InputFixed;
9228 } else if (IncomingInputs.size() == 2) {
9229 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9230 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9231 // We have two non-adjacent or clobbered inputs we need to extract from
9232 // the source half. To do this, we need to map them into some adjacent
9233 // dword slot in the source mask.
9234 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9235 IncomingInputs[1] - SourceOffset};
9237 // If there is a free slot in the source half mask adjacent to one of
9238 // the inputs, place the other input in it. We use (Index XOR 1) to
9239 // compute an adjacent index.
9240 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9241 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9242 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9243 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9244 InputsFixed[1] = InputsFixed[0] ^ 1;
9245 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9246 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9247 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9248 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9249 InputsFixed[0] = InputsFixed[1] ^ 1;
9250 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9251 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9252 // The two inputs are in the same DWord but it is clobbered and the
9253 // adjacent DWord isn't used at all. Move both inputs to the free
9255 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9256 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9257 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9258 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9260 // The only way we hit this point is if there is no clobbering
9261 // (because there are no off-half inputs to this half) and there is no
9262 // free slot adjacent to one of the inputs. In this case, we have to
9263 // swap an input with a non-input.
9264 for (int i = 0; i < 4; ++i)
9265 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9266 "We can't handle any clobbers here!");
9267 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9268 "Cannot have adjacent inputs here!");
9270 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9271 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9273 // We also have to update the final source mask in this case because
9274 // it may need to undo the above swap.
9275 for (int &M : FinalSourceHalfMask)
9276 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9277 M = InputsFixed[1] + SourceOffset;
9278 else if (M == InputsFixed[1] + SourceOffset)
9279 M = (InputsFixed[0] ^ 1) + SourceOffset;
9281 InputsFixed[1] = InputsFixed[0] ^ 1;
9284 // Point everything at the fixed inputs.
9285 for (int &M : HalfMask)
9286 if (M == IncomingInputs[0])
9287 M = InputsFixed[0] + SourceOffset;
9288 else if (M == IncomingInputs[1])
9289 M = InputsFixed[1] + SourceOffset;
9291 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9292 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9295 llvm_unreachable("Unhandled input size!");
9298 // Now hoist the DWord down to the right half.
9299 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9300 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9301 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9302 for (int &M : HalfMask)
9303 for (int Input : IncomingInputs)
9305 M = FreeDWord * 2 + Input % 2;
9307 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9308 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9309 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9310 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9312 // Now enact all the shuffles we've computed to move the inputs into their
9314 if (!isNoopShuffleMask(PSHUFLMask))
9315 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9316 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9317 if (!isNoopShuffleMask(PSHUFHMask))
9318 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9319 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9320 if (!isNoopShuffleMask(PSHUFDMask))
9323 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9324 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9326 // At this point, each half should contain all its inputs, and we can then
9327 // just shuffle them into their final position.
9328 assert(std::count_if(LoMask.begin(), LoMask.end(),
9329 [](int M) { return M >= 4; }) == 0 &&
9330 "Failed to lift all the high half inputs to the low mask!");
9331 assert(std::count_if(HiMask.begin(), HiMask.end(),
9332 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9333 "Failed to lift all the low half inputs to the high mask!");
9335 // Do a half shuffle for the low mask.
9336 if (!isNoopShuffleMask(LoMask))
9337 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9338 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9340 // Do a half shuffle with the high mask after shifting its values down.
9341 for (int &M : HiMask)
9344 if (!isNoopShuffleMask(HiMask))
9345 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9346 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9351 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9352 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9353 SDValue V2, ArrayRef<int> Mask,
9354 SelectionDAG &DAG, bool &V1InUse,
9356 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9362 int Size = Mask.size();
9363 int Scale = 16 / Size;
9364 for (int i = 0; i < 16; ++i) {
9365 if (Mask[i / Scale] == -1) {
9366 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9368 const int ZeroMask = 0x80;
9369 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9371 int V2Idx = Mask[i / Scale] < Size
9373 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9374 if (Zeroable[i / Scale])
9375 V1Idx = V2Idx = ZeroMask;
9376 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9377 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9378 V1InUse |= (ZeroMask != V1Idx);
9379 V2InUse |= (ZeroMask != V2Idx);
9384 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9385 DAG.getBitcast(MVT::v16i8, V1),
9386 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9388 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9389 DAG.getBitcast(MVT::v16i8, V2),
9390 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9392 // If we need shuffled inputs from both, blend the two.
9394 if (V1InUse && V2InUse)
9395 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9397 V = V1InUse ? V1 : V2;
9399 // Cast the result back to the correct type.
9400 return DAG.getBitcast(VT, V);
9403 /// \brief Generic lowering of 8-lane i16 shuffles.
9405 /// This handles both single-input shuffles and combined shuffle/blends with
9406 /// two inputs. The single input shuffles are immediately delegated to
9407 /// a dedicated lowering routine.
9409 /// The blends are lowered in one of three fundamental ways. If there are few
9410 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9411 /// of the input is significantly cheaper when lowered as an interleaving of
9412 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9413 /// halves of the inputs separately (making them have relatively few inputs)
9414 /// and then concatenate them.
9415 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9416 const X86Subtarget *Subtarget,
9417 SelectionDAG &DAG) {
9419 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9420 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9421 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9422 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9423 ArrayRef<int> OrigMask = SVOp->getMask();
9424 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9425 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9426 MutableArrayRef<int> Mask(MaskStorage);
9428 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9430 // Whenever we can lower this as a zext, that instruction is strictly faster
9431 // than any alternative.
9432 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9433 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9436 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9438 auto isV2 = [](int M) { return M >= 8; };
9440 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9442 if (NumV2Inputs == 0) {
9443 // Check for being able to broadcast a single element.
9444 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9445 Mask, Subtarget, DAG))
9448 // Try to use shift instructions.
9450 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9453 // Use dedicated unpack instructions for masks that match their pattern.
9455 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9458 // Try to use byte rotation instructions.
9459 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9460 Mask, Subtarget, DAG))
9463 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9467 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9468 "All single-input shuffles should be canonicalized to be V1-input "
9471 // Try to use shift instructions.
9473 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9476 // See if we can use SSE4A Extraction / Insertion.
9477 if (Subtarget->hasSSE4A())
9478 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9481 // There are special ways we can lower some single-element blends.
9482 if (NumV2Inputs == 1)
9483 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9484 Mask, Subtarget, DAG))
9487 // We have different paths for blend lowering, but they all must use the
9488 // *exact* same predicate.
9489 bool IsBlendSupported = Subtarget->hasSSE41();
9490 if (IsBlendSupported)
9491 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9495 if (SDValue Masked =
9496 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9499 // Use dedicated unpack instructions for masks that match their pattern.
9501 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9504 // Try to use byte rotation instructions.
9505 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9506 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9509 if (SDValue BitBlend =
9510 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9513 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9517 // If we can't directly blend but can use PSHUFB, that will be better as it
9518 // can both shuffle and set up the inefficient blend.
9519 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9520 bool V1InUse, V2InUse;
9521 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9525 // We can always bit-blend if we have to so the fallback strategy is to
9526 // decompose into single-input permutes and blends.
9527 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9531 /// \brief Check whether a compaction lowering can be done by dropping even
9532 /// elements and compute how many times even elements must be dropped.
9534 /// This handles shuffles which take every Nth element where N is a power of
9535 /// two. Example shuffle masks:
9537 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9538 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9539 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9540 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9541 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9542 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9544 /// Any of these lanes can of course be undef.
9546 /// This routine only supports N <= 3.
9547 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9550 /// \returns N above, or the number of times even elements must be dropped if
9551 /// there is such a number. Otherwise returns zero.
9552 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9553 // Figure out whether we're looping over two inputs or just one.
9554 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9556 // The modulus for the shuffle vector entries is based on whether this is
9557 // a single input or not.
9558 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9559 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9560 "We should only be called with masks with a power-of-2 size!");
9562 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9564 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9565 // and 2^3 simultaneously. This is because we may have ambiguity with
9566 // partially undef inputs.
9567 bool ViableForN[3] = {true, true, true};
9569 for (int i = 0, e = Mask.size(); i < e; ++i) {
9570 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9575 bool IsAnyViable = false;
9576 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9577 if (ViableForN[j]) {
9580 // The shuffle mask must be equal to (i * 2^N) % M.
9581 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9584 ViableForN[j] = false;
9586 // Early exit if we exhaust the possible powers of two.
9591 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9595 // Return 0 as there is no viable power of two.
9599 /// \brief Generic lowering of v16i8 shuffles.
9601 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9602 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9603 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9604 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9606 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9607 const X86Subtarget *Subtarget,
9608 SelectionDAG &DAG) {
9610 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9611 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9612 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9613 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9614 ArrayRef<int> Mask = SVOp->getMask();
9615 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9617 // Try to use shift instructions.
9619 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9622 // Try to use byte rotation instructions.
9623 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9624 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9627 // Try to use a zext lowering.
9628 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9629 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9632 // See if we can use SSE4A Extraction / Insertion.
9633 if (Subtarget->hasSSE4A())
9634 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9638 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9640 // For single-input shuffles, there are some nicer lowering tricks we can use.
9641 if (NumV2Elements == 0) {
9642 // Check for being able to broadcast a single element.
9643 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9644 Mask, Subtarget, DAG))
9647 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9648 // Notably, this handles splat and partial-splat shuffles more efficiently.
9649 // However, it only makes sense if the pre-duplication shuffle simplifies
9650 // things significantly. Currently, this means we need to be able to
9651 // express the pre-duplication shuffle as an i16 shuffle.
9653 // FIXME: We should check for other patterns which can be widened into an
9654 // i16 shuffle as well.
9655 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9656 for (int i = 0; i < 16; i += 2)
9657 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9662 auto tryToWidenViaDuplication = [&]() -> SDValue {
9663 if (!canWidenViaDuplication(Mask))
9665 SmallVector<int, 4> LoInputs;
9666 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9667 [](int M) { return M >= 0 && M < 8; });
9668 std::sort(LoInputs.begin(), LoInputs.end());
9669 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9671 SmallVector<int, 4> HiInputs;
9672 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9673 [](int M) { return M >= 8; });
9674 std::sort(HiInputs.begin(), HiInputs.end());
9675 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9678 bool TargetLo = LoInputs.size() >= HiInputs.size();
9679 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9680 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9682 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9683 SmallDenseMap<int, int, 8> LaneMap;
9684 for (int I : InPlaceInputs) {
9685 PreDupI16Shuffle[I/2] = I/2;
9688 int j = TargetLo ? 0 : 4, je = j + 4;
9689 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9690 // Check if j is already a shuffle of this input. This happens when
9691 // there are two adjacent bytes after we move the low one.
9692 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9693 // If we haven't yet mapped the input, search for a slot into which
9695 while (j < je && PreDupI16Shuffle[j] != -1)
9699 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9702 // Map this input with the i16 shuffle.
9703 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9706 // Update the lane map based on the mapping we ended up with.
9707 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9709 V1 = DAG.getBitcast(
9711 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9712 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9714 // Unpack the bytes to form the i16s that will be shuffled into place.
9715 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9716 MVT::v16i8, V1, V1);
9718 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9719 for (int i = 0; i < 16; ++i)
9720 if (Mask[i] != -1) {
9721 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9722 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9723 if (PostDupI16Shuffle[i / 2] == -1)
9724 PostDupI16Shuffle[i / 2] = MappedMask;
9726 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9727 "Conflicting entrties in the original shuffle!");
9729 return DAG.getBitcast(
9731 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9732 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9734 if (SDValue V = tryToWidenViaDuplication())
9738 if (SDValue Masked =
9739 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9742 // Use dedicated unpack instructions for masks that match their pattern.
9744 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9747 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9748 // with PSHUFB. It is important to do this before we attempt to generate any
9749 // blends but after all of the single-input lowerings. If the single input
9750 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9751 // want to preserve that and we can DAG combine any longer sequences into
9752 // a PSHUFB in the end. But once we start blending from multiple inputs,
9753 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9754 // and there are *very* few patterns that would actually be faster than the
9755 // PSHUFB approach because of its ability to zero lanes.
9757 // FIXME: The only exceptions to the above are blends which are exact
9758 // interleavings with direct instructions supporting them. We currently don't
9759 // handle those well here.
9760 if (Subtarget->hasSSSE3()) {
9761 bool V1InUse = false;
9762 bool V2InUse = false;
9764 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9765 DAG, V1InUse, V2InUse);
9767 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9768 // do so. This avoids using them to handle blends-with-zero which is
9769 // important as a single pshufb is significantly faster for that.
9770 if (V1InUse && V2InUse) {
9771 if (Subtarget->hasSSE41())
9772 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9773 Mask, Subtarget, DAG))
9776 // We can use an unpack to do the blending rather than an or in some
9777 // cases. Even though the or may be (very minorly) more efficient, we
9778 // preference this lowering because there are common cases where part of
9779 // the complexity of the shuffles goes away when we do the final blend as
9781 // FIXME: It might be worth trying to detect if the unpack-feeding
9782 // shuffles will both be pshufb, in which case we shouldn't bother with
9784 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9785 DL, MVT::v16i8, V1, V2, Mask, DAG))
9792 // There are special ways we can lower some single-element blends.
9793 if (NumV2Elements == 1)
9794 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9795 Mask, Subtarget, DAG))
9798 if (SDValue BitBlend =
9799 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9802 // Check whether a compaction lowering can be done. This handles shuffles
9803 // which take every Nth element for some even N. See the helper function for
9806 // We special case these as they can be particularly efficiently handled with
9807 // the PACKUSB instruction on x86 and they show up in common patterns of
9808 // rearranging bytes to truncate wide elements.
9809 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9810 // NumEvenDrops is the power of two stride of the elements. Another way of
9811 // thinking about it is that we need to drop the even elements this many
9812 // times to get the original input.
9813 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9815 // First we need to zero all the dropped bytes.
9816 assert(NumEvenDrops <= 3 &&
9817 "No support for dropping even elements more than 3 times.");
9818 // We use the mask type to pick which bytes are preserved based on how many
9819 // elements are dropped.
9820 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9821 SDValue ByteClearMask = DAG.getBitcast(
9822 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9823 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9825 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9827 // Now pack things back together.
9828 V1 = DAG.getBitcast(MVT::v8i16, V1);
9829 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9830 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9831 for (int i = 1; i < NumEvenDrops; ++i) {
9832 Result = DAG.getBitcast(MVT::v8i16, Result);
9833 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9839 // Handle multi-input cases by blending single-input shuffles.
9840 if (NumV2Elements > 0)
9841 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9844 // The fallback path for single-input shuffles widens this into two v8i16
9845 // vectors with unpacks, shuffles those, and then pulls them back together
9849 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9850 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9851 for (int i = 0; i < 16; ++i)
9853 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9855 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9857 SDValue VLoHalf, VHiHalf;
9858 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9859 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9861 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9862 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9863 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9864 [](int M) { return M >= 0 && M % 2 == 1; })) {
9865 // Use a mask to drop the high bytes.
9866 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9867 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9868 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9870 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9871 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9873 // Squash the masks to point directly into VLoHalf.
9874 for (int &M : LoBlendMask)
9877 for (int &M : HiBlendMask)
9881 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9882 // VHiHalf so that we can blend them as i16s.
9883 VLoHalf = DAG.getBitcast(
9884 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9885 VHiHalf = DAG.getBitcast(
9886 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9889 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9890 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9892 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9895 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9897 /// This routine breaks down the specific type of 128-bit shuffle and
9898 /// dispatches to the lowering routines accordingly.
9899 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9900 MVT VT, const X86Subtarget *Subtarget,
9901 SelectionDAG &DAG) {
9902 switch (VT.SimpleTy) {
9904 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9906 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9908 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9910 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9912 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9914 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9917 llvm_unreachable("Unimplemented!");
9921 /// \brief Helper function to test whether a shuffle mask could be
9922 /// simplified by widening the elements being shuffled.
9924 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9925 /// leaves it in an unspecified state.
9927 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9928 /// shuffle masks. The latter have the special property of a '-2' representing
9929 /// a zero-ed lane of a vector.
9930 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9931 SmallVectorImpl<int> &WidenedMask) {
9932 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9933 // If both elements are undef, its trivial.
9934 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9935 WidenedMask.push_back(SM_SentinelUndef);
9939 // Check for an undef mask and a mask value properly aligned to fit with
9940 // a pair of values. If we find such a case, use the non-undef mask's value.
9941 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9942 WidenedMask.push_back(Mask[i + 1] / 2);
9945 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9946 WidenedMask.push_back(Mask[i] / 2);
9950 // When zeroing, we need to spread the zeroing across both lanes to widen.
9951 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9952 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9953 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9954 WidenedMask.push_back(SM_SentinelZero);
9960 // Finally check if the two mask values are adjacent and aligned with
9962 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9963 WidenedMask.push_back(Mask[i] / 2);
9967 // Otherwise we can't safely widen the elements used in this shuffle.
9970 assert(WidenedMask.size() == Mask.size() / 2 &&
9971 "Incorrect size of mask after widening the elements!");
9976 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9978 /// This routine just extracts two subvectors, shuffles them independently, and
9979 /// then concatenates them back together. This should work effectively with all
9980 /// AVX vector shuffle types.
9981 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9982 SDValue V2, ArrayRef<int> Mask,
9983 SelectionDAG &DAG) {
9984 assert(VT.getSizeInBits() >= 256 &&
9985 "Only for 256-bit or wider vector shuffles!");
9986 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9987 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9989 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9990 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9992 int NumElements = VT.getVectorNumElements();
9993 int SplitNumElements = NumElements / 2;
9994 MVT ScalarVT = VT.getVectorElementType();
9995 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9997 // Rather than splitting build-vectors, just build two narrower build
9998 // vectors. This helps shuffling with splats and zeros.
9999 auto SplitVector = [&](SDValue V) {
10000 while (V.getOpcode() == ISD::BITCAST)
10001 V = V->getOperand(0);
10003 MVT OrigVT = V.getSimpleValueType();
10004 int OrigNumElements = OrigVT.getVectorNumElements();
10005 int OrigSplitNumElements = OrigNumElements / 2;
10006 MVT OrigScalarVT = OrigVT.getVectorElementType();
10007 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
10011 auto *BV = dyn_cast<BuildVectorSDNode>(V);
10013 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10014 DAG.getIntPtrConstant(0, DL));
10015 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10016 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
10019 SmallVector<SDValue, 16> LoOps, HiOps;
10020 for (int i = 0; i < OrigSplitNumElements; ++i) {
10021 LoOps.push_back(BV->getOperand(i));
10022 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
10024 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
10025 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
10027 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
10028 DAG.getBitcast(SplitVT, HiV));
10031 SDValue LoV1, HiV1, LoV2, HiV2;
10032 std::tie(LoV1, HiV1) = SplitVector(V1);
10033 std::tie(LoV2, HiV2) = SplitVector(V2);
10035 // Now create two 4-way blends of these half-width vectors.
10036 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
10037 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
10038 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
10039 for (int i = 0; i < SplitNumElements; ++i) {
10040 int M = HalfMask[i];
10041 if (M >= NumElements) {
10042 if (M >= NumElements + SplitNumElements)
10046 V2BlendMask.push_back(M - NumElements);
10047 V1BlendMask.push_back(-1);
10048 BlendMask.push_back(SplitNumElements + i);
10049 } else if (M >= 0) {
10050 if (M >= SplitNumElements)
10054 V2BlendMask.push_back(-1);
10055 V1BlendMask.push_back(M);
10056 BlendMask.push_back(i);
10058 V2BlendMask.push_back(-1);
10059 V1BlendMask.push_back(-1);
10060 BlendMask.push_back(-1);
10064 // Because the lowering happens after all combining takes place, we need to
10065 // manually combine these blend masks as much as possible so that we create
10066 // a minimal number of high-level vector shuffle nodes.
10068 // First try just blending the halves of V1 or V2.
10069 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10070 return DAG.getUNDEF(SplitVT);
10071 if (!UseLoV2 && !UseHiV2)
10072 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10073 if (!UseLoV1 && !UseHiV1)
10074 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10076 SDValue V1Blend, V2Blend;
10077 if (UseLoV1 && UseHiV1) {
10079 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10081 // We only use half of V1 so map the usage down into the final blend mask.
10082 V1Blend = UseLoV1 ? LoV1 : HiV1;
10083 for (int i = 0; i < SplitNumElements; ++i)
10084 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10085 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10087 if (UseLoV2 && UseHiV2) {
10089 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10091 // We only use half of V2 so map the usage down into the final blend mask.
10092 V2Blend = UseLoV2 ? LoV2 : HiV2;
10093 for (int i = 0; i < SplitNumElements; ++i)
10094 if (BlendMask[i] >= SplitNumElements)
10095 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10097 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10099 SDValue Lo = HalfBlend(LoMask);
10100 SDValue Hi = HalfBlend(HiMask);
10101 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10104 /// \brief Either split a vector in halves or decompose the shuffles and the
10107 /// This is provided as a good fallback for many lowerings of non-single-input
10108 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10109 /// between splitting the shuffle into 128-bit components and stitching those
10110 /// back together vs. extracting the single-input shuffles and blending those
10112 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10113 SDValue V2, ArrayRef<int> Mask,
10114 SelectionDAG &DAG) {
10115 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10116 "lower single-input shuffles as it "
10117 "could then recurse on itself.");
10118 int Size = Mask.size();
10120 // If this can be modeled as a broadcast of two elements followed by a blend,
10121 // prefer that lowering. This is especially important because broadcasts can
10122 // often fold with memory operands.
10123 auto DoBothBroadcast = [&] {
10124 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10127 if (V2BroadcastIdx == -1)
10128 V2BroadcastIdx = M - Size;
10129 else if (M - Size != V2BroadcastIdx)
10131 } else if (M >= 0) {
10132 if (V1BroadcastIdx == -1)
10133 V1BroadcastIdx = M;
10134 else if (M != V1BroadcastIdx)
10139 if (DoBothBroadcast())
10140 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10143 // If the inputs all stem from a single 128-bit lane of each input, then we
10144 // split them rather than blending because the split will decompose to
10145 // unusually few instructions.
10146 int LaneCount = VT.getSizeInBits() / 128;
10147 int LaneSize = Size / LaneCount;
10148 SmallBitVector LaneInputs[2];
10149 LaneInputs[0].resize(LaneCount, false);
10150 LaneInputs[1].resize(LaneCount, false);
10151 for (int i = 0; i < Size; ++i)
10153 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10154 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10155 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10157 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10158 // that the decomposed single-input shuffles don't end up here.
10159 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10162 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10163 /// a permutation and blend of those lanes.
10165 /// This essentially blends the out-of-lane inputs to each lane into the lane
10166 /// from a permuted copy of the vector. This lowering strategy results in four
10167 /// instructions in the worst case for a single-input cross lane shuffle which
10168 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10169 /// of. Special cases for each particular shuffle pattern should be handled
10170 /// prior to trying this lowering.
10171 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10172 SDValue V1, SDValue V2,
10173 ArrayRef<int> Mask,
10174 SelectionDAG &DAG) {
10175 // FIXME: This should probably be generalized for 512-bit vectors as well.
10176 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
10177 int LaneSize = Mask.size() / 2;
10179 // If there are only inputs from one 128-bit lane, splitting will in fact be
10180 // less expensive. The flags track whether the given lane contains an element
10181 // that crosses to another lane.
10182 bool LaneCrossing[2] = {false, false};
10183 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10184 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10185 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10186 if (!LaneCrossing[0] || !LaneCrossing[1])
10187 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10189 if (isSingleInputShuffleMask(Mask)) {
10190 SmallVector<int, 32> FlippedBlendMask;
10191 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10192 FlippedBlendMask.push_back(
10193 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10195 : Mask[i] % LaneSize +
10196 (i / LaneSize) * LaneSize + Size));
10198 // Flip the vector, and blend the results which should now be in-lane. The
10199 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10200 // 5 for the high source. The value 3 selects the high half of source 2 and
10201 // the value 2 selects the low half of source 2. We only use source 2 to
10202 // allow folding it into a memory operand.
10203 unsigned PERMMask = 3 | 2 << 4;
10204 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10205 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10206 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10209 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10210 // will be handled by the above logic and a blend of the results, much like
10211 // other patterns in AVX.
10212 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10215 /// \brief Handle lowering 2-lane 128-bit shuffles.
10216 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10217 SDValue V2, ArrayRef<int> Mask,
10218 const X86Subtarget *Subtarget,
10219 SelectionDAG &DAG) {
10220 // TODO: If minimizing size and one of the inputs is a zero vector and the
10221 // the zero vector has only one use, we could use a VPERM2X128 to save the
10222 // instruction bytes needed to explicitly generate the zero vector.
10224 // Blends are faster and handle all the non-lane-crossing cases.
10225 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10229 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10230 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10232 // If either input operand is a zero vector, use VPERM2X128 because its mask
10233 // allows us to replace the zero input with an implicit zero.
10234 if (!IsV1Zero && !IsV2Zero) {
10235 // Check for patterns which can be matched with a single insert of a 128-bit
10237 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10238 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10239 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10240 VT.getVectorNumElements() / 2);
10241 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10242 DAG.getIntPtrConstant(0, DL));
10243 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10244 OnlyUsesV1 ? V1 : V2,
10245 DAG.getIntPtrConstant(0, DL));
10246 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10250 // Otherwise form a 128-bit permutation. After accounting for undefs,
10251 // convert the 64-bit shuffle mask selection values into 128-bit
10252 // selection bits by dividing the indexes by 2 and shifting into positions
10253 // defined by a vperm2*128 instruction's immediate control byte.
10255 // The immediate permute control byte looks like this:
10256 // [1:0] - select 128 bits from sources for low half of destination
10258 // [3] - zero low half of destination
10259 // [5:4] - select 128 bits from sources for high half of destination
10261 // [7] - zero high half of destination
10263 int MaskLO = Mask[0];
10264 if (MaskLO == SM_SentinelUndef)
10265 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10267 int MaskHI = Mask[2];
10268 if (MaskHI == SM_SentinelUndef)
10269 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10271 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10273 // If either input is a zero vector, replace it with an undef input.
10274 // Shuffle mask values < 4 are selecting elements of V1.
10275 // Shuffle mask values >= 4 are selecting elements of V2.
10276 // Adjust each half of the permute mask by clearing the half that was
10277 // selecting the zero vector and setting the zero mask bit.
10279 V1 = DAG.getUNDEF(VT);
10281 PermMask = (PermMask & 0xf0) | 0x08;
10283 PermMask = (PermMask & 0x0f) | 0x80;
10286 V2 = DAG.getUNDEF(VT);
10288 PermMask = (PermMask & 0xf0) | 0x08;
10290 PermMask = (PermMask & 0x0f) | 0x80;
10293 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10294 DAG.getConstant(PermMask, DL, MVT::i8));
10297 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10298 /// shuffling each lane.
10300 /// This will only succeed when the result of fixing the 128-bit lanes results
10301 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10302 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10303 /// the lane crosses early and then use simpler shuffles within each lane.
10305 /// FIXME: It might be worthwhile at some point to support this without
10306 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10307 /// in x86 only floating point has interesting non-repeating shuffles, and even
10308 /// those are still *marginally* more expensive.
10309 static SDValue lowerVectorShuffleByMerging128BitLanes(
10310 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10311 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10312 assert(!isSingleInputShuffleMask(Mask) &&
10313 "This is only useful with multiple inputs.");
10315 int Size = Mask.size();
10316 int LaneSize = 128 / VT.getScalarSizeInBits();
10317 int NumLanes = Size / LaneSize;
10318 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10320 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10321 // check whether the in-128-bit lane shuffles share a repeating pattern.
10322 SmallVector<int, 4> Lanes;
10323 Lanes.resize(NumLanes, -1);
10324 SmallVector<int, 4> InLaneMask;
10325 InLaneMask.resize(LaneSize, -1);
10326 for (int i = 0; i < Size; ++i) {
10330 int j = i / LaneSize;
10332 if (Lanes[j] < 0) {
10333 // First entry we've seen for this lane.
10334 Lanes[j] = Mask[i] / LaneSize;
10335 } else if (Lanes[j] != Mask[i] / LaneSize) {
10336 // This doesn't match the lane selected previously!
10340 // Check that within each lane we have a consistent shuffle mask.
10341 int k = i % LaneSize;
10342 if (InLaneMask[k] < 0) {
10343 InLaneMask[k] = Mask[i] % LaneSize;
10344 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10345 // This doesn't fit a repeating in-lane mask.
10350 // First shuffle the lanes into place.
10351 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10352 VT.getSizeInBits() / 64);
10353 SmallVector<int, 8> LaneMask;
10354 LaneMask.resize(NumLanes * 2, -1);
10355 for (int i = 0; i < NumLanes; ++i)
10356 if (Lanes[i] >= 0) {
10357 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10358 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10361 V1 = DAG.getBitcast(LaneVT, V1);
10362 V2 = DAG.getBitcast(LaneVT, V2);
10363 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10365 // Cast it back to the type we actually want.
10366 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10368 // Now do a simple shuffle that isn't lane crossing.
10369 SmallVector<int, 8> NewMask;
10370 NewMask.resize(Size, -1);
10371 for (int i = 0; i < Size; ++i)
10373 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10374 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10375 "Must not introduce lane crosses at this point!");
10377 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10380 /// Lower shuffles where an entire half of a 256-bit vector is UNDEF.
10381 /// This allows for fast cases such as subvector extraction/insertion
10382 /// or shuffling smaller vector types which can lower more efficiently.
10383 static SDValue lowerVectorShuffleWithUndefHalf(SDLoc DL, MVT VT, SDValue V1,
10384 SDValue V2, ArrayRef<int> Mask,
10385 const X86Subtarget *Subtarget,
10386 SelectionDAG &DAG) {
10387 assert(VT.getSizeInBits() == 256 && "Expected 256-bit vector");
10389 unsigned NumElts = VT.getVectorNumElements();
10390 unsigned HalfNumElts = NumElts / 2;
10391 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(), HalfNumElts);
10393 bool UndefLower = isUndefInRange(Mask, 0, HalfNumElts);
10394 bool UndefUpper = isUndefInRange(Mask, HalfNumElts, HalfNumElts);
10395 if (!UndefLower && !UndefUpper)
10398 // Upper half is undef and lower half is whole upper subvector.
10399 // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
10401 isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) {
10402 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10403 DAG.getIntPtrConstant(HalfNumElts, DL));
10404 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10405 DAG.getIntPtrConstant(0, DL));
10408 // Lower half is undef and upper half is whole lower subvector.
10409 // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
10411 isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) {
10412 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10413 DAG.getIntPtrConstant(0, DL));
10414 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10415 DAG.getIntPtrConstant(HalfNumElts, DL));
10418 // AVX2 supports efficient immediate 64-bit element cross-lane shuffles.
10419 if (UndefLower && Subtarget->hasAVX2() &&
10420 (VT == MVT::v4f64 || VT == MVT::v4i64))
10423 // If the shuffle only uses the lower halves of the input operands,
10424 // then extract them and perform the 'half' shuffle at half width.
10425 // e.g. vector_shuffle <X, X, X, X, u, u, u, u> or <X, X, u, u>
10426 int HalfIdx1 = -1, HalfIdx2 = -1;
10427 SmallVector<int, 8> HalfMask;
10428 unsigned Offset = UndefLower ? HalfNumElts : 0;
10429 for (unsigned i = 0; i != HalfNumElts; ++i) {
10430 int M = Mask[i + Offset];
10432 HalfMask.push_back(M);
10436 // Determine which of the 4 half vectors this element is from.
10437 // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2.
10438 int HalfIdx = M / HalfNumElts;
10440 // Only shuffle using the lower halves of the inputs.
10441 // TODO: Investigate usefulness of shuffling with upper halves.
10442 if (HalfIdx != 0 && HalfIdx != 2)
10445 // Determine the element index into its half vector source.
10446 int HalfElt = M % HalfNumElts;
10448 // We can shuffle with up to 2 half vectors, set the new 'half'
10449 // shuffle mask accordingly.
10450 if (-1 == HalfIdx1 || HalfIdx1 == HalfIdx) {
10451 HalfMask.push_back(HalfElt);
10452 HalfIdx1 = HalfIdx;
10455 if (-1 == HalfIdx2 || HalfIdx2 == HalfIdx) {
10456 HalfMask.push_back(HalfElt + HalfNumElts);
10457 HalfIdx2 = HalfIdx;
10461 // Too many half vectors referenced.
10464 assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length");
10466 auto GetHalfVector = [&](int HalfIdx) {
10468 return DAG.getUNDEF(HalfVT);
10469 SDValue V = (HalfIdx < 2 ? V1 : V2);
10470 HalfIdx = (HalfIdx % 2) * HalfNumElts;
10471 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V,
10472 DAG.getIntPtrConstant(HalfIdx, DL));
10475 SDValue Half1 = GetHalfVector(HalfIdx1);
10476 SDValue Half2 = GetHalfVector(HalfIdx2);
10477 SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask);
10478 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V,
10479 DAG.getIntPtrConstant(Offset, DL));
10482 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10485 /// This returns true if the elements from a particular input are already in the
10486 /// slot required by the given mask and require no permutation.
10487 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10488 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10489 int Size = Mask.size();
10490 for (int i = 0; i < Size; ++i)
10491 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10497 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10498 ArrayRef<int> Mask, SDValue V1,
10499 SDValue V2, SelectionDAG &DAG) {
10501 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10502 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10503 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10504 int NumElts = VT.getVectorNumElements();
10505 bool ShufpdMask = true;
10506 bool CommutableMask = true;
10507 unsigned Immediate = 0;
10508 for (int i = 0; i < NumElts; ++i) {
10511 int Val = (i & 6) + NumElts * (i & 1);
10512 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10513 if (Mask[i] < Val || Mask[i] > Val + 1)
10514 ShufpdMask = false;
10515 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10516 CommutableMask = false;
10517 Immediate |= (Mask[i] % 2) << i;
10520 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10521 DAG.getConstant(Immediate, DL, MVT::i8));
10522 if (CommutableMask)
10523 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10524 DAG.getConstant(Immediate, DL, MVT::i8));
10528 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10530 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10531 /// isn't available.
10532 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10533 const X86Subtarget *Subtarget,
10534 SelectionDAG &DAG) {
10536 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10537 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10538 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10539 ArrayRef<int> Mask = SVOp->getMask();
10540 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10542 SmallVector<int, 4> WidenedMask;
10543 if (canWidenShuffleElements(Mask, WidenedMask))
10544 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10547 if (isSingleInputShuffleMask(Mask)) {
10548 // Check for being able to broadcast a single element.
10549 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10550 Mask, Subtarget, DAG))
10553 // Use low duplicate instructions for masks that match their pattern.
10554 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10555 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10557 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10558 // Non-half-crossing single input shuffles can be lowerid with an
10559 // interleaved permutation.
10560 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10561 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10562 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10563 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10566 // With AVX2 we have direct support for this permutation.
10567 if (Subtarget->hasAVX2())
10568 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10569 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10571 // Otherwise, fall back.
10572 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10576 // Use dedicated unpack instructions for masks that match their pattern.
10578 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10581 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10585 // Check if the blend happens to exactly fit that of SHUFPD.
10587 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10590 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10591 // shuffle. However, if we have AVX2 and either inputs are already in place,
10592 // we will be able to shuffle even across lanes the other input in a single
10593 // instruction so skip this pattern.
10594 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10595 isShuffleMaskInputInPlace(1, Mask))))
10596 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10597 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10600 // If we have AVX2 then we always want to lower with a blend because an v4 we
10601 // can fully permute the elements.
10602 if (Subtarget->hasAVX2())
10603 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10606 // Otherwise fall back on generic lowering.
10607 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10610 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10612 /// This routine is only called when we have AVX2 and thus a reasonable
10613 /// instruction set for v4i64 shuffling..
10614 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10615 const X86Subtarget *Subtarget,
10616 SelectionDAG &DAG) {
10618 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10619 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10620 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10621 ArrayRef<int> Mask = SVOp->getMask();
10622 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10623 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10625 SmallVector<int, 4> WidenedMask;
10626 if (canWidenShuffleElements(Mask, WidenedMask))
10627 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10630 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10634 // Check for being able to broadcast a single element.
10635 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10636 Mask, Subtarget, DAG))
10639 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10640 // use lower latency instructions that will operate on both 128-bit lanes.
10641 SmallVector<int, 2> RepeatedMask;
10642 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10643 if (isSingleInputShuffleMask(Mask)) {
10644 int PSHUFDMask[] = {-1, -1, -1, -1};
10645 for (int i = 0; i < 2; ++i)
10646 if (RepeatedMask[i] >= 0) {
10647 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10648 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10650 return DAG.getBitcast(
10652 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10653 DAG.getBitcast(MVT::v8i32, V1),
10654 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10658 // AVX2 provides a direct instruction for permuting a single input across
10660 if (isSingleInputShuffleMask(Mask))
10661 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10662 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10664 // Try to use shift instructions.
10665 if (SDValue Shift =
10666 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10669 // Use dedicated unpack instructions for masks that match their pattern.
10671 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10674 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10675 // shuffle. However, if we have AVX2 and either inputs are already in place,
10676 // we will be able to shuffle even across lanes the other input in a single
10677 // instruction so skip this pattern.
10678 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10679 isShuffleMaskInputInPlace(1, Mask))))
10680 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10681 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10684 // Otherwise fall back on generic blend lowering.
10685 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10689 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10691 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10692 /// isn't available.
10693 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10694 const X86Subtarget *Subtarget,
10695 SelectionDAG &DAG) {
10697 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10698 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10699 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10700 ArrayRef<int> Mask = SVOp->getMask();
10701 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10703 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10707 // Check for being able to broadcast a single element.
10708 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10709 Mask, Subtarget, DAG))
10712 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10713 // options to efficiently lower the shuffle.
10714 SmallVector<int, 4> RepeatedMask;
10715 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10716 assert(RepeatedMask.size() == 4 &&
10717 "Repeated masks must be half the mask width!");
10719 // Use even/odd duplicate instructions for masks that match their pattern.
10720 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10721 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10722 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10723 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10725 if (isSingleInputShuffleMask(Mask))
10726 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10727 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10729 // Use dedicated unpack instructions for masks that match their pattern.
10731 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10734 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10735 // have already handled any direct blends. We also need to squash the
10736 // repeated mask into a simulated v4f32 mask.
10737 for (int i = 0; i < 4; ++i)
10738 if (RepeatedMask[i] >= 8)
10739 RepeatedMask[i] -= 4;
10740 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10743 // If we have a single input shuffle with different shuffle patterns in the
10744 // two 128-bit lanes use the variable mask to VPERMILPS.
10745 if (isSingleInputShuffleMask(Mask)) {
10746 SDValue VPermMask[8];
10747 for (int i = 0; i < 8; ++i)
10748 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10749 : DAG.getConstant(Mask[i], DL, MVT::i32);
10750 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10751 return DAG.getNode(
10752 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10753 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10755 if (Subtarget->hasAVX2())
10756 return DAG.getNode(
10757 X86ISD::VPERMV, DL, MVT::v8f32,
10758 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10760 // Otherwise, fall back.
10761 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10765 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10767 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10768 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10771 // If we have AVX2 then we always want to lower with a blend because at v8 we
10772 // can fully permute the elements.
10773 if (Subtarget->hasAVX2())
10774 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10777 // Otherwise fall back on generic lowering.
10778 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10781 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10783 /// This routine is only called when we have AVX2 and thus a reasonable
10784 /// instruction set for v8i32 shuffling..
10785 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10786 const X86Subtarget *Subtarget,
10787 SelectionDAG &DAG) {
10789 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10790 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10791 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10792 ArrayRef<int> Mask = SVOp->getMask();
10793 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10794 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10796 // Whenever we can lower this as a zext, that instruction is strictly faster
10797 // than any alternative. It also allows us to fold memory operands into the
10798 // shuffle in many cases.
10799 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10800 Mask, Subtarget, DAG))
10803 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10807 // Check for being able to broadcast a single element.
10808 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10809 Mask, Subtarget, DAG))
10812 // If the shuffle mask is repeated in each 128-bit lane we can use more
10813 // efficient instructions that mirror the shuffles across the two 128-bit
10815 SmallVector<int, 4> RepeatedMask;
10816 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10817 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10818 if (isSingleInputShuffleMask(Mask))
10819 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10820 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10822 // Use dedicated unpack instructions for masks that match their pattern.
10824 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10828 // Try to use shift instructions.
10829 if (SDValue Shift =
10830 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10833 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10834 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10837 // If the shuffle patterns aren't repeated but it is a single input, directly
10838 // generate a cross-lane VPERMD instruction.
10839 if (isSingleInputShuffleMask(Mask)) {
10840 SDValue VPermMask[8];
10841 for (int i = 0; i < 8; ++i)
10842 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10843 : DAG.getConstant(Mask[i], DL, MVT::i32);
10844 return DAG.getNode(
10845 X86ISD::VPERMV, DL, MVT::v8i32,
10846 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10849 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10851 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10852 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10855 // Otherwise fall back on generic blend lowering.
10856 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10860 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10862 /// This routine is only called when we have AVX2 and thus a reasonable
10863 /// instruction set for v16i16 shuffling..
10864 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10865 const X86Subtarget *Subtarget,
10866 SelectionDAG &DAG) {
10868 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10869 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10870 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10871 ArrayRef<int> Mask = SVOp->getMask();
10872 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10873 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10875 // Whenever we can lower this as a zext, that instruction is strictly faster
10876 // than any alternative. It also allows us to fold memory operands into the
10877 // shuffle in many cases.
10878 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10879 Mask, Subtarget, DAG))
10882 // Check for being able to broadcast a single element.
10883 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10884 Mask, Subtarget, DAG))
10887 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10891 // Use dedicated unpack instructions for masks that match their pattern.
10893 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10896 // Try to use shift instructions.
10897 if (SDValue Shift =
10898 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10901 // Try to use byte rotation instructions.
10902 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10903 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10906 if (isSingleInputShuffleMask(Mask)) {
10907 // There are no generalized cross-lane shuffle operations available on i16
10909 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10910 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10913 SmallVector<int, 8> RepeatedMask;
10914 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10915 // As this is a single-input shuffle, the repeated mask should be
10916 // a strictly valid v8i16 mask that we can pass through to the v8i16
10917 // lowering to handle even the v16 case.
10918 return lowerV8I16GeneralSingleInputVectorShuffle(
10919 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10922 SDValue PSHUFBMask[32];
10923 for (int i = 0; i < 16; ++i) {
10924 if (Mask[i] == -1) {
10925 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10929 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10930 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10931 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10932 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10934 return DAG.getBitcast(MVT::v16i16,
10935 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10936 DAG.getBitcast(MVT::v32i8, V1),
10937 DAG.getNode(ISD::BUILD_VECTOR, DL,
10938 MVT::v32i8, PSHUFBMask)));
10941 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10943 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10944 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10947 // Otherwise fall back on generic lowering.
10948 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10951 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10953 /// This routine is only called when we have AVX2 and thus a reasonable
10954 /// instruction set for v32i8 shuffling..
10955 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10956 const X86Subtarget *Subtarget,
10957 SelectionDAG &DAG) {
10959 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10960 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10961 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10962 ArrayRef<int> Mask = SVOp->getMask();
10963 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10964 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10966 // Whenever we can lower this as a zext, that instruction is strictly faster
10967 // than any alternative. It also allows us to fold memory operands into the
10968 // shuffle in many cases.
10969 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10970 Mask, Subtarget, DAG))
10973 // Check for being able to broadcast a single element.
10974 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10975 Mask, Subtarget, DAG))
10978 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10982 // Use dedicated unpack instructions for masks that match their pattern.
10984 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10987 // Try to use shift instructions.
10988 if (SDValue Shift =
10989 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10992 // Try to use byte rotation instructions.
10993 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10994 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10997 if (isSingleInputShuffleMask(Mask)) {
10998 // There are no generalized cross-lane shuffle operations available on i8
11000 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
11001 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
11004 SDValue PSHUFBMask[32];
11005 for (int i = 0; i < 32; ++i)
11008 ? DAG.getUNDEF(MVT::i8)
11009 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
11012 return DAG.getNode(
11013 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
11014 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
11017 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11019 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11020 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
11023 // Otherwise fall back on generic lowering.
11024 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
11027 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
11029 /// This routine either breaks down the specific type of a 256-bit x86 vector
11030 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
11031 /// together based on the available instructions.
11032 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11033 MVT VT, const X86Subtarget *Subtarget,
11034 SelectionDAG &DAG) {
11036 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11037 ArrayRef<int> Mask = SVOp->getMask();
11039 // If we have a single input to the zero element, insert that into V1 if we
11040 // can do so cheaply.
11041 int NumElts = VT.getVectorNumElements();
11042 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
11043 return M >= NumElts;
11046 if (NumV2Elements == 1 && Mask[0] >= NumElts)
11047 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
11048 DL, VT, V1, V2, Mask, Subtarget, DAG))
11051 // Handle special cases where the lower or upper half is UNDEF.
11053 lowerVectorShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG))
11056 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
11057 // can check for those subtargets here and avoid much of the subtarget
11058 // querying in the per-vector-type lowering routines. With AVX1 we have
11059 // essentially *zero* ability to manipulate a 256-bit vector with integer
11060 // types. Since we'll use floating point types there eventually, just
11061 // immediately cast everything to a float and operate entirely in that domain.
11062 if (VT.isInteger() && !Subtarget->hasAVX2()) {
11063 int ElementBits = VT.getScalarSizeInBits();
11064 if (ElementBits < 32)
11065 // No floating point type available, decompose into 128-bit vectors.
11066 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11068 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
11069 VT.getVectorNumElements());
11070 V1 = DAG.getBitcast(FpVT, V1);
11071 V2 = DAG.getBitcast(FpVT, V2);
11072 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
11075 switch (VT.SimpleTy) {
11077 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11079 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11081 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11083 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11085 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11087 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11090 llvm_unreachable("Not a valid 256-bit x86 vector type!");
11094 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
11095 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
11096 ArrayRef<int> Mask,
11097 SDValue V1, SDValue V2,
11098 SelectionDAG &DAG) {
11099 assert(VT.getScalarSizeInBits() == 64 &&
11100 "Unexpected element type size for 128bit shuffle.");
11102 // To handle 256 bit vector requires VLX and most probably
11103 // function lowerV2X128VectorShuffle() is better solution.
11104 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
11106 SmallVector<int, 4> WidenedMask;
11107 if (!canWidenShuffleElements(Mask, WidenedMask))
11110 // Form a 128-bit permutation.
11111 // Convert the 64-bit shuffle mask selection values into 128-bit selection
11112 // bits defined by a vshuf64x2 instruction's immediate control byte.
11113 unsigned PermMask = 0, Imm = 0;
11114 unsigned ControlBitsNum = WidenedMask.size() / 2;
11116 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
11117 if (WidenedMask[i] == SM_SentinelZero)
11120 // Use first element in place of undef mask.
11121 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
11122 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
11125 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
11126 DAG.getConstant(PermMask, DL, MVT::i8));
11129 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
11130 ArrayRef<int> Mask, SDValue V1,
11131 SDValue V2, SelectionDAG &DAG) {
11133 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
11135 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
11136 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
11138 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
11139 if (isSingleInputShuffleMask(Mask))
11140 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
11142 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
11145 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
11146 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11147 const X86Subtarget *Subtarget,
11148 SelectionDAG &DAG) {
11150 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11151 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11152 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11153 ArrayRef<int> Mask = SVOp->getMask();
11154 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11156 if (SDValue Shuf128 =
11157 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
11160 if (SDValue Unpck =
11161 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
11164 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
11167 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
11168 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11169 const X86Subtarget *Subtarget,
11170 SelectionDAG &DAG) {
11172 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11173 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11174 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11175 ArrayRef<int> Mask = SVOp->getMask();
11176 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11178 if (SDValue Unpck =
11179 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
11182 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
11185 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11186 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11187 const X86Subtarget *Subtarget,
11188 SelectionDAG &DAG) {
11190 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11191 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11192 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11193 ArrayRef<int> Mask = SVOp->getMask();
11194 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11196 if (SDValue Shuf128 =
11197 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
11200 if (SDValue Unpck =
11201 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
11204 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
11207 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11208 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11209 const X86Subtarget *Subtarget,
11210 SelectionDAG &DAG) {
11212 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11213 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11214 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11215 ArrayRef<int> Mask = SVOp->getMask();
11216 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11218 if (SDValue Unpck =
11219 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
11222 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
11225 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11226 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11227 const X86Subtarget *Subtarget,
11228 SelectionDAG &DAG) {
11230 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11231 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11232 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11233 ArrayRef<int> Mask = SVOp->getMask();
11234 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11235 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11237 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
11240 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11241 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11242 const X86Subtarget *Subtarget,
11243 SelectionDAG &DAG) {
11245 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11246 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11247 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11248 ArrayRef<int> Mask = SVOp->getMask();
11249 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11250 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11252 // FIXME: Implement direct support for this type!
11253 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11256 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11258 /// This routine either breaks down the specific type of a 512-bit x86 vector
11259 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11260 /// together based on the available instructions.
11261 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11262 MVT VT, const X86Subtarget *Subtarget,
11263 SelectionDAG &DAG) {
11265 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11266 ArrayRef<int> Mask = SVOp->getMask();
11267 assert(Subtarget->hasAVX512() &&
11268 "Cannot lower 512-bit vectors w/ basic ISA!");
11270 // Check for being able to broadcast a single element.
11271 if (SDValue Broadcast =
11272 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
11275 // Dispatch to each element type for lowering. If we don't have supprot for
11276 // specific element type shuffles at 512 bits, immediately split them and
11277 // lower them. Each lowering routine of a given type is allowed to assume that
11278 // the requisite ISA extensions for that element type are available.
11279 switch (VT.SimpleTy) {
11281 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11283 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11285 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11287 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11289 if (Subtarget->hasBWI())
11290 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11293 if (Subtarget->hasBWI())
11294 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11298 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11301 // Otherwise fall back on splitting.
11302 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11305 // Lower vXi1 vector shuffles.
11306 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11307 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11308 // vector, shuffle and then truncate it back.
11309 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11310 MVT VT, const X86Subtarget *Subtarget,
11311 SelectionDAG &DAG) {
11313 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11314 ArrayRef<int> Mask = SVOp->getMask();
11315 assert(Subtarget->hasAVX512() &&
11316 "Cannot lower 512-bit vectors w/o basic ISA!");
11318 switch (VT.SimpleTy) {
11320 llvm_unreachable("Expected a vector of i1 elements");
11322 ExtVT = MVT::v2i64;
11325 ExtVT = MVT::v4i32;
11328 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11331 ExtVT = MVT::v16i32;
11334 ExtVT = MVT::v32i16;
11337 ExtVT = MVT::v64i8;
11341 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11342 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11343 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11344 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11346 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11349 V2 = DAG.getUNDEF(ExtVT);
11350 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11351 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11352 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11353 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11355 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11356 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11357 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11359 /// \brief Top-level lowering for x86 vector shuffles.
11361 /// This handles decomposition, canonicalization, and lowering of all x86
11362 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11363 /// above in helper routines. The canonicalization attempts to widen shuffles
11364 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11365 /// s.t. only one of the two inputs needs to be tested, etc.
11366 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11367 SelectionDAG &DAG) {
11368 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11369 ArrayRef<int> Mask = SVOp->getMask();
11370 SDValue V1 = Op.getOperand(0);
11371 SDValue V2 = Op.getOperand(1);
11372 MVT VT = Op.getSimpleValueType();
11373 int NumElements = VT.getVectorNumElements();
11375 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11377 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11378 "Can't lower MMX shuffles");
11380 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11381 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11382 if (V1IsUndef && V2IsUndef)
11383 return DAG.getUNDEF(VT);
11385 // When we create a shuffle node we put the UNDEF node to second operand,
11386 // but in some cases the first operand may be transformed to UNDEF.
11387 // In this case we should just commute the node.
11389 return DAG.getCommutedVectorShuffle(*SVOp);
11391 // Check for non-undef masks pointing at an undef vector and make the masks
11392 // undef as well. This makes it easier to match the shuffle based solely on
11396 if (M >= NumElements) {
11397 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11398 for (int &M : NewMask)
11399 if (M >= NumElements)
11401 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11404 // We actually see shuffles that are entirely re-arrangements of a set of
11405 // zero inputs. This mostly happens while decomposing complex shuffles into
11406 // simple ones. Directly lower these as a buildvector of zeros.
11407 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11408 if (Zeroable.all())
11409 return getZeroVector(VT, Subtarget, DAG, dl);
11411 // Try to collapse shuffles into using a vector type with fewer elements but
11412 // wider element types. We cap this to not form integers or floating point
11413 // elements wider than 64 bits, but it might be interesting to form i128
11414 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11415 SmallVector<int, 16> WidenedMask;
11416 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11417 canWidenShuffleElements(Mask, WidenedMask)) {
11418 MVT NewEltVT = VT.isFloatingPoint()
11419 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11420 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11421 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11422 // Make sure that the new vector type is legal. For example, v2f64 isn't
11424 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11425 V1 = DAG.getBitcast(NewVT, V1);
11426 V2 = DAG.getBitcast(NewVT, V2);
11427 return DAG.getBitcast(
11428 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11432 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11433 for (int M : SVOp->getMask())
11435 ++NumUndefElements;
11436 else if (M < NumElements)
11441 // Commute the shuffle as needed such that more elements come from V1 than
11442 // V2. This allows us to match the shuffle pattern strictly on how many
11443 // elements come from V1 without handling the symmetric cases.
11444 if (NumV2Elements > NumV1Elements)
11445 return DAG.getCommutedVectorShuffle(*SVOp);
11447 // When the number of V1 and V2 elements are the same, try to minimize the
11448 // number of uses of V2 in the low half of the vector. When that is tied,
11449 // ensure that the sum of indices for V1 is equal to or lower than the sum
11450 // indices for V2. When those are equal, try to ensure that the number of odd
11451 // indices for V1 is lower than the number of odd indices for V2.
11452 if (NumV1Elements == NumV2Elements) {
11453 int LowV1Elements = 0, LowV2Elements = 0;
11454 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11455 if (M >= NumElements)
11459 if (LowV2Elements > LowV1Elements) {
11460 return DAG.getCommutedVectorShuffle(*SVOp);
11461 } else if (LowV2Elements == LowV1Elements) {
11462 int SumV1Indices = 0, SumV2Indices = 0;
11463 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11464 if (SVOp->getMask()[i] >= NumElements)
11466 else if (SVOp->getMask()[i] >= 0)
11468 if (SumV2Indices < SumV1Indices) {
11469 return DAG.getCommutedVectorShuffle(*SVOp);
11470 } else if (SumV2Indices == SumV1Indices) {
11471 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11472 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11473 if (SVOp->getMask()[i] >= NumElements)
11474 NumV2OddIndices += i % 2;
11475 else if (SVOp->getMask()[i] >= 0)
11476 NumV1OddIndices += i % 2;
11477 if (NumV2OddIndices < NumV1OddIndices)
11478 return DAG.getCommutedVectorShuffle(*SVOp);
11483 // For each vector width, delegate to a specialized lowering routine.
11484 if (VT.is128BitVector())
11485 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11487 if (VT.is256BitVector())
11488 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11490 if (VT.is512BitVector())
11491 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11494 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11495 llvm_unreachable("Unimplemented!");
11498 // This function assumes its argument is a BUILD_VECTOR of constants or
11499 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11501 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11502 unsigned &MaskValue) {
11504 unsigned NumElems = BuildVector->getNumOperands();
11506 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11507 // We don't handle the >2 lanes case right now.
11508 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11512 unsigned NumElemsInLane = NumElems / NumLanes;
11514 // Blend for v16i16 should be symmetric for the both lanes.
11515 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11516 SDValue EltCond = BuildVector->getOperand(i);
11517 SDValue SndLaneEltCond =
11518 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11520 int Lane1Cond = -1, Lane2Cond = -1;
11521 if (isa<ConstantSDNode>(EltCond))
11522 Lane1Cond = !isNullConstant(EltCond);
11523 if (isa<ConstantSDNode>(SndLaneEltCond))
11524 Lane2Cond = !isNullConstant(SndLaneEltCond);
11526 unsigned LaneMask = 0;
11527 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11528 // Lane1Cond != 0, means we want the first argument.
11529 // Lane1Cond == 0, means we want the second argument.
11530 // The encoding of this argument is 0 for the first argument, 1
11531 // for the second. Therefore, invert the condition.
11532 LaneMask = !Lane1Cond << i;
11533 else if (Lane1Cond < 0)
11534 LaneMask = !Lane2Cond << i;
11538 MaskValue |= LaneMask;
11540 MaskValue |= LaneMask << NumElemsInLane;
11545 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11546 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11547 const X86Subtarget *Subtarget,
11548 SelectionDAG &DAG) {
11549 SDValue Cond = Op.getOperand(0);
11550 SDValue LHS = Op.getOperand(1);
11551 SDValue RHS = Op.getOperand(2);
11553 MVT VT = Op.getSimpleValueType();
11555 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11557 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11559 // Only non-legal VSELECTs reach this lowering, convert those into generic
11560 // shuffles and re-use the shuffle lowering path for blends.
11561 SmallVector<int, 32> Mask;
11562 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11563 SDValue CondElt = CondBV->getOperand(i);
11565 isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
11568 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11571 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11572 // A vselect where all conditions and data are constants can be optimized into
11573 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11574 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11575 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11576 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11579 // Try to lower this to a blend-style vector shuffle. This can handle all
11580 // constant condition cases.
11581 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11584 // Variable blends are only legal from SSE4.1 onward.
11585 if (!Subtarget->hasSSE41())
11588 // Only some types will be legal on some subtargets. If we can emit a legal
11589 // VSELECT-matching blend, return Op, and but if we need to expand, return
11591 switch (Op.getSimpleValueType().SimpleTy) {
11593 // Most of the vector types have blends past SSE4.1.
11597 // The byte blends for AVX vectors were introduced only in AVX2.
11598 if (Subtarget->hasAVX2())
11605 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11606 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11609 // FIXME: We should custom lower this by fixing the condition and using i8
11615 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11616 MVT VT = Op.getSimpleValueType();
11619 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11622 if (VT.getSizeInBits() == 8) {
11623 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11624 Op.getOperand(0), Op.getOperand(1));
11625 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11626 DAG.getValueType(VT));
11627 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11630 if (VT.getSizeInBits() == 16) {
11631 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11632 if (isNullConstant(Op.getOperand(1)))
11633 return DAG.getNode(
11634 ISD::TRUNCATE, dl, MVT::i16,
11635 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11636 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11637 Op.getOperand(1)));
11638 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11639 Op.getOperand(0), Op.getOperand(1));
11640 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11641 DAG.getValueType(VT));
11642 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11645 if (VT == MVT::f32) {
11646 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11647 // the result back to FR32 register. It's only worth matching if the
11648 // result has a single use which is a store or a bitcast to i32. And in
11649 // the case of a store, it's not worth it if the index is a constant 0,
11650 // because a MOVSSmr can be used instead, which is smaller and faster.
11651 if (!Op.hasOneUse())
11653 SDNode *User = *Op.getNode()->use_begin();
11654 if ((User->getOpcode() != ISD::STORE ||
11655 isNullConstant(Op.getOperand(1))) &&
11656 (User->getOpcode() != ISD::BITCAST ||
11657 User->getValueType(0) != MVT::i32))
11659 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11660 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11662 return DAG.getBitcast(MVT::f32, Extract);
11665 if (VT == MVT::i32 || VT == MVT::i64) {
11666 // ExtractPS/pextrq works with constant index.
11667 if (isa<ConstantSDNode>(Op.getOperand(1)))
11673 /// Extract one bit from mask vector, like v16i1 or v8i1.
11674 /// AVX-512 feature.
11676 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11677 SDValue Vec = Op.getOperand(0);
11679 MVT VecVT = Vec.getSimpleValueType();
11680 SDValue Idx = Op.getOperand(1);
11681 MVT EltVT = Op.getSimpleValueType();
11683 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11684 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11685 "Unexpected vector type in ExtractBitFromMaskVector");
11687 // variable index can't be handled in mask registers,
11688 // extend vector to VR512
11689 if (!isa<ConstantSDNode>(Idx)) {
11690 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11691 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11692 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11693 ExtVT.getVectorElementType(), Ext, Idx);
11694 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11697 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11698 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11699 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11700 rc = getRegClassFor(MVT::v16i1);
11701 unsigned MaxSift = rc->getSize()*8 - 1;
11702 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11703 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11704 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11705 DAG.getConstant(MaxSift, dl, MVT::i8));
11706 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11707 DAG.getIntPtrConstant(0, dl));
11711 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11712 SelectionDAG &DAG) const {
11714 SDValue Vec = Op.getOperand(0);
11715 MVT VecVT = Vec.getSimpleValueType();
11716 SDValue Idx = Op.getOperand(1);
11718 if (Op.getSimpleValueType() == MVT::i1)
11719 return ExtractBitFromMaskVector(Op, DAG);
11721 if (!isa<ConstantSDNode>(Idx)) {
11722 if (VecVT.is512BitVector() ||
11723 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11724 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11727 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11728 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11729 MaskEltVT.getSizeInBits());
11731 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11732 auto PtrVT = getPointerTy(DAG.getDataLayout());
11733 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11734 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11735 DAG.getConstant(0, dl, PtrVT));
11736 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11737 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11738 DAG.getConstant(0, dl, PtrVT));
11743 // If this is a 256-bit vector result, first extract the 128-bit vector and
11744 // then extract the element from the 128-bit vector.
11745 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11747 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11748 // Get the 128-bit vector.
11749 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11750 MVT EltVT = VecVT.getVectorElementType();
11752 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11753 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11755 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11756 // this can be done with a mask.
11757 IdxVal &= ElemsPerChunk - 1;
11758 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11759 DAG.getConstant(IdxVal, dl, MVT::i32));
11762 assert(VecVT.is128BitVector() && "Unexpected vector length");
11764 if (Subtarget->hasSSE41())
11765 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11768 MVT VT = Op.getSimpleValueType();
11769 // TODO: handle v16i8.
11770 if (VT.getSizeInBits() == 16) {
11771 SDValue Vec = Op.getOperand(0);
11772 if (isNullConstant(Op.getOperand(1)))
11773 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11774 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11775 DAG.getBitcast(MVT::v4i32, Vec),
11776 Op.getOperand(1)));
11777 // Transform it so it match pextrw which produces a 32-bit result.
11778 MVT EltVT = MVT::i32;
11779 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11780 Op.getOperand(0), Op.getOperand(1));
11781 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11782 DAG.getValueType(VT));
11783 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11786 if (VT.getSizeInBits() == 32) {
11787 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11791 // SHUFPS the element to the lowest double word, then movss.
11792 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11793 MVT VVT = Op.getOperand(0).getSimpleValueType();
11794 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11795 DAG.getUNDEF(VVT), Mask);
11796 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11797 DAG.getIntPtrConstant(0, dl));
11800 if (VT.getSizeInBits() == 64) {
11801 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11802 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11803 // to match extract_elt for f64.
11804 if (isNullConstant(Op.getOperand(1)))
11807 // UNPCKHPD the element to the lowest double word, then movsd.
11808 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11809 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11810 int Mask[2] = { 1, -1 };
11811 MVT VVT = Op.getOperand(0).getSimpleValueType();
11812 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11813 DAG.getUNDEF(VVT), Mask);
11814 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11815 DAG.getIntPtrConstant(0, dl));
11821 /// Insert one bit to mask vector, like v16i1 or v8i1.
11822 /// AVX-512 feature.
11824 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11826 SDValue Vec = Op.getOperand(0);
11827 SDValue Elt = Op.getOperand(1);
11828 SDValue Idx = Op.getOperand(2);
11829 MVT VecVT = Vec.getSimpleValueType();
11831 if (!isa<ConstantSDNode>(Idx)) {
11832 // Non constant index. Extend source and destination,
11833 // insert element and then truncate the result.
11834 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11835 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11836 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11837 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11838 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11839 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11842 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11843 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11845 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11846 DAG.getConstant(IdxVal, dl, MVT::i8));
11847 if (Vec.getOpcode() == ISD::UNDEF)
11849 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11852 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11853 SelectionDAG &DAG) const {
11854 MVT VT = Op.getSimpleValueType();
11855 MVT EltVT = VT.getVectorElementType();
11857 if (EltVT == MVT::i1)
11858 return InsertBitToMaskVector(Op, DAG);
11861 SDValue N0 = Op.getOperand(0);
11862 SDValue N1 = Op.getOperand(1);
11863 SDValue N2 = Op.getOperand(2);
11864 if (!isa<ConstantSDNode>(N2))
11866 auto *N2C = cast<ConstantSDNode>(N2);
11867 unsigned IdxVal = N2C->getZExtValue();
11869 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11870 // into that, and then insert the subvector back into the result.
11871 if (VT.is256BitVector() || VT.is512BitVector()) {
11872 // With a 256-bit vector, we can insert into the zero element efficiently
11873 // using a blend if we have AVX or AVX2 and the right data type.
11874 if (VT.is256BitVector() && IdxVal == 0) {
11875 // TODO: It is worthwhile to cast integer to floating point and back
11876 // and incur a domain crossing penalty if that's what we'll end up
11877 // doing anyway after extracting to a 128-bit vector.
11878 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11879 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11880 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11881 N2 = DAG.getIntPtrConstant(1, dl);
11882 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11886 // Get the desired 128-bit vector chunk.
11887 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11889 // Insert the element into the desired chunk.
11890 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11891 assert(isPowerOf2_32(NumEltsIn128));
11892 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11893 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11895 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11896 DAG.getConstant(IdxIn128, dl, MVT::i32));
11898 // Insert the changed part back into the bigger vector
11899 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11901 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11903 if (Subtarget->hasSSE41()) {
11904 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11906 if (VT == MVT::v8i16) {
11907 Opc = X86ISD::PINSRW;
11909 assert(VT == MVT::v16i8);
11910 Opc = X86ISD::PINSRB;
11913 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11915 if (N1.getValueType() != MVT::i32)
11916 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11917 if (N2.getValueType() != MVT::i32)
11918 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11919 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11922 if (EltVT == MVT::f32) {
11923 // Bits [7:6] of the constant are the source select. This will always be
11924 // zero here. The DAG Combiner may combine an extract_elt index into
11925 // these bits. For example (insert (extract, 3), 2) could be matched by
11926 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11927 // Bits [5:4] of the constant are the destination select. This is the
11928 // value of the incoming immediate.
11929 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11930 // combine either bitwise AND or insert of float 0.0 to set these bits.
11932 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11933 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11934 // If this is an insertion of 32-bits into the low 32-bits of
11935 // a vector, we prefer to generate a blend with immediate rather
11936 // than an insertps. Blends are simpler operations in hardware and so
11937 // will always have equal or better performance than insertps.
11938 // But if optimizing for size and there's a load folding opportunity,
11939 // generate insertps because blendps does not have a 32-bit memory
11941 N2 = DAG.getIntPtrConstant(1, dl);
11942 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11943 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11945 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11946 // Create this as a scalar to vector..
11947 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11948 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11951 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11952 // PINSR* works with constant index.
11957 if (EltVT == MVT::i8)
11960 if (EltVT.getSizeInBits() == 16) {
11961 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11962 // as its second argument.
11963 if (N1.getValueType() != MVT::i32)
11964 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11965 if (N2.getValueType() != MVT::i32)
11966 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11967 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11972 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11974 MVT OpVT = Op.getSimpleValueType();
11976 // If this is a 256-bit vector result, first insert into a 128-bit
11977 // vector and then insert into the 256-bit vector.
11978 if (!OpVT.is128BitVector()) {
11979 // Insert into a 128-bit vector.
11980 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11981 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11982 OpVT.getVectorNumElements() / SizeFactor);
11984 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11986 // Insert the 128-bit vector.
11987 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11990 if (OpVT == MVT::v1i64 &&
11991 Op.getOperand(0).getValueType() == MVT::i64)
11992 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11994 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11995 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11996 return DAG.getBitcast(
11997 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
12000 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
12001 // a simple subregister reference or explicit instructions to grab
12002 // upper bits of a vector.
12003 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12004 SelectionDAG &DAG) {
12006 SDValue In = Op.getOperand(0);
12007 SDValue Idx = Op.getOperand(1);
12008 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12009 MVT ResVT = Op.getSimpleValueType();
12010 MVT InVT = In.getSimpleValueType();
12012 if (Subtarget->hasFp256()) {
12013 if (ResVT.is128BitVector() &&
12014 (InVT.is256BitVector() || InVT.is512BitVector()) &&
12015 isa<ConstantSDNode>(Idx)) {
12016 return Extract128BitVector(In, IdxVal, DAG, dl);
12018 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
12019 isa<ConstantSDNode>(Idx)) {
12020 return Extract256BitVector(In, IdxVal, DAG, dl);
12026 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
12027 // simple superregister reference or explicit instructions to insert
12028 // the upper bits of a vector.
12029 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12030 SelectionDAG &DAG) {
12031 if (!Subtarget->hasAVX())
12035 SDValue Vec = Op.getOperand(0);
12036 SDValue SubVec = Op.getOperand(1);
12037 SDValue Idx = Op.getOperand(2);
12039 if (!isa<ConstantSDNode>(Idx))
12042 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12043 MVT OpVT = Op.getSimpleValueType();
12044 MVT SubVecVT = SubVec.getSimpleValueType();
12046 // Fold two 16-byte subvector loads into one 32-byte load:
12047 // (insert_subvector (insert_subvector undef, (load addr), 0),
12048 // (load addr + 16), Elts/2)
12050 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
12051 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
12052 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
12053 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
12054 if (Idx2 && Idx2->getZExtValue() == 0) {
12055 SDValue SubVec2 = Vec.getOperand(1);
12056 // If needed, look through a bitcast to get to the load.
12057 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
12058 SubVec2 = SubVec2.getOperand(0);
12060 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
12062 unsigned Alignment = FirstLd->getAlignment();
12063 unsigned AS = FirstLd->getAddressSpace();
12064 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
12065 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
12066 OpVT, AS, Alignment, &Fast) && Fast) {
12067 SDValue Ops[] = { SubVec2, SubVec };
12068 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
12075 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
12076 SubVecVT.is128BitVector())
12077 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
12079 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
12080 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
12082 if (OpVT.getVectorElementType() == MVT::i1)
12083 return Insert1BitVector(Op, DAG);
12088 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
12089 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
12090 // one of the above mentioned nodes. It has to be wrapped because otherwise
12091 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
12092 // be used to form addressing mode. These wrapped nodes will be selected
12095 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
12096 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
12098 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12099 // global base reg.
12100 unsigned char OpFlag = 0;
12101 unsigned WrapperKind = X86ISD::Wrapper;
12102 CodeModel::Model M = DAG.getTarget().getCodeModel();
12104 if (Subtarget->isPICStyleRIPRel() &&
12105 (M == CodeModel::Small || M == CodeModel::Kernel))
12106 WrapperKind = X86ISD::WrapperRIP;
12107 else if (Subtarget->isPICStyleGOT())
12108 OpFlag = X86II::MO_GOTOFF;
12109 else if (Subtarget->isPICStyleStubPIC())
12110 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12112 auto PtrVT = getPointerTy(DAG.getDataLayout());
12113 SDValue Result = DAG.getTargetConstantPool(
12114 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
12116 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12117 // With PIC, the address is actually $g + Offset.
12120 DAG.getNode(ISD::ADD, DL, PtrVT,
12121 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12127 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
12128 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
12130 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12131 // global base reg.
12132 unsigned char OpFlag = 0;
12133 unsigned WrapperKind = X86ISD::Wrapper;
12134 CodeModel::Model M = DAG.getTarget().getCodeModel();
12136 if (Subtarget->isPICStyleRIPRel() &&
12137 (M == CodeModel::Small || M == CodeModel::Kernel))
12138 WrapperKind = X86ISD::WrapperRIP;
12139 else if (Subtarget->isPICStyleGOT())
12140 OpFlag = X86II::MO_GOTOFF;
12141 else if (Subtarget->isPICStyleStubPIC())
12142 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12144 auto PtrVT = getPointerTy(DAG.getDataLayout());
12145 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
12147 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12149 // With PIC, the address is actually $g + Offset.
12152 DAG.getNode(ISD::ADD, DL, PtrVT,
12153 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12159 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
12160 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
12162 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12163 // global base reg.
12164 unsigned char OpFlag = 0;
12165 unsigned WrapperKind = X86ISD::Wrapper;
12166 CodeModel::Model M = DAG.getTarget().getCodeModel();
12168 if (Subtarget->isPICStyleRIPRel() &&
12169 (M == CodeModel::Small || M == CodeModel::Kernel)) {
12170 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
12171 OpFlag = X86II::MO_GOTPCREL;
12172 WrapperKind = X86ISD::WrapperRIP;
12173 } else if (Subtarget->isPICStyleGOT()) {
12174 OpFlag = X86II::MO_GOT;
12175 } else if (Subtarget->isPICStyleStubPIC()) {
12176 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
12177 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12178 OpFlag = X86II::MO_DARWIN_NONLAZY;
12181 auto PtrVT = getPointerTy(DAG.getDataLayout());
12182 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
12185 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12187 // With PIC, the address is actually $g + Offset.
12188 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12189 !Subtarget->is64Bit()) {
12191 DAG.getNode(ISD::ADD, DL, PtrVT,
12192 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12195 // For symbols that require a load from a stub to get the address, emit the
12197 if (isGlobalStubReference(OpFlag))
12198 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
12199 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12200 false, false, false, 0);
12206 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12207 // Create the TargetBlockAddressAddress node.
12208 unsigned char OpFlags =
12209 Subtarget->ClassifyBlockAddressReference();
12210 CodeModel::Model M = DAG.getTarget().getCodeModel();
12211 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12212 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12214 auto PtrVT = getPointerTy(DAG.getDataLayout());
12215 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
12217 if (Subtarget->isPICStyleRIPRel() &&
12218 (M == CodeModel::Small || M == CodeModel::Kernel))
12219 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12221 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12223 // With PIC, the address is actually $g + Offset.
12224 if (isGlobalRelativeToPICBase(OpFlags)) {
12225 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12226 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12233 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12234 int64_t Offset, SelectionDAG &DAG) const {
12235 // Create the TargetGlobalAddress node, folding in the constant
12236 // offset if it is legal.
12237 unsigned char OpFlags =
12238 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12239 CodeModel::Model M = DAG.getTarget().getCodeModel();
12240 auto PtrVT = getPointerTy(DAG.getDataLayout());
12242 if (OpFlags == X86II::MO_NO_FLAG &&
12243 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12244 // A direct static reference to a global.
12245 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
12248 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
12251 if (Subtarget->isPICStyleRIPRel() &&
12252 (M == CodeModel::Small || M == CodeModel::Kernel))
12253 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12255 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12257 // With PIC, the address is actually $g + Offset.
12258 if (isGlobalRelativeToPICBase(OpFlags)) {
12259 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12260 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12263 // For globals that require a load from a stub to get the address, emit the
12265 if (isGlobalStubReference(OpFlags))
12266 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12267 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12268 false, false, false, 0);
12270 // If there was a non-zero offset that we didn't fold, create an explicit
12271 // addition for it.
12273 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12274 DAG.getConstant(Offset, dl, PtrVT));
12280 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12281 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12282 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12283 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12287 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12288 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12289 unsigned char OperandFlags, bool LocalDynamic = false) {
12290 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12291 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12293 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12294 GA->getValueType(0),
12298 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12302 SDValue Ops[] = { Chain, TGA, *InFlag };
12303 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12305 SDValue Ops[] = { Chain, TGA };
12306 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12309 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12310 MFI->setAdjustsStack(true);
12311 MFI->setHasCalls(true);
12313 SDValue Flag = Chain.getValue(1);
12314 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12317 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12319 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12322 SDLoc dl(GA); // ? function entry point might be better
12323 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12324 DAG.getNode(X86ISD::GlobalBaseReg,
12325 SDLoc(), PtrVT), InFlag);
12326 InFlag = Chain.getValue(1);
12328 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12331 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12333 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12335 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12336 X86::RAX, X86II::MO_TLSGD);
12339 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12345 // Get the start address of the TLS block for this module.
12346 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12347 .getInfo<X86MachineFunctionInfo>();
12348 MFI->incNumLocalDynamicTLSAccesses();
12352 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12353 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12356 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12357 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12358 InFlag = Chain.getValue(1);
12359 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12360 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12363 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12367 unsigned char OperandFlags = X86II::MO_DTPOFF;
12368 unsigned WrapperKind = X86ISD::Wrapper;
12369 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12370 GA->getValueType(0),
12371 GA->getOffset(), OperandFlags);
12372 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12374 // Add x@dtpoff with the base.
12375 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12378 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12379 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12380 const EVT PtrVT, TLSModel::Model model,
12381 bool is64Bit, bool isPIC) {
12384 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12385 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12386 is64Bit ? 257 : 256));
12388 SDValue ThreadPointer =
12389 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12390 MachinePointerInfo(Ptr), false, false, false, 0);
12392 unsigned char OperandFlags = 0;
12393 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12395 unsigned WrapperKind = X86ISD::Wrapper;
12396 if (model == TLSModel::LocalExec) {
12397 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12398 } else if (model == TLSModel::InitialExec) {
12400 OperandFlags = X86II::MO_GOTTPOFF;
12401 WrapperKind = X86ISD::WrapperRIP;
12403 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12406 llvm_unreachable("Unexpected model");
12409 // emit "addl x@ntpoff,%eax" (local exec)
12410 // or "addl x@indntpoff,%eax" (initial exec)
12411 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12413 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12414 GA->getOffset(), OperandFlags);
12415 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12417 if (model == TLSModel::InitialExec) {
12418 if (isPIC && !is64Bit) {
12419 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12420 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12424 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12425 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12426 false, false, false, 0);
12429 // The address of the thread local variable is the add of the thread
12430 // pointer with the offset of the variable.
12431 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12435 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12437 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12439 // Cygwin uses emutls.
12440 // FIXME: It may be EmulatedTLS-generic also for X86-Android.
12441 if (Subtarget->isTargetWindowsCygwin())
12442 return LowerToTLSEmulatedModel(GA, DAG);
12444 const GlobalValue *GV = GA->getGlobal();
12445 auto PtrVT = getPointerTy(DAG.getDataLayout());
12447 if (Subtarget->isTargetELF()) {
12448 if (DAG.getTarget().Options.EmulatedTLS)
12449 return LowerToTLSEmulatedModel(GA, DAG);
12450 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12452 case TLSModel::GeneralDynamic:
12453 if (Subtarget->is64Bit())
12454 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12455 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12456 case TLSModel::LocalDynamic:
12457 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12458 Subtarget->is64Bit());
12459 case TLSModel::InitialExec:
12460 case TLSModel::LocalExec:
12461 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12462 DAG.getTarget().getRelocationModel() ==
12465 llvm_unreachable("Unknown TLS model.");
12468 if (Subtarget->isTargetDarwin()) {
12469 // Darwin only has one model of TLS. Lower to that.
12470 unsigned char OpFlag = 0;
12471 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12472 X86ISD::WrapperRIP : X86ISD::Wrapper;
12474 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12475 // global base reg.
12476 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12477 !Subtarget->is64Bit();
12479 OpFlag = X86II::MO_TLVP_PIC_BASE;
12481 OpFlag = X86II::MO_TLVP;
12483 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12484 GA->getValueType(0),
12485 GA->getOffset(), OpFlag);
12486 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12488 // With PIC32, the address is actually $g + Offset.
12490 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12491 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12494 // Lowering the machine isd will make sure everything is in the right
12496 SDValue Chain = DAG.getEntryNode();
12497 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12498 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, DL, true), DL);
12499 SDValue Args[] = { Chain, Offset };
12500 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12502 DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true),
12503 DAG.getIntPtrConstant(0, DL, true), SDValue(), DL);
12505 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12506 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12507 MFI->setAdjustsStack(true);
12509 // And our return value (tls address) is in the standard call return value
12511 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12512 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12515 if (Subtarget->isTargetKnownWindowsMSVC() ||
12516 Subtarget->isTargetWindowsGNU()) {
12517 // Just use the implicit TLS architecture
12518 // Need to generate someting similar to:
12519 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12521 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12522 // mov rcx, qword [rdx+rcx*8]
12523 // mov eax, .tls$:tlsvar
12524 // [rax+rcx] contains the address
12525 // Windows 64bit: gs:0x58
12526 // Windows 32bit: fs:__tls_array
12529 SDValue Chain = DAG.getEntryNode();
12531 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12532 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12533 // use its literal value of 0x2C.
12534 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12535 ? Type::getInt8PtrTy(*DAG.getContext(),
12537 : Type::getInt32PtrTy(*DAG.getContext(),
12540 SDValue TlsArray = Subtarget->is64Bit()
12541 ? DAG.getIntPtrConstant(0x58, dl)
12542 : (Subtarget->isTargetWindowsGNU()
12543 ? DAG.getIntPtrConstant(0x2C, dl)
12544 : DAG.getExternalSymbol("_tls_array", PtrVT));
12546 SDValue ThreadPointer =
12547 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12551 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12552 res = ThreadPointer;
12554 // Load the _tls_index variable
12555 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12556 if (Subtarget->is64Bit())
12557 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12558 MachinePointerInfo(), MVT::i32, false, false,
12561 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12564 auto &DL = DAG.getDataLayout();
12566 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12567 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12569 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12572 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12575 // Get the offset of start of .tls section
12576 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12577 GA->getValueType(0),
12578 GA->getOffset(), X86II::MO_SECREL);
12579 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12581 // The address of the thread local variable is the add of the thread
12582 // pointer with the offset of the variable.
12583 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12586 llvm_unreachable("TLS not implemented for this target.");
12589 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12590 /// and take a 2 x i32 value to shift plus a shift amount.
12591 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12592 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12593 MVT VT = Op.getSimpleValueType();
12594 unsigned VTBits = VT.getSizeInBits();
12596 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12597 SDValue ShOpLo = Op.getOperand(0);
12598 SDValue ShOpHi = Op.getOperand(1);
12599 SDValue ShAmt = Op.getOperand(2);
12600 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12601 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12603 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12604 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12605 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12606 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12607 : DAG.getConstant(0, dl, VT);
12609 SDValue Tmp2, Tmp3;
12610 if (Op.getOpcode() == ISD::SHL_PARTS) {
12611 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12612 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12614 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12615 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12618 // If the shift amount is larger or equal than the width of a part we can't
12619 // rely on the results of shld/shrd. Insert a test and select the appropriate
12620 // values for large shift amounts.
12621 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12622 DAG.getConstant(VTBits, dl, MVT::i8));
12623 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12624 AndNode, DAG.getConstant(0, dl, MVT::i8));
12627 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12628 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12629 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12631 if (Op.getOpcode() == ISD::SHL_PARTS) {
12632 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12633 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12635 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12636 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12639 SDValue Ops[2] = { Lo, Hi };
12640 return DAG.getMergeValues(Ops, dl);
12643 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12644 SelectionDAG &DAG) const {
12645 SDValue Src = Op.getOperand(0);
12646 MVT SrcVT = Src.getSimpleValueType();
12647 MVT VT = Op.getSimpleValueType();
12650 if (SrcVT.isVector()) {
12651 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12652 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12653 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12654 DAG.getUNDEF(SrcVT)));
12656 if (SrcVT.getVectorElementType() == MVT::i1) {
12657 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12658 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12659 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12664 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12665 "Unknown SINT_TO_FP to lower!");
12667 // These are really Legal; return the operand so the caller accepts it as
12669 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12671 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12672 Subtarget->is64Bit()) {
12676 SDValue ValueToStore = Op.getOperand(0);
12677 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12678 !Subtarget->is64Bit())
12679 // Bitcasting to f64 here allows us to do a single 64-bit store from
12680 // an SSE register, avoiding the store forwarding penalty that would come
12681 // with two 32-bit stores.
12682 ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
12684 unsigned Size = SrcVT.getSizeInBits()/8;
12685 MachineFunction &MF = DAG.getMachineFunction();
12686 auto PtrVT = getPointerTy(MF.getDataLayout());
12687 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12688 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12689 SDValue Chain = DAG.getStore(
12690 DAG.getEntryNode(), dl, ValueToStore, StackSlot,
12691 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12693 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12696 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12698 SelectionDAG &DAG) const {
12702 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12704 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12706 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12708 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12710 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12711 MachineMemOperand *MMO;
12713 int SSFI = FI->getIndex();
12714 MMO = DAG.getMachineFunction().getMachineMemOperand(
12715 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12716 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12718 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12719 StackSlot = StackSlot.getOperand(1);
12721 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12722 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12724 Tys, Ops, SrcVT, MMO);
12727 Chain = Result.getValue(1);
12728 SDValue InFlag = Result.getValue(2);
12730 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12731 // shouldn't be necessary except that RFP cannot be live across
12732 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12733 MachineFunction &MF = DAG.getMachineFunction();
12734 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12735 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12736 auto PtrVT = getPointerTy(MF.getDataLayout());
12737 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12738 Tys = DAG.getVTList(MVT::Other);
12740 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12742 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12743 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12744 MachineMemOperand::MOStore, SSFISize, SSFISize);
12746 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12747 Ops, Op.getValueType(), MMO);
12748 Result = DAG.getLoad(
12749 Op.getValueType(), DL, Chain, StackSlot,
12750 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12751 false, false, false, 0);
12757 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12758 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12759 SelectionDAG &DAG) const {
12760 // This algorithm is not obvious. Here it is what we're trying to output:
12763 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12764 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12766 haddpd %xmm0, %xmm0
12768 pshufd $0x4e, %xmm0, %xmm1
12774 LLVMContext *Context = DAG.getContext();
12776 // Build some magic constants.
12777 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12778 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12779 auto PtrVT = getPointerTy(DAG.getDataLayout());
12780 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12782 SmallVector<Constant*,2> CV1;
12784 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12785 APInt(64, 0x4330000000000000ULL))));
12787 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12788 APInt(64, 0x4530000000000000ULL))));
12789 Constant *C1 = ConstantVector::get(CV1);
12790 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12792 // Load the 64-bit value into an XMM register.
12793 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12796 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12797 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12798 false, false, false, 16);
12800 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12803 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12804 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12805 false, false, false, 16);
12806 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12807 // TODO: Are there any fast-math-flags to propagate here?
12808 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12811 if (Subtarget->hasSSE3()) {
12812 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12813 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12815 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12816 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12818 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12819 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12822 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12823 DAG.getIntPtrConstant(0, dl));
12826 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12827 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12828 SelectionDAG &DAG) const {
12830 // FP constant to bias correct the final result.
12831 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12834 // Load the 32-bit value into an XMM register.
12835 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12838 // Zero out the upper parts of the register.
12839 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12841 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12842 DAG.getBitcast(MVT::v2f64, Load),
12843 DAG.getIntPtrConstant(0, dl));
12845 // Or the load with the bias.
12846 SDValue Or = DAG.getNode(
12847 ISD::OR, dl, MVT::v2i64,
12848 DAG.getBitcast(MVT::v2i64,
12849 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12850 DAG.getBitcast(MVT::v2i64,
12851 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12853 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12854 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12856 // Subtract the bias.
12857 // TODO: Are there any fast-math-flags to propagate here?
12858 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12860 // Handle final rounding.
12861 MVT DestVT = Op.getSimpleValueType();
12863 if (DestVT.bitsLT(MVT::f64))
12864 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12865 DAG.getIntPtrConstant(0, dl));
12866 if (DestVT.bitsGT(MVT::f64))
12867 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12869 // Handle final rounding.
12873 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12874 const X86Subtarget &Subtarget) {
12875 // The algorithm is the following:
12876 // #ifdef __SSE4_1__
12877 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12878 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12879 // (uint4) 0x53000000, 0xaa);
12881 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12882 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12884 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12885 // return (float4) lo + fhi;
12887 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12888 // reassociate the two FADDs, and if we do that, the algorithm fails
12889 // spectacularly (PR24512).
12890 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12891 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12892 // there's also the MachineCombiner reassociations happening on Machine IR.
12893 if (DAG.getTarget().Options.UnsafeFPMath)
12897 SDValue V = Op->getOperand(0);
12898 MVT VecIntVT = V.getSimpleValueType();
12899 bool Is128 = VecIntVT == MVT::v4i32;
12900 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12901 // If we convert to something else than the supported type, e.g., to v4f64,
12903 if (VecFloatVT != Op->getSimpleValueType(0))
12906 unsigned NumElts = VecIntVT.getVectorNumElements();
12907 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12908 "Unsupported custom type");
12909 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12911 // In the #idef/#else code, we have in common:
12912 // - The vector of constants:
12918 // Create the splat vector for 0x4b000000.
12919 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12920 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12921 CstLow, CstLow, CstLow, CstLow};
12922 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12923 makeArrayRef(&CstLowArray[0], NumElts));
12924 // Create the splat vector for 0x53000000.
12925 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12926 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12927 CstHigh, CstHigh, CstHigh, CstHigh};
12928 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12929 makeArrayRef(&CstHighArray[0], NumElts));
12931 // Create the right shift.
12932 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12933 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12934 CstShift, CstShift, CstShift, CstShift};
12935 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12936 makeArrayRef(&CstShiftArray[0], NumElts));
12937 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12940 if (Subtarget.hasSSE41()) {
12941 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12942 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12943 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12944 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12945 // Low will be bitcasted right away, so do not bother bitcasting back to its
12947 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12948 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12949 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12950 // (uint4) 0x53000000, 0xaa);
12951 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12952 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12953 // High will be bitcasted right away, so do not bother bitcasting back to
12954 // its original type.
12955 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12956 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12958 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12959 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12960 CstMask, CstMask, CstMask);
12961 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12962 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12963 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12965 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12966 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12969 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12970 SDValue CstFAdd = DAG.getConstantFP(
12971 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12972 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12973 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12974 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12975 makeArrayRef(&CstFAddArray[0], NumElts));
12977 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12978 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12979 // TODO: Are there any fast-math-flags to propagate here?
12981 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12982 // return (float4) lo + fhi;
12983 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12984 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12987 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12988 SelectionDAG &DAG) const {
12989 SDValue N0 = Op.getOperand(0);
12990 MVT SVT = N0.getSimpleValueType();
12993 switch (SVT.SimpleTy) {
12995 llvm_unreachable("Custom UINT_TO_FP is not supported!");
13000 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
13001 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
13002 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
13006 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
13009 assert(Subtarget->hasAVX512());
13010 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
13011 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
13015 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
13016 SelectionDAG &DAG) const {
13017 SDValue N0 = Op.getOperand(0);
13019 auto PtrVT = getPointerTy(DAG.getDataLayout());
13021 if (Op.getSimpleValueType().isVector())
13022 return lowerUINT_TO_FP_vec(Op, DAG);
13024 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
13025 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
13026 // the optimization here.
13027 if (DAG.SignBitIsZero(N0))
13028 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
13030 MVT SrcVT = N0.getSimpleValueType();
13031 MVT DstVT = Op.getSimpleValueType();
13033 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
13034 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
13035 // Conversions from unsigned i32 to f32/f64 are legal,
13036 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
13040 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
13041 return LowerUINT_TO_FP_i64(Op, DAG);
13042 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
13043 return LowerUINT_TO_FP_i32(Op, DAG);
13044 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
13047 // Make a 64-bit buffer, and use it to build an FILD.
13048 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
13049 if (SrcVT == MVT::i32) {
13050 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
13051 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
13052 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13053 StackSlot, MachinePointerInfo(),
13055 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
13056 OffsetSlot, MachinePointerInfo(),
13058 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
13062 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
13063 SDValue ValueToStore = Op.getOperand(0);
13064 if (isScalarFPTypeInSSEReg(Op.getValueType()) && !Subtarget->is64Bit())
13065 // Bitcasting to f64 here allows us to do a single 64-bit store from
13066 // an SSE register, avoiding the store forwarding penalty that would come
13067 // with two 32-bit stores.
13068 ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
13069 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, ValueToStore,
13070 StackSlot, MachinePointerInfo(),
13072 // For i64 source, we need to add the appropriate power of 2 if the input
13073 // was negative. This is the same as the optimization in
13074 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
13075 // we must be careful to do the computation in x87 extended precision, not
13076 // in SSE. (The generic code can't know it's OK to do this, or how to.)
13077 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
13078 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
13079 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
13080 MachineMemOperand::MOLoad, 8, 8);
13082 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
13083 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
13084 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
13087 APInt FF(32, 0x5F800000ULL);
13089 // Check whether the sign bit is set.
13090 SDValue SignSet = DAG.getSetCC(
13091 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
13092 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
13094 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
13095 SDValue FudgePtr = DAG.getConstantPool(
13096 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
13098 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
13099 SDValue Zero = DAG.getIntPtrConstant(0, dl);
13100 SDValue Four = DAG.getIntPtrConstant(4, dl);
13101 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
13103 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
13105 // Load the value out, extending it from f32 to f80.
13106 // FIXME: Avoid the extend by constructing the right constant pool?
13107 SDValue Fudge = DAG.getExtLoad(
13108 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
13109 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
13110 false, false, false, 4);
13111 // Extend everything to 80 bits to force it to be done on x87.
13112 // TODO: Are there any fast-math-flags to propagate here?
13113 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
13114 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
13115 DAG.getIntPtrConstant(0, dl));
13118 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
13119 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
13120 // just return an <SDValue(), SDValue()> pair.
13121 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
13122 // to i16, i32 or i64, and we lower it to a legal sequence.
13123 // If lowered to the final integer result we return a <result, SDValue()> pair.
13124 // Otherwise we lower it to a sequence ending with a FIST, return a
13125 // <FIST, StackSlot> pair, and the caller is responsible for loading
13126 // the final integer result from StackSlot.
13127 std::pair<SDValue,SDValue>
13128 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
13129 bool IsSigned, bool IsReplace) const {
13132 EVT DstTy = Op.getValueType();
13133 EVT TheVT = Op.getOperand(0).getValueType();
13134 auto PtrVT = getPointerTy(DAG.getDataLayout());
13136 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
13137 // f16 must be promoted before using the lowering in this routine.
13138 // fp128 does not use this lowering.
13139 return std::make_pair(SDValue(), SDValue());
13142 // If using FIST to compute an unsigned i64, we'll need some fixup
13143 // to handle values above the maximum signed i64. A FIST is always
13144 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
13145 bool UnsignedFixup = !IsSigned &&
13146 DstTy == MVT::i64 &&
13147 (!Subtarget->is64Bit() ||
13148 !isScalarFPTypeInSSEReg(TheVT));
13150 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
13151 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
13152 // The low 32 bits of the fist result will have the correct uint32 result.
13153 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
13157 assert(DstTy.getSimpleVT() <= MVT::i64 &&
13158 DstTy.getSimpleVT() >= MVT::i16 &&
13159 "Unknown FP_TO_INT to lower!");
13161 // These are really Legal.
13162 if (DstTy == MVT::i32 &&
13163 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13164 return std::make_pair(SDValue(), SDValue());
13165 if (Subtarget->is64Bit() &&
13166 DstTy == MVT::i64 &&
13167 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13168 return std::make_pair(SDValue(), SDValue());
13170 // We lower FP->int64 into FISTP64 followed by a load from a temporary
13172 MachineFunction &MF = DAG.getMachineFunction();
13173 unsigned MemSize = DstTy.getSizeInBits()/8;
13174 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13175 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13178 switch (DstTy.getSimpleVT().SimpleTy) {
13179 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
13180 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
13181 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
13182 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
13185 SDValue Chain = DAG.getEntryNode();
13186 SDValue Value = Op.getOperand(0);
13187 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
13189 if (UnsignedFixup) {
13191 // Conversion to unsigned i64 is implemented with a select,
13192 // depending on whether the source value fits in the range
13193 // of a signed i64. Let Thresh be the FP equivalent of
13194 // 0x8000000000000000ULL.
13196 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
13197 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
13198 // Fist-to-mem64 FistSrc
13199 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
13200 // to XOR'ing the high 32 bits with Adjust.
13202 // Being a power of 2, Thresh is exactly representable in all FP formats.
13203 // For X87 we'd like to use the smallest FP type for this constant, but
13204 // for DAG type consistency we have to match the FP operand type.
13206 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
13207 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
13208 bool LosesInfo = false;
13209 if (TheVT == MVT::f64)
13210 // The rounding mode is irrelevant as the conversion should be exact.
13211 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
13213 else if (TheVT == MVT::f80)
13214 Status = Thresh.convert(APFloat::x87DoubleExtended,
13215 APFloat::rmNearestTiesToEven, &LosesInfo);
13217 assert(Status == APFloat::opOK && !LosesInfo &&
13218 "FP conversion should have been exact");
13220 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
13222 SDValue Cmp = DAG.getSetCC(DL,
13223 getSetCCResultType(DAG.getDataLayout(),
13224 *DAG.getContext(), TheVT),
13225 Value, ThreshVal, ISD::SETLT);
13226 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
13227 DAG.getConstant(0, DL, MVT::i32),
13228 DAG.getConstant(0x80000000, DL, MVT::i32));
13229 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
13230 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
13231 *DAG.getContext(), TheVT),
13232 Value, ThreshVal, ISD::SETLT);
13233 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
13236 // FIXME This causes a redundant load/store if the SSE-class value is already
13237 // in memory, such as if it is on the callstack.
13238 if (isScalarFPTypeInSSEReg(TheVT)) {
13239 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13240 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13241 MachinePointerInfo::getFixedStack(MF, SSFI), false,
13243 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13245 Chain, StackSlot, DAG.getValueType(TheVT)
13248 MachineMemOperand *MMO =
13249 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13250 MachineMemOperand::MOLoad, MemSize, MemSize);
13251 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13252 Chain = Value.getValue(1);
13253 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13254 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13257 MachineMemOperand *MMO =
13258 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13259 MachineMemOperand::MOStore, MemSize, MemSize);
13261 if (UnsignedFixup) {
13263 // Insert the FIST, load its result as two i32's,
13264 // and XOR the high i32 with Adjust.
13266 SDValue FistOps[] = { Chain, Value, StackSlot };
13267 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13268 FistOps, DstTy, MMO);
13270 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
13271 MachinePointerInfo(),
13272 false, false, false, 0);
13273 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
13274 DAG.getConstant(4, DL, PtrVT));
13276 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
13277 MachinePointerInfo(),
13278 false, false, false, 0);
13279 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13281 if (Subtarget->is64Bit()) {
13282 // Join High32 and Low32 into a 64-bit result.
13283 // (High32 << 32) | Low32
13284 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13285 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13286 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13287 DAG.getConstant(32, DL, MVT::i8));
13288 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13289 return std::make_pair(Result, SDValue());
13292 SDValue ResultOps[] = { Low32, High32 };
13294 SDValue pair = IsReplace
13295 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13296 : DAG.getMergeValues(ResultOps, DL);
13297 return std::make_pair(pair, SDValue());
13299 // Build the FP_TO_INT*_IN_MEM
13300 SDValue Ops[] = { Chain, Value, StackSlot };
13301 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13303 return std::make_pair(FIST, StackSlot);
13307 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13308 const X86Subtarget *Subtarget) {
13309 MVT VT = Op->getSimpleValueType(0);
13310 SDValue In = Op->getOperand(0);
13311 MVT InVT = In.getSimpleValueType();
13314 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13315 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13317 // Optimize vectors in AVX mode:
13320 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13321 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13322 // Concat upper and lower parts.
13325 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13326 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13327 // Concat upper and lower parts.
13330 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13331 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13332 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13335 if (Subtarget->hasInt256())
13336 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13338 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13339 SDValue Undef = DAG.getUNDEF(InVT);
13340 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13341 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13342 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13344 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13345 VT.getVectorNumElements()/2);
13347 OpLo = DAG.getBitcast(HVT, OpLo);
13348 OpHi = DAG.getBitcast(HVT, OpHi);
13350 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13353 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13354 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13355 MVT VT = Op->getSimpleValueType(0);
13356 SDValue In = Op->getOperand(0);
13357 MVT InVT = In.getSimpleValueType();
13359 unsigned int NumElts = VT.getVectorNumElements();
13360 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13363 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13364 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13366 assert(InVT.getVectorElementType() == MVT::i1);
13367 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13369 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13371 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13373 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13374 if (VT.is512BitVector())
13376 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13379 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13380 SelectionDAG &DAG) {
13381 if (Subtarget->hasFp256())
13382 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13388 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13389 SelectionDAG &DAG) {
13391 MVT VT = Op.getSimpleValueType();
13392 SDValue In = Op.getOperand(0);
13393 MVT SVT = In.getSimpleValueType();
13395 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13396 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13398 if (Subtarget->hasFp256())
13399 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13402 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13403 VT.getVectorNumElements() != SVT.getVectorNumElements());
13407 static SDValue LowerTruncateVecI1(SDValue Op, SelectionDAG &DAG,
13408 const X86Subtarget *Subtarget) {
13411 MVT VT = Op.getSimpleValueType();
13412 SDValue In = Op.getOperand(0);
13413 MVT InVT = In.getSimpleValueType();
13415 assert(VT.getVectorElementType() == MVT::i1 && "Unexected vector type.");
13417 // Shift LSB to MSB and use VPMOVB2M - SKX.
13418 unsigned ShiftInx = InVT.getScalarSizeInBits() - 1;
13419 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13420 Subtarget->hasBWI()) || // legal, will go to VPMOVB2M, VPMOVW2M
13421 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13422 InVT.getScalarSizeInBits() <= 16 && Subtarget->hasBWI() &&
13423 Subtarget->hasVLX())) { // legal, will go to VPMOVB2M, VPMOVW2M
13424 // Shift packed bytes not supported natively, bitcast to dword
13425 MVT ExtVT = MVT::getVectorVT(MVT::i16, InVT.getSizeInBits()/16);
13426 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, ExtVT,
13427 DAG.getBitcast(ExtVT, In),
13428 DAG.getConstant(ShiftInx, DL, ExtVT));
13429 ShiftNode = DAG.getBitcast(InVT, ShiftNode);
13430 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13432 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13433 Subtarget->hasDQI()) || // legal, will go to VPMOVD2M, VPMOVQ2M
13434 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13435 InVT.getScalarSizeInBits() >= 32 && Subtarget->hasDQI() &&
13436 Subtarget->hasVLX())) { // legal, will go to VPMOVD2M, VPMOVQ2M
13438 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13439 DAG.getConstant(ShiftInx, DL, InVT));
13440 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13443 // Shift LSB to MSB, extend if necessary and use TESTM.
13444 unsigned NumElts = InVT.getVectorNumElements();
13445 if (InVT.getSizeInBits() < 512 &&
13446 (InVT.getScalarType() == MVT::i8 || InVT.getScalarType() == MVT::i16 ||
13447 !Subtarget->hasVLX())) {
13448 assert((NumElts == 8 || NumElts == 16) && "Unexected vector type.");
13450 // TESTD/Q should be used (if BW supported we use CVT2MASK above),
13451 // so vector should be extended to packed dword/qword.
13452 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts);
13453 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13455 ShiftInx = InVT.getScalarSizeInBits() - 1;
13458 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13459 DAG.getConstant(ShiftInx, DL, InVT));
13460 return DAG.getNode(X86ISD::TESTM, DL, VT, ShiftNode, ShiftNode);
13463 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13465 MVT VT = Op.getSimpleValueType();
13466 SDValue In = Op.getOperand(0);
13467 MVT InVT = In.getSimpleValueType();
13469 if (VT == MVT::i1) {
13470 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13471 "Invalid scalar TRUNCATE operation");
13472 if (InVT.getSizeInBits() >= 32)
13474 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13475 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13477 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13478 "Invalid TRUNCATE operation");
13480 if (VT.getVectorElementType() == MVT::i1)
13481 return LowerTruncateVecI1(Op, DAG, Subtarget);
13483 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13484 if (Subtarget->hasAVX512()) {
13485 // word to byte only under BWI
13486 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13487 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13488 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13489 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13491 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13492 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13493 if (Subtarget->hasInt256()) {
13494 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13495 In = DAG.getBitcast(MVT::v8i32, In);
13496 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13498 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13499 DAG.getIntPtrConstant(0, DL));
13502 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13503 DAG.getIntPtrConstant(0, DL));
13504 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13505 DAG.getIntPtrConstant(2, DL));
13506 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13507 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13508 static const int ShufMask[] = {0, 2, 4, 6};
13509 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13512 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13513 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13514 if (Subtarget->hasInt256()) {
13515 In = DAG.getBitcast(MVT::v32i8, In);
13517 SmallVector<SDValue,32> pshufbMask;
13518 for (unsigned i = 0; i < 2; ++i) {
13519 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13520 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13521 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13522 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13523 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13524 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13525 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13526 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13527 for (unsigned j = 0; j < 8; ++j)
13528 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13530 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13531 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13532 In = DAG.getBitcast(MVT::v4i64, In);
13534 static const int ShufMask[] = {0, 2, -1, -1};
13535 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13537 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13538 DAG.getIntPtrConstant(0, DL));
13539 return DAG.getBitcast(VT, In);
13542 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13543 DAG.getIntPtrConstant(0, DL));
13545 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13546 DAG.getIntPtrConstant(4, DL));
13548 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13549 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13551 // The PSHUFB mask:
13552 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13553 -1, -1, -1, -1, -1, -1, -1, -1};
13555 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13556 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13557 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13559 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13560 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13562 // The MOVLHPS Mask:
13563 static const int ShufMask2[] = {0, 1, 4, 5};
13564 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13565 return DAG.getBitcast(MVT::v8i16, res);
13568 // Handle truncation of V256 to V128 using shuffles.
13569 if (!VT.is128BitVector() || !InVT.is256BitVector())
13572 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13574 unsigned NumElems = VT.getVectorNumElements();
13575 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13577 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13578 // Prepare truncation shuffle mask
13579 for (unsigned i = 0; i != NumElems; ++i)
13580 MaskVec[i] = i * 2;
13581 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13582 DAG.getUNDEF(NVT), &MaskVec[0]);
13583 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13584 DAG.getIntPtrConstant(0, DL));
13587 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13588 SelectionDAG &DAG) const {
13589 assert(!Op.getSimpleValueType().isVector());
13591 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13592 /*IsSigned=*/ true, /*IsReplace=*/ false);
13593 SDValue FIST = Vals.first, StackSlot = Vals.second;
13594 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13595 if (!FIST.getNode())
13598 if (StackSlot.getNode())
13599 // Load the result.
13600 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13601 FIST, StackSlot, MachinePointerInfo(),
13602 false, false, false, 0);
13604 // The node is the result.
13608 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13609 SelectionDAG &DAG) const {
13610 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13611 /*IsSigned=*/ false, /*IsReplace=*/ false);
13612 SDValue FIST = Vals.first, StackSlot = Vals.second;
13613 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13614 if (!FIST.getNode())
13617 if (StackSlot.getNode())
13618 // Load the result.
13619 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13620 FIST, StackSlot, MachinePointerInfo(),
13621 false, false, false, 0);
13623 // The node is the result.
13627 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13629 MVT VT = Op.getSimpleValueType();
13630 SDValue In = Op.getOperand(0);
13631 MVT SVT = In.getSimpleValueType();
13633 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13635 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13636 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13637 In, DAG.getUNDEF(SVT)));
13640 /// The only differences between FABS and FNEG are the mask and the logic op.
13641 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13642 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13643 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13644 "Wrong opcode for lowering FABS or FNEG.");
13646 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13648 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13649 // into an FNABS. We'll lower the FABS after that if it is still in use.
13651 for (SDNode *User : Op->uses())
13652 if (User->getOpcode() == ISD::FNEG)
13656 MVT VT = Op.getSimpleValueType();
13658 bool IsF128 = (VT == MVT::f128);
13660 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13661 // decide if we should generate a 16-byte constant mask when we only need 4 or
13662 // 8 bytes for the scalar case.
13668 if (VT.isVector()) {
13670 EltVT = VT.getVectorElementType();
13671 NumElts = VT.getVectorNumElements();
13672 } else if (IsF128) {
13673 // SSE instructions are used for optimized f128 logical operations.
13674 LogicVT = MVT::f128;
13678 // There are no scalar bitwise logical SSE/AVX instructions, so we
13679 // generate a 16-byte vector constant and logic op even for the scalar case.
13680 // Using a 16-byte mask allows folding the load of the mask with
13681 // the logic op, so it can save (~4 bytes) on code size.
13682 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13684 NumElts = (VT == MVT::f64) ? 2 : 4;
13687 unsigned EltBits = EltVT.getSizeInBits();
13688 LLVMContext *Context = DAG.getContext();
13689 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13691 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13692 Constant *C = ConstantInt::get(*Context, MaskElt);
13693 C = ConstantVector::getSplat(NumElts, C);
13694 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13695 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13696 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13698 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13699 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13700 false, false, false, Alignment);
13702 SDValue Op0 = Op.getOperand(0);
13703 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13705 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13706 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13708 if (VT.isVector() || IsF128)
13709 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13711 // For the scalar case extend to a 128-bit vector, perform the logic op,
13712 // and extract the scalar result back out.
13713 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13714 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13715 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13716 DAG.getIntPtrConstant(0, dl));
13719 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13721 LLVMContext *Context = DAG.getContext();
13722 SDValue Op0 = Op.getOperand(0);
13723 SDValue Op1 = Op.getOperand(1);
13725 MVT VT = Op.getSimpleValueType();
13726 MVT SrcVT = Op1.getSimpleValueType();
13727 bool IsF128 = (VT == MVT::f128);
13729 // If second operand is smaller, extend it first.
13730 if (SrcVT.bitsLT(VT)) {
13731 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13734 // And if it is bigger, shrink it first.
13735 if (SrcVT.bitsGT(VT)) {
13736 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13740 // At this point the operands and the result should have the same
13741 // type, and that won't be f80 since that is not custom lowered.
13742 assert((VT == MVT::f64 || VT == MVT::f32 || IsF128) &&
13743 "Unexpected type in LowerFCOPYSIGN");
13745 const fltSemantics &Sem =
13746 VT == MVT::f64 ? APFloat::IEEEdouble :
13747 (IsF128 ? APFloat::IEEEquad : APFloat::IEEEsingle);
13748 const unsigned SizeInBits = VT.getSizeInBits();
13750 SmallVector<Constant *, 4> CV(
13751 VT == MVT::f64 ? 2 : (IsF128 ? 1 : 4),
13752 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13754 // First, clear all bits but the sign bit from the second operand (sign).
13755 CV[0] = ConstantFP::get(*Context,
13756 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13757 Constant *C = ConstantVector::get(CV);
13758 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13759 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13761 // Perform all logic operations as 16-byte vectors because there are no
13762 // scalar FP logic instructions in SSE. This allows load folding of the
13763 // constants into the logic instructions.
13764 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : (IsF128 ? MVT::f128 : MVT::v4f32);
13766 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13767 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13768 false, false, false, 16);
13770 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13771 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13773 // Next, clear the sign bit from the first operand (magnitude).
13774 // If it's a constant, we can clear it here.
13775 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13776 APFloat APF = Op0CN->getValueAPF();
13777 // If the magnitude is a positive zero, the sign bit alone is enough.
13778 if (APF.isPosZero())
13779 return IsF128 ? SignBit :
13780 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13781 DAG.getIntPtrConstant(0, dl));
13783 CV[0] = ConstantFP::get(*Context, APF);
13785 CV[0] = ConstantFP::get(
13787 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13789 C = ConstantVector::get(CV);
13790 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13792 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13793 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13794 false, false, false, 16);
13795 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13796 if (!isa<ConstantFPSDNode>(Op0)) {
13798 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13799 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13801 // OR the magnitude value with the sign bit.
13802 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13803 return IsF128 ? Val :
13804 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13805 DAG.getIntPtrConstant(0, dl));
13808 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13809 SDValue N0 = Op.getOperand(0);
13811 MVT VT = Op.getSimpleValueType();
13813 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13814 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13815 DAG.getConstant(1, dl, VT));
13816 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13819 // Check whether an OR'd tree is PTEST-able.
13820 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13821 SelectionDAG &DAG) {
13822 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13824 if (!Subtarget->hasSSE41())
13827 if (!Op->hasOneUse())
13830 SDNode *N = Op.getNode();
13833 SmallVector<SDValue, 8> Opnds;
13834 DenseMap<SDValue, unsigned> VecInMap;
13835 SmallVector<SDValue, 8> VecIns;
13836 EVT VT = MVT::Other;
13838 // Recognize a special case where a vector is casted into wide integer to
13840 Opnds.push_back(N->getOperand(0));
13841 Opnds.push_back(N->getOperand(1));
13843 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13844 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13845 // BFS traverse all OR'd operands.
13846 if (I->getOpcode() == ISD::OR) {
13847 Opnds.push_back(I->getOperand(0));
13848 Opnds.push_back(I->getOperand(1));
13849 // Re-evaluate the number of nodes to be traversed.
13850 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13854 // Quit if a non-EXTRACT_VECTOR_ELT
13855 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13858 // Quit if without a constant index.
13859 SDValue Idx = I->getOperand(1);
13860 if (!isa<ConstantSDNode>(Idx))
13863 SDValue ExtractedFromVec = I->getOperand(0);
13864 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13865 if (M == VecInMap.end()) {
13866 VT = ExtractedFromVec.getValueType();
13867 // Quit if not 128/256-bit vector.
13868 if (!VT.is128BitVector() && !VT.is256BitVector())
13870 // Quit if not the same type.
13871 if (VecInMap.begin() != VecInMap.end() &&
13872 VT != VecInMap.begin()->first.getValueType())
13874 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13875 VecIns.push_back(ExtractedFromVec);
13877 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13880 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13881 "Not extracted from 128-/256-bit vector.");
13883 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13885 for (DenseMap<SDValue, unsigned>::const_iterator
13886 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13887 // Quit if not all elements are used.
13888 if (I->second != FullMask)
13892 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13894 // Cast all vectors into TestVT for PTEST.
13895 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13896 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13898 // If more than one full vectors are evaluated, OR them first before PTEST.
13899 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13900 // Each iteration will OR 2 nodes and append the result until there is only
13901 // 1 node left, i.e. the final OR'd value of all vectors.
13902 SDValue LHS = VecIns[Slot];
13903 SDValue RHS = VecIns[Slot + 1];
13904 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13907 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13908 VecIns.back(), VecIns.back());
13911 /// \brief return true if \c Op has a use that doesn't just read flags.
13912 static bool hasNonFlagsUse(SDValue Op) {
13913 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13915 SDNode *User = *UI;
13916 unsigned UOpNo = UI.getOperandNo();
13917 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13918 // Look pass truncate.
13919 UOpNo = User->use_begin().getOperandNo();
13920 User = *User->use_begin();
13923 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13924 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13930 /// Emit nodes that will be selected as "test Op0,Op0", or something
13932 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13933 SelectionDAG &DAG) const {
13934 if (Op.getValueType() == MVT::i1) {
13935 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13936 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13937 DAG.getConstant(0, dl, MVT::i8));
13939 // CF and OF aren't always set the way we want. Determine which
13940 // of these we need.
13941 bool NeedCF = false;
13942 bool NeedOF = false;
13945 case X86::COND_A: case X86::COND_AE:
13946 case X86::COND_B: case X86::COND_BE:
13949 case X86::COND_G: case X86::COND_GE:
13950 case X86::COND_L: case X86::COND_LE:
13951 case X86::COND_O: case X86::COND_NO: {
13952 // Check if we really need to set the
13953 // Overflow flag. If NoSignedWrap is present
13954 // that is not actually needed.
13955 switch (Op->getOpcode()) {
13960 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13961 if (BinNode->Flags.hasNoSignedWrap())
13971 // See if we can use the EFLAGS value from the operand instead of
13972 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13973 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13974 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13975 // Emit a CMP with 0, which is the TEST pattern.
13976 //if (Op.getValueType() == MVT::i1)
13977 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13978 // DAG.getConstant(0, MVT::i1));
13979 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13980 DAG.getConstant(0, dl, Op.getValueType()));
13982 unsigned Opcode = 0;
13983 unsigned NumOperands = 0;
13985 // Truncate operations may prevent the merge of the SETCC instruction
13986 // and the arithmetic instruction before it. Attempt to truncate the operands
13987 // of the arithmetic instruction and use a reduced bit-width instruction.
13988 bool NeedTruncation = false;
13989 SDValue ArithOp = Op;
13990 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13991 SDValue Arith = Op->getOperand(0);
13992 // Both the trunc and the arithmetic op need to have one user each.
13993 if (Arith->hasOneUse())
13994 switch (Arith.getOpcode()) {
14001 NeedTruncation = true;
14007 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
14008 // which may be the result of a CAST. We use the variable 'Op', which is the
14009 // non-casted variable when we check for possible users.
14010 switch (ArithOp.getOpcode()) {
14012 // Due to an isel shortcoming, be conservative if this add is likely to be
14013 // selected as part of a load-modify-store instruction. When the root node
14014 // in a match is a store, isel doesn't know how to remap non-chain non-flag
14015 // uses of other nodes in the match, such as the ADD in this case. This
14016 // leads to the ADD being left around and reselected, with the result being
14017 // two adds in the output. Alas, even if none our users are stores, that
14018 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
14019 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
14020 // climbing the DAG back to the root, and it doesn't seem to be worth the
14022 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14023 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14024 if (UI->getOpcode() != ISD::CopyToReg &&
14025 UI->getOpcode() != ISD::SETCC &&
14026 UI->getOpcode() != ISD::STORE)
14029 if (ConstantSDNode *C =
14030 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
14031 // An add of one will be selected as an INC.
14032 if (C->isOne() && !Subtarget->slowIncDec()) {
14033 Opcode = X86ISD::INC;
14038 // An add of negative one (subtract of one) will be selected as a DEC.
14039 if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
14040 Opcode = X86ISD::DEC;
14046 // Otherwise use a regular EFLAGS-setting add.
14047 Opcode = X86ISD::ADD;
14052 // If we have a constant logical shift that's only used in a comparison
14053 // against zero turn it into an equivalent AND. This allows turning it into
14054 // a TEST instruction later.
14055 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
14056 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
14057 EVT VT = Op.getValueType();
14058 unsigned BitWidth = VT.getSizeInBits();
14059 unsigned ShAmt = Op->getConstantOperandVal(1);
14060 if (ShAmt >= BitWidth) // Avoid undefined shifts.
14062 APInt Mask = ArithOp.getOpcode() == ISD::SRL
14063 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
14064 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
14065 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
14067 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
14068 DAG.getConstant(Mask, dl, VT));
14069 DAG.ReplaceAllUsesWith(Op, New);
14075 // If the primary and result isn't used, don't bother using X86ISD::AND,
14076 // because a TEST instruction will be better.
14077 if (!hasNonFlagsUse(Op))
14083 // Due to the ISEL shortcoming noted above, be conservative if this op is
14084 // likely to be selected as part of a load-modify-store instruction.
14085 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14086 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14087 if (UI->getOpcode() == ISD::STORE)
14090 // Otherwise use a regular EFLAGS-setting instruction.
14091 switch (ArithOp.getOpcode()) {
14092 default: llvm_unreachable("unexpected operator!");
14093 case ISD::SUB: Opcode = X86ISD::SUB; break;
14094 case ISD::XOR: Opcode = X86ISD::XOR; break;
14095 case ISD::AND: Opcode = X86ISD::AND; break;
14097 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
14098 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
14099 if (EFLAGS.getNode())
14102 Opcode = X86ISD::OR;
14116 return SDValue(Op.getNode(), 1);
14122 // If we found that truncation is beneficial, perform the truncation and
14124 if (NeedTruncation) {
14125 EVT VT = Op.getValueType();
14126 SDValue WideVal = Op->getOperand(0);
14127 EVT WideVT = WideVal.getValueType();
14128 unsigned ConvertedOp = 0;
14129 // Use a target machine opcode to prevent further DAGCombine
14130 // optimizations that may separate the arithmetic operations
14131 // from the setcc node.
14132 switch (WideVal.getOpcode()) {
14134 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
14135 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
14136 case ISD::AND: ConvertedOp = X86ISD::AND; break;
14137 case ISD::OR: ConvertedOp = X86ISD::OR; break;
14138 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
14142 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14143 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
14144 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
14145 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
14146 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
14152 // Emit a CMP with 0, which is the TEST pattern.
14153 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14154 DAG.getConstant(0, dl, Op.getValueType()));
14156 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14157 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
14159 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
14160 DAG.ReplaceAllUsesWith(Op, New);
14161 return SDValue(New.getNode(), 1);
14164 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
14166 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
14167 SDLoc dl, SelectionDAG &DAG) const {
14168 if (isNullConstant(Op1))
14169 return EmitTest(Op0, X86CC, dl, DAG);
14171 assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
14172 "Unexpected comparison operation for MVT::i1 operands");
14174 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
14175 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
14176 // Do the comparison at i32 if it's smaller, besides the Atom case.
14177 // This avoids subregister aliasing issues. Keep the smaller reference
14178 // if we're optimizing for size, however, as that'll allow better folding
14179 // of memory operations.
14180 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
14181 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
14182 !Subtarget->isAtom()) {
14183 unsigned ExtendOp =
14184 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
14185 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
14186 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
14188 // Use SUB instead of CMP to enable CSE between SUB and CMP.
14189 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
14190 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
14192 return SDValue(Sub.getNode(), 1);
14194 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
14197 /// Convert a comparison if required by the subtarget.
14198 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
14199 SelectionDAG &DAG) const {
14200 // If the subtarget does not support the FUCOMI instruction, floating-point
14201 // comparisons have to be converted.
14202 if (Subtarget->hasCMov() ||
14203 Cmp.getOpcode() != X86ISD::CMP ||
14204 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
14205 !Cmp.getOperand(1).getValueType().isFloatingPoint())
14208 // The instruction selector will select an FUCOM instruction instead of
14209 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
14210 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
14211 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
14213 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
14214 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
14215 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
14216 DAG.getConstant(8, dl, MVT::i8));
14217 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
14219 // Some 64-bit targets lack SAHF support, but they do support FCOMI.
14220 assert(Subtarget->hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?");
14221 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
14224 /// The minimum architected relative accuracy is 2^-12. We need one
14225 /// Newton-Raphson step to have a good float result (24 bits of precision).
14226 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
14227 DAGCombinerInfo &DCI,
14228 unsigned &RefinementSteps,
14229 bool &UseOneConstNR) const {
14230 EVT VT = Op.getValueType();
14231 const char *RecipOp;
14233 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
14234 // TODO: Add support for AVX512 (v16f32).
14235 // It is likely not profitable to do this for f64 because a double-precision
14236 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
14237 // instructions: convert to single, rsqrtss, convert back to double, refine
14238 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
14239 // along with FMA, this could be a throughput win.
14240 if (VT == MVT::f32 && Subtarget->hasSSE1())
14242 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14243 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14244 RecipOp = "vec-sqrtf";
14248 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14249 if (!Recips.isEnabled(RecipOp))
14252 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14253 UseOneConstNR = false;
14254 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
14257 /// The minimum architected relative accuracy is 2^-12. We need one
14258 /// Newton-Raphson step to have a good float result (24 bits of precision).
14259 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
14260 DAGCombinerInfo &DCI,
14261 unsigned &RefinementSteps) const {
14262 EVT VT = Op.getValueType();
14263 const char *RecipOp;
14265 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
14266 // TODO: Add support for AVX512 (v16f32).
14267 // It is likely not profitable to do this for f64 because a double-precision
14268 // reciprocal estimate with refinement on x86 prior to FMA requires
14269 // 15 instructions: convert to single, rcpss, convert back to double, refine
14270 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
14271 // along with FMA, this could be a throughput win.
14272 if (VT == MVT::f32 && Subtarget->hasSSE1())
14274 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14275 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14276 RecipOp = "vec-divf";
14280 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14281 if (!Recips.isEnabled(RecipOp))
14284 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14285 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
14288 /// If we have at least two divisions that use the same divisor, convert to
14289 /// multplication by a reciprocal. This may need to be adjusted for a given
14290 /// CPU if a division's cost is not at least twice the cost of a multiplication.
14291 /// This is because we still need one division to calculate the reciprocal and
14292 /// then we need two multiplies by that reciprocal as replacements for the
14293 /// original divisions.
14294 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
14298 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
14299 /// if it's possible.
14300 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
14301 SDLoc dl, SelectionDAG &DAG) const {
14302 SDValue Op0 = And.getOperand(0);
14303 SDValue Op1 = And.getOperand(1);
14304 if (Op0.getOpcode() == ISD::TRUNCATE)
14305 Op0 = Op0.getOperand(0);
14306 if (Op1.getOpcode() == ISD::TRUNCATE)
14307 Op1 = Op1.getOperand(0);
14310 if (Op1.getOpcode() == ISD::SHL)
14311 std::swap(Op0, Op1);
14312 if (Op0.getOpcode() == ISD::SHL) {
14313 if (isOneConstant(Op0.getOperand(0))) {
14314 // If we looked past a truncate, check that it's only truncating away
14316 unsigned BitWidth = Op0.getValueSizeInBits();
14317 unsigned AndBitWidth = And.getValueSizeInBits();
14318 if (BitWidth > AndBitWidth) {
14320 DAG.computeKnownBits(Op0, Zeros, Ones);
14321 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14325 RHS = Op0.getOperand(1);
14327 } else if (Op1.getOpcode() == ISD::Constant) {
14328 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14329 uint64_t AndRHSVal = AndRHS->getZExtValue();
14330 SDValue AndLHS = Op0;
14332 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14333 LHS = AndLHS.getOperand(0);
14334 RHS = AndLHS.getOperand(1);
14337 // Use BT if the immediate can't be encoded in a TEST instruction.
14338 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14340 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14344 if (LHS.getNode()) {
14345 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14346 // instruction. Since the shift amount is in-range-or-undefined, we know
14347 // that doing a bittest on the i32 value is ok. We extend to i32 because
14348 // the encoding for the i16 version is larger than the i32 version.
14349 // Also promote i16 to i32 for performance / code size reason.
14350 if (LHS.getValueType() == MVT::i8 ||
14351 LHS.getValueType() == MVT::i16)
14352 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14354 // If the operand types disagree, extend the shift amount to match. Since
14355 // BT ignores high bits (like shifts) we can use anyextend.
14356 if (LHS.getValueType() != RHS.getValueType())
14357 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14359 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14360 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14361 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14362 DAG.getConstant(Cond, dl, MVT::i8), BT);
14368 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14370 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14375 // SSE Condition code mapping:
14384 switch (SetCCOpcode) {
14385 default: llvm_unreachable("Unexpected SETCC condition");
14387 case ISD::SETEQ: SSECC = 0; break;
14389 case ISD::SETGT: Swap = true; // Fallthrough
14391 case ISD::SETOLT: SSECC = 1; break;
14393 case ISD::SETGE: Swap = true; // Fallthrough
14395 case ISD::SETOLE: SSECC = 2; break;
14396 case ISD::SETUO: SSECC = 3; break;
14398 case ISD::SETNE: SSECC = 4; break;
14399 case ISD::SETULE: Swap = true; // Fallthrough
14400 case ISD::SETUGE: SSECC = 5; break;
14401 case ISD::SETULT: Swap = true; // Fallthrough
14402 case ISD::SETUGT: SSECC = 6; break;
14403 case ISD::SETO: SSECC = 7; break;
14405 case ISD::SETONE: SSECC = 8; break;
14408 std::swap(Op0, Op1);
14413 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14414 // ones, and then concatenate the result back.
14415 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14416 MVT VT = Op.getSimpleValueType();
14418 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14419 "Unsupported value type for operation");
14421 unsigned NumElems = VT.getVectorNumElements();
14423 SDValue CC = Op.getOperand(2);
14425 // Extract the LHS vectors
14426 SDValue LHS = Op.getOperand(0);
14427 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14428 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14430 // Extract the RHS vectors
14431 SDValue RHS = Op.getOperand(1);
14432 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14433 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14435 // Issue the operation on the smaller types and concatenate the result back
14436 MVT EltVT = VT.getVectorElementType();
14437 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14438 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14439 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14440 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14443 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14444 SDValue Op0 = Op.getOperand(0);
14445 SDValue Op1 = Op.getOperand(1);
14446 SDValue CC = Op.getOperand(2);
14447 MVT VT = Op.getSimpleValueType();
14450 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14451 "Unexpected type for boolean compare operation");
14452 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14453 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14454 DAG.getConstant(-1, dl, VT));
14455 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14456 DAG.getConstant(-1, dl, VT));
14457 switch (SetCCOpcode) {
14458 default: llvm_unreachable("Unexpected SETCC condition");
14460 // (x == y) -> ~(x ^ y)
14461 return DAG.getNode(ISD::XOR, dl, VT,
14462 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14463 DAG.getConstant(-1, dl, VT));
14465 // (x != y) -> (x ^ y)
14466 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14469 // (x > y) -> (x & ~y)
14470 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14473 // (x < y) -> (~x & y)
14474 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14477 // (x <= y) -> (~x | y)
14478 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14481 // (x >=y) -> (x | ~y)
14482 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14486 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14487 const X86Subtarget *Subtarget) {
14488 SDValue Op0 = Op.getOperand(0);
14489 SDValue Op1 = Op.getOperand(1);
14490 SDValue CC = Op.getOperand(2);
14491 MVT VT = Op.getSimpleValueType();
14494 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14495 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14496 "Cannot set masked compare for this operation");
14498 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14500 bool Unsigned = false;
14503 switch (SetCCOpcode) {
14504 default: llvm_unreachable("Unexpected SETCC condition");
14505 case ISD::SETNE: SSECC = 4; break;
14506 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14507 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14508 case ISD::SETLT: Swap = true; //fall-through
14509 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14510 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14511 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14512 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14513 case ISD::SETULE: Unsigned = true; //fall-through
14514 case ISD::SETLE: SSECC = 2; break;
14518 std::swap(Op0, Op1);
14520 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14521 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14522 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14523 DAG.getConstant(SSECC, dl, MVT::i8));
14526 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14527 /// operand \p Op1. If non-trivial (for example because it's not constant)
14528 /// return an empty value.
14529 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14531 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14535 MVT VT = Op1.getSimpleValueType();
14536 MVT EVT = VT.getVectorElementType();
14537 unsigned n = VT.getVectorNumElements();
14538 SmallVector<SDValue, 8> ULTOp1;
14540 for (unsigned i = 0; i < n; ++i) {
14541 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14542 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14545 // Avoid underflow.
14546 APInt Val = Elt->getAPIntValue();
14550 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14553 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14556 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14557 SelectionDAG &DAG) {
14558 SDValue Op0 = Op.getOperand(0);
14559 SDValue Op1 = Op.getOperand(1);
14560 SDValue CC = Op.getOperand(2);
14561 MVT VT = Op.getSimpleValueType();
14562 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14563 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14568 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14569 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14572 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14573 unsigned Opc = X86ISD::CMPP;
14574 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14575 assert(VT.getVectorNumElements() <= 16);
14576 Opc = X86ISD::CMPM;
14578 // In the two special cases we can't handle, emit two comparisons.
14581 unsigned CombineOpc;
14582 if (SetCCOpcode == ISD::SETUEQ) {
14583 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14585 assert(SetCCOpcode == ISD::SETONE);
14586 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14589 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14590 DAG.getConstant(CC0, dl, MVT::i8));
14591 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14592 DAG.getConstant(CC1, dl, MVT::i8));
14593 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14595 // Handle all other FP comparisons here.
14596 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14597 DAG.getConstant(SSECC, dl, MVT::i8));
14600 MVT VTOp0 = Op0.getSimpleValueType();
14601 assert(VTOp0 == Op1.getSimpleValueType() &&
14602 "Expected operands with same type!");
14603 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14604 "Invalid number of packed elements for source and destination!");
14606 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14607 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14608 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14609 // legalizer firstly checks if the first operand in input to the setcc has
14610 // a legal type. If so, then it promotes the return type to that same type.
14611 // Otherwise, the return type is promoted to the 'next legal type' which,
14612 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14614 // We reach this code only if the following two conditions are met:
14615 // 1. Both return type and operand type have been promoted to wider types
14616 // by the type legalizer.
14617 // 2. The original operand type has been promoted to a 256-bit vector.
14619 // Note that condition 2. only applies for AVX targets.
14620 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14621 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14624 // The non-AVX512 code below works under the assumption that source and
14625 // destination types are the same.
14626 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14627 "Value types for source and destination must be the same!");
14629 // Break 256-bit integer vector compare into smaller ones.
14630 if (VT.is256BitVector() && !Subtarget->hasInt256())
14631 return Lower256IntVSETCC(Op, DAG);
14633 MVT OpVT = Op1.getSimpleValueType();
14634 if (OpVT.getVectorElementType() == MVT::i1)
14635 return LowerBoolVSETCC_AVX512(Op, DAG);
14637 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14638 if (Subtarget->hasAVX512()) {
14639 if (Op1.getSimpleValueType().is512BitVector() ||
14640 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14641 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14642 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14644 // In AVX-512 architecture setcc returns mask with i1 elements,
14645 // But there is no compare instruction for i8 and i16 elements in KNL.
14646 // We are not talking about 512-bit operands in this case, these
14647 // types are illegal.
14649 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14650 OpVT.getVectorElementType().getSizeInBits() >= 8))
14651 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14652 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14655 // Lower using XOP integer comparisons.
14656 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14657 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14658 // Translate compare code to XOP PCOM compare mode.
14659 unsigned CmpMode = 0;
14660 switch (SetCCOpcode) {
14661 default: llvm_unreachable("Unexpected SETCC condition");
14663 case ISD::SETLT: CmpMode = 0x00; break;
14665 case ISD::SETLE: CmpMode = 0x01; break;
14667 case ISD::SETGT: CmpMode = 0x02; break;
14669 case ISD::SETGE: CmpMode = 0x03; break;
14670 case ISD::SETEQ: CmpMode = 0x04; break;
14671 case ISD::SETNE: CmpMode = 0x05; break;
14674 // Are we comparing unsigned or signed integers?
14675 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14676 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14678 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14679 DAG.getConstant(CmpMode, dl, MVT::i8));
14682 // We are handling one of the integer comparisons here. Since SSE only has
14683 // GT and EQ comparisons for integer, swapping operands and multiple
14684 // operations may be required for some comparisons.
14686 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14687 bool Subus = false;
14689 switch (SetCCOpcode) {
14690 default: llvm_unreachable("Unexpected SETCC condition");
14691 case ISD::SETNE: Invert = true;
14692 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14693 case ISD::SETLT: Swap = true;
14694 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14695 case ISD::SETGE: Swap = true;
14696 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14697 Invert = true; break;
14698 case ISD::SETULT: Swap = true;
14699 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14700 FlipSigns = true; break;
14701 case ISD::SETUGE: Swap = true;
14702 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14703 FlipSigns = true; Invert = true; break;
14706 // Special case: Use min/max operations for SETULE/SETUGE
14707 MVT VET = VT.getVectorElementType();
14709 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14710 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14713 switch (SetCCOpcode) {
14715 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14716 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14719 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14722 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14723 if (!MinMax && hasSubus) {
14724 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14726 // t = psubus Op0, Op1
14727 // pcmpeq t, <0..0>
14728 switch (SetCCOpcode) {
14730 case ISD::SETULT: {
14731 // If the comparison is against a constant we can turn this into a
14732 // setule. With psubus, setule does not require a swap. This is
14733 // beneficial because the constant in the register is no longer
14734 // destructed as the destination so it can be hoisted out of a loop.
14735 // Only do this pre-AVX since vpcmp* is no longer destructive.
14736 if (Subtarget->hasAVX())
14738 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14739 if (ULEOp1.getNode()) {
14741 Subus = true; Invert = false; Swap = false;
14745 // Psubus is better than flip-sign because it requires no inversion.
14746 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14747 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14751 Opc = X86ISD::SUBUS;
14757 std::swap(Op0, Op1);
14759 // Check that the operation in question is available (most are plain SSE2,
14760 // but PCMPGTQ and PCMPEQQ have different requirements).
14761 if (VT == MVT::v2i64) {
14762 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14763 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14765 // First cast everything to the right type.
14766 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14767 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14769 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14770 // bits of the inputs before performing those operations. The lower
14771 // compare is always unsigned.
14774 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14776 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14777 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14778 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14779 Sign, Zero, Sign, Zero);
14781 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14782 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14784 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14785 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14786 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14788 // Create masks for only the low parts/high parts of the 64 bit integers.
14789 static const int MaskHi[] = { 1, 1, 3, 3 };
14790 static const int MaskLo[] = { 0, 0, 2, 2 };
14791 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14792 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14793 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14795 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14796 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14799 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14801 return DAG.getBitcast(VT, Result);
14804 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14805 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14806 // pcmpeqd + pshufd + pand.
14807 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14809 // First cast everything to the right type.
14810 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14811 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14814 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14816 // Make sure the lower and upper halves are both all-ones.
14817 static const int Mask[] = { 1, 0, 3, 2 };
14818 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14819 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14822 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14824 return DAG.getBitcast(VT, Result);
14828 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14829 // bits of the inputs before performing those operations.
14831 MVT EltVT = VT.getVectorElementType();
14832 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14834 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14835 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14838 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14840 // If the logical-not of the result is required, perform that now.
14842 Result = DAG.getNOT(dl, Result, VT);
14845 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14848 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14849 getZeroVector(VT, Subtarget, DAG, dl));
14854 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14856 MVT VT = Op.getSimpleValueType();
14858 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14860 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14861 && "SetCC type must be 8-bit or 1-bit integer");
14862 SDValue Op0 = Op.getOperand(0);
14863 SDValue Op1 = Op.getOperand(1);
14865 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14867 // Optimize to BT if possible.
14868 // Lower (X & (1 << N)) == 0 to BT(X, N).
14869 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14870 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14871 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14872 isNullConstant(Op1) &&
14873 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14874 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14876 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14881 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14883 if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
14884 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14886 // If the input is a setcc, then reuse the input setcc or use a new one with
14887 // the inverted condition.
14888 if (Op0.getOpcode() == X86ISD::SETCC) {
14889 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14890 bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
14894 CCode = X86::GetOppositeBranchCondition(CCode);
14895 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14896 DAG.getConstant(CCode, dl, MVT::i8),
14897 Op0.getOperand(1));
14899 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14903 if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
14904 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14906 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14907 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14910 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14911 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14912 if (X86CC == X86::COND_INVALID)
14915 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14916 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14917 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14918 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14920 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14924 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14925 SDValue LHS = Op.getOperand(0);
14926 SDValue RHS = Op.getOperand(1);
14927 SDValue Carry = Op.getOperand(2);
14928 SDValue Cond = Op.getOperand(3);
14931 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14932 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14934 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14935 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14936 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14937 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14938 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14941 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14942 static bool isX86LogicalCmp(SDValue Op) {
14943 unsigned Opc = Op.getNode()->getOpcode();
14944 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14945 Opc == X86ISD::SAHF)
14947 if (Op.getResNo() == 1 &&
14948 (Opc == X86ISD::ADD ||
14949 Opc == X86ISD::SUB ||
14950 Opc == X86ISD::ADC ||
14951 Opc == X86ISD::SBB ||
14952 Opc == X86ISD::SMUL ||
14953 Opc == X86ISD::UMUL ||
14954 Opc == X86ISD::INC ||
14955 Opc == X86ISD::DEC ||
14956 Opc == X86ISD::OR ||
14957 Opc == X86ISD::XOR ||
14958 Opc == X86ISD::AND))
14961 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14967 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14968 if (V.getOpcode() != ISD::TRUNCATE)
14971 SDValue VOp0 = V.getOperand(0);
14972 unsigned InBits = VOp0.getValueSizeInBits();
14973 unsigned Bits = V.getValueSizeInBits();
14974 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14977 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14978 bool addTest = true;
14979 SDValue Cond = Op.getOperand(0);
14980 SDValue Op1 = Op.getOperand(1);
14981 SDValue Op2 = Op.getOperand(2);
14983 MVT VT = Op1.getSimpleValueType();
14986 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14987 // are available or VBLENDV if AVX is available.
14988 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14989 if (Cond.getOpcode() == ISD::SETCC &&
14990 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14991 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14992 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14993 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14994 int SSECC = translateX86FSETCC(
14995 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14998 if (Subtarget->hasAVX512()) {
14999 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
15000 DAG.getConstant(SSECC, DL, MVT::i8));
15001 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
15004 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
15005 DAG.getConstant(SSECC, DL, MVT::i8));
15007 // If we have AVX, we can use a variable vector select (VBLENDV) instead
15008 // of 3 logic instructions for size savings and potentially speed.
15009 // Unfortunately, there is no scalar form of VBLENDV.
15011 // If either operand is a constant, don't try this. We can expect to
15012 // optimize away at least one of the logic instructions later in that
15013 // case, so that sequence would be faster than a variable blend.
15015 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
15016 // uses XMM0 as the selection register. That may need just as many
15017 // instructions as the AND/ANDN/OR sequence due to register moves, so
15020 if (Subtarget->hasAVX() &&
15021 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
15023 // Convert to vectors, do a VSELECT, and convert back to scalar.
15024 // All of the conversions should be optimized away.
15026 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
15027 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
15028 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
15029 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
15031 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
15032 VCmp = DAG.getBitcast(VCmpVT, VCmp);
15034 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
15036 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
15037 VSel, DAG.getIntPtrConstant(0, DL));
15039 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
15040 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
15041 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
15045 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
15047 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
15048 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
15049 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
15050 Op1Scalar = Op1.getOperand(0);
15052 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
15053 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
15054 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
15055 Op2Scalar = Op2.getOperand(0);
15056 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
15057 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
15058 Op1Scalar.getValueType(),
15059 Cond, Op1Scalar, Op2Scalar);
15060 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
15061 return DAG.getBitcast(VT, newSelect);
15062 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
15063 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
15064 DAG.getIntPtrConstant(0, DL));
15068 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
15069 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
15070 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15071 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
15072 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15073 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
15074 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
15076 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
15079 if (Cond.getOpcode() == ISD::SETCC) {
15080 SDValue NewCond = LowerSETCC(Cond, DAG);
15081 if (NewCond.getNode())
15085 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
15086 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
15087 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
15088 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
15089 if (Cond.getOpcode() == X86ISD::SETCC &&
15090 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
15091 isNullConstant(Cond.getOperand(1).getOperand(1))) {
15092 SDValue Cmp = Cond.getOperand(1);
15094 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
15096 if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15097 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
15098 SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
15100 SDValue CmpOp0 = Cmp.getOperand(0);
15101 // Apply further optimizations for special cases
15102 // (select (x != 0), -1, 0) -> neg & sbb
15103 // (select (x == 0), 0, -1) -> neg & sbb
15104 if (isNullConstant(Y) &&
15105 (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
15106 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
15107 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
15108 DAG.getConstant(0, DL,
15109 CmpOp0.getValueType()),
15111 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15112 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15113 SDValue(Neg.getNode(), 1));
15117 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
15118 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
15119 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15121 SDValue Res = // Res = 0 or -1.
15122 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15123 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
15125 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
15126 Res = DAG.getNOT(DL, Res, Res.getValueType());
15128 if (!isNullConstant(Op2))
15129 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
15134 // Look past (and (setcc_carry (cmp ...)), 1).
15135 if (Cond.getOpcode() == ISD::AND &&
15136 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15137 isOneConstant(Cond.getOperand(1)))
15138 Cond = Cond.getOperand(0);
15140 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15141 // setting operand in place of the X86ISD::SETCC.
15142 unsigned CondOpcode = Cond.getOpcode();
15143 if (CondOpcode == X86ISD::SETCC ||
15144 CondOpcode == X86ISD::SETCC_CARRY) {
15145 CC = Cond.getOperand(0);
15147 SDValue Cmp = Cond.getOperand(1);
15148 unsigned Opc = Cmp.getOpcode();
15149 MVT VT = Op.getSimpleValueType();
15151 bool IllegalFPCMov = false;
15152 if (VT.isFloatingPoint() && !VT.isVector() &&
15153 !isScalarFPTypeInSSEReg(VT)) // FPStack?
15154 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
15156 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
15157 Opc == X86ISD::BT) { // FIXME
15161 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15162 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15163 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15164 Cond.getOperand(0).getValueType() != MVT::i8)) {
15165 SDValue LHS = Cond.getOperand(0);
15166 SDValue RHS = Cond.getOperand(1);
15167 unsigned X86Opcode;
15170 switch (CondOpcode) {
15171 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15172 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15173 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15174 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15175 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15176 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15177 default: llvm_unreachable("unexpected overflowing operator");
15179 if (CondOpcode == ISD::UMULO)
15180 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15183 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15185 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
15187 if (CondOpcode == ISD::UMULO)
15188 Cond = X86Op.getValue(2);
15190 Cond = X86Op.getValue(1);
15192 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
15197 // Look past the truncate if the high bits are known zero.
15198 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15199 Cond = Cond.getOperand(0);
15201 // We know the result of AND is compared against zero. Try to match
15203 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15204 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
15205 CC = NewSetCC.getOperand(0);
15206 Cond = NewSetCC.getOperand(1);
15213 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
15214 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
15217 // a < b ? -1 : 0 -> RES = ~setcc_carry
15218 // a < b ? 0 : -1 -> RES = setcc_carry
15219 // a >= b ? -1 : 0 -> RES = setcc_carry
15220 // a >= b ? 0 : -1 -> RES = ~setcc_carry
15221 if (Cond.getOpcode() == X86ISD::SUB) {
15222 Cond = ConvertCmpIfNecessary(Cond, DAG);
15223 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
15225 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
15226 (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15227 (isNullConstant(Op1) || isNullConstant(Op2))) {
15228 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15229 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15231 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
15232 return DAG.getNOT(DL, Res, Res.getValueType());
15237 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15238 // widen the cmov and push the truncate through. This avoids introducing a new
15239 // branch during isel and doesn't add any extensions.
15240 if (Op.getValueType() == MVT::i8 &&
15241 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15242 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15243 if (T1.getValueType() == T2.getValueType() &&
15244 // Blacklist CopyFromReg to avoid partial register stalls.
15245 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15246 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15247 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15248 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15252 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15253 // condition is true.
15254 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15255 SDValue Ops[] = { Op2, Op1, CC, Cond };
15256 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15259 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
15260 const X86Subtarget *Subtarget,
15261 SelectionDAG &DAG) {
15262 MVT VT = Op->getSimpleValueType(0);
15263 SDValue In = Op->getOperand(0);
15264 MVT InVT = In.getSimpleValueType();
15265 MVT VTElt = VT.getVectorElementType();
15266 MVT InVTElt = InVT.getVectorElementType();
15270 if ((InVTElt == MVT::i1) &&
15271 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15272 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15274 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15275 VTElt.getSizeInBits() <= 16)) ||
15277 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15278 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15280 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15281 VTElt.getSizeInBits() >= 32))))
15282 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15284 unsigned int NumElts = VT.getVectorNumElements();
15286 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
15289 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15290 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15291 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15292 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15295 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15296 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
15298 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
15301 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
15303 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
15304 if (VT.is512BitVector())
15306 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
15309 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
15310 const X86Subtarget *Subtarget,
15311 SelectionDAG &DAG) {
15312 SDValue In = Op->getOperand(0);
15313 MVT VT = Op->getSimpleValueType(0);
15314 MVT InVT = In.getSimpleValueType();
15315 assert(VT.getSizeInBits() == InVT.getSizeInBits());
15317 MVT InSVT = InVT.getVectorElementType();
15318 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
15320 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
15322 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15327 // SSE41 targets can use the pmovsx* instructions directly.
15328 if (Subtarget->hasSSE41())
15329 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15331 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15335 // As SRAI is only available on i16/i32 types, we expand only up to i32
15336 // and handle i64 separately.
15337 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15338 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15339 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15340 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15341 Curr = DAG.getBitcast(CurrVT, Curr);
15344 SDValue SignExt = Curr;
15345 if (CurrVT != InVT) {
15346 unsigned SignExtShift =
15347 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15348 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15349 DAG.getConstant(SignExtShift, dl, MVT::i8));
15355 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15356 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15357 DAG.getConstant(31, dl, MVT::i8));
15358 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15359 return DAG.getBitcast(VT, Ext);
15365 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15366 SelectionDAG &DAG) {
15367 MVT VT = Op->getSimpleValueType(0);
15368 SDValue In = Op->getOperand(0);
15369 MVT InVT = In.getSimpleValueType();
15372 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15373 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15375 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15376 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15377 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15380 if (Subtarget->hasInt256())
15381 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15383 // Optimize vectors in AVX mode
15384 // Sign extend v8i16 to v8i32 and
15387 // Divide input vector into two parts
15388 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15389 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15390 // concat the vectors to original VT
15392 unsigned NumElems = InVT.getVectorNumElements();
15393 SDValue Undef = DAG.getUNDEF(InVT);
15395 SmallVector<int,8> ShufMask1(NumElems, -1);
15396 for (unsigned i = 0; i != NumElems/2; ++i)
15399 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15401 SmallVector<int,8> ShufMask2(NumElems, -1);
15402 for (unsigned i = 0; i != NumElems/2; ++i)
15403 ShufMask2[i] = i + NumElems/2;
15405 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15407 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15408 VT.getVectorNumElements()/2);
15410 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15411 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15413 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15416 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15417 // may emit an illegal shuffle but the expansion is still better than scalar
15418 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15419 // we'll emit a shuffle and a arithmetic shift.
15420 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15421 // TODO: It is possible to support ZExt by zeroing the undef values during
15422 // the shuffle phase or after the shuffle.
15423 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15424 SelectionDAG &DAG) {
15425 MVT RegVT = Op.getSimpleValueType();
15426 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15427 assert(RegVT.isInteger() &&
15428 "We only custom lower integer vector sext loads.");
15430 // Nothing useful we can do without SSE2 shuffles.
15431 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15433 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15435 EVT MemVT = Ld->getMemoryVT();
15436 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15437 unsigned RegSz = RegVT.getSizeInBits();
15439 ISD::LoadExtType Ext = Ld->getExtensionType();
15441 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15442 && "Only anyext and sext are currently implemented.");
15443 assert(MemVT != RegVT && "Cannot extend to the same type");
15444 assert(MemVT.isVector() && "Must load a vector from memory");
15446 unsigned NumElems = RegVT.getVectorNumElements();
15447 unsigned MemSz = MemVT.getSizeInBits();
15448 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15450 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15451 // The only way in which we have a legal 256-bit vector result but not the
15452 // integer 256-bit operations needed to directly lower a sextload is if we
15453 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15454 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15455 // correctly legalized. We do this late to allow the canonical form of
15456 // sextload to persist throughout the rest of the DAG combiner -- it wants
15457 // to fold together any extensions it can, and so will fuse a sign_extend
15458 // of an sextload into a sextload targeting a wider value.
15460 if (MemSz == 128) {
15461 // Just switch this to a normal load.
15462 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15463 "it must be a legal 128-bit vector "
15465 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15466 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15467 Ld->isInvariant(), Ld->getAlignment());
15469 assert(MemSz < 128 &&
15470 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15471 // Do an sext load to a 128-bit vector type. We want to use the same
15472 // number of elements, but elements half as wide. This will end up being
15473 // recursively lowered by this routine, but will succeed as we definitely
15474 // have all the necessary features if we're using AVX1.
15476 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15477 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15479 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15480 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15481 Ld->isNonTemporal(), Ld->isInvariant(),
15482 Ld->getAlignment());
15485 // Replace chain users with the new chain.
15486 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15487 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15489 // Finally, do a normal sign-extend to the desired register.
15490 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15493 // All sizes must be a power of two.
15494 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15495 "Non-power-of-two elements are not custom lowered!");
15497 // Attempt to load the original value using scalar loads.
15498 // Find the largest scalar type that divides the total loaded size.
15499 MVT SclrLoadTy = MVT::i8;
15500 for (MVT Tp : MVT::integer_valuetypes()) {
15501 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15506 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15507 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15509 SclrLoadTy = MVT::f64;
15511 // Calculate the number of scalar loads that we need to perform
15512 // in order to load our vector from memory.
15513 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15515 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15516 "Can only lower sext loads with a single scalar load!");
15518 unsigned loadRegZize = RegSz;
15519 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15522 // Represent our vector as a sequence of elements which are the
15523 // largest scalar that we can load.
15524 EVT LoadUnitVecVT = EVT::getVectorVT(
15525 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15527 // Represent the data using the same element type that is stored in
15528 // memory. In practice, we ''widen'' MemVT.
15530 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15531 loadRegZize / MemVT.getScalarSizeInBits());
15533 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15534 "Invalid vector type");
15536 // We can't shuffle using an illegal type.
15537 assert(TLI.isTypeLegal(WideVecVT) &&
15538 "We only lower types that form legal widened vector types");
15540 SmallVector<SDValue, 8> Chains;
15541 SDValue Ptr = Ld->getBasePtr();
15542 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15543 TLI.getPointerTy(DAG.getDataLayout()));
15544 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15546 for (unsigned i = 0; i < NumLoads; ++i) {
15547 // Perform a single load.
15548 SDValue ScalarLoad =
15549 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15550 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15551 Ld->getAlignment());
15552 Chains.push_back(ScalarLoad.getValue(1));
15553 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15554 // another round of DAGCombining.
15556 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15558 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15559 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15561 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15564 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15566 // Bitcast the loaded value to a vector of the original element type, in
15567 // the size of the target vector type.
15568 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15569 unsigned SizeRatio = RegSz / MemSz;
15571 if (Ext == ISD::SEXTLOAD) {
15572 // If we have SSE4.1, we can directly emit a VSEXT node.
15573 if (Subtarget->hasSSE41()) {
15574 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15575 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15579 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15581 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15582 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15584 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15585 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15589 // Redistribute the loaded elements into the different locations.
15590 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15591 for (unsigned i = 0; i != NumElems; ++i)
15592 ShuffleVec[i * SizeRatio] = i;
15594 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15595 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15597 // Bitcast to the requested type.
15598 Shuff = DAG.getBitcast(RegVT, Shuff);
15599 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15603 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15604 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15605 // from the AND / OR.
15606 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15607 Opc = Op.getOpcode();
15608 if (Opc != ISD::OR && Opc != ISD::AND)
15610 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15611 Op.getOperand(0).hasOneUse() &&
15612 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15613 Op.getOperand(1).hasOneUse());
15616 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15617 // 1 and that the SETCC node has a single use.
15618 static bool isXor1OfSetCC(SDValue Op) {
15619 if (Op.getOpcode() != ISD::XOR)
15621 if (isOneConstant(Op.getOperand(1)))
15622 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15623 Op.getOperand(0).hasOneUse();
15627 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15628 bool addTest = true;
15629 SDValue Chain = Op.getOperand(0);
15630 SDValue Cond = Op.getOperand(1);
15631 SDValue Dest = Op.getOperand(2);
15634 bool Inverted = false;
15636 if (Cond.getOpcode() == ISD::SETCC) {
15637 // Check for setcc([su]{add,sub,mul}o == 0).
15638 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15639 isNullConstant(Cond.getOperand(1)) &&
15640 Cond.getOperand(0).getResNo() == 1 &&
15641 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15642 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15643 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15644 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15645 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15646 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15648 Cond = Cond.getOperand(0);
15650 SDValue NewCond = LowerSETCC(Cond, DAG);
15651 if (NewCond.getNode())
15656 // FIXME: LowerXALUO doesn't handle these!!
15657 else if (Cond.getOpcode() == X86ISD::ADD ||
15658 Cond.getOpcode() == X86ISD::SUB ||
15659 Cond.getOpcode() == X86ISD::SMUL ||
15660 Cond.getOpcode() == X86ISD::UMUL)
15661 Cond = LowerXALUO(Cond, DAG);
15664 // Look pass (and (setcc_carry (cmp ...)), 1).
15665 if (Cond.getOpcode() == ISD::AND &&
15666 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15667 isOneConstant(Cond.getOperand(1)))
15668 Cond = Cond.getOperand(0);
15670 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15671 // setting operand in place of the X86ISD::SETCC.
15672 unsigned CondOpcode = Cond.getOpcode();
15673 if (CondOpcode == X86ISD::SETCC ||
15674 CondOpcode == X86ISD::SETCC_CARRY) {
15675 CC = Cond.getOperand(0);
15677 SDValue Cmp = Cond.getOperand(1);
15678 unsigned Opc = Cmp.getOpcode();
15679 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15680 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15684 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15688 // These can only come from an arithmetic instruction with overflow,
15689 // e.g. SADDO, UADDO.
15690 Cond = Cond.getNode()->getOperand(1);
15696 CondOpcode = Cond.getOpcode();
15697 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15698 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15699 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15700 Cond.getOperand(0).getValueType() != MVT::i8)) {
15701 SDValue LHS = Cond.getOperand(0);
15702 SDValue RHS = Cond.getOperand(1);
15703 unsigned X86Opcode;
15706 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15707 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15709 switch (CondOpcode) {
15710 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15712 if (isOneConstant(RHS)) {
15713 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15716 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15717 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15719 if (isOneConstant(RHS)) {
15720 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15723 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15724 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15725 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15726 default: llvm_unreachable("unexpected overflowing operator");
15729 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15730 if (CondOpcode == ISD::UMULO)
15731 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15734 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15736 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15738 if (CondOpcode == ISD::UMULO)
15739 Cond = X86Op.getValue(2);
15741 Cond = X86Op.getValue(1);
15743 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15747 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15748 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15749 if (CondOpc == ISD::OR) {
15750 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15751 // two branches instead of an explicit OR instruction with a
15753 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15754 isX86LogicalCmp(Cmp)) {
15755 CC = Cond.getOperand(0).getOperand(0);
15756 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15757 Chain, Dest, CC, Cmp);
15758 CC = Cond.getOperand(1).getOperand(0);
15762 } else { // ISD::AND
15763 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15764 // two branches instead of an explicit AND instruction with a
15765 // separate test. However, we only do this if this block doesn't
15766 // have a fall-through edge, because this requires an explicit
15767 // jmp when the condition is false.
15768 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15769 isX86LogicalCmp(Cmp) &&
15770 Op.getNode()->hasOneUse()) {
15771 X86::CondCode CCode =
15772 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15773 CCode = X86::GetOppositeBranchCondition(CCode);
15774 CC = DAG.getConstant(CCode, dl, MVT::i8);
15775 SDNode *User = *Op.getNode()->use_begin();
15776 // Look for an unconditional branch following this conditional branch.
15777 // We need this because we need to reverse the successors in order
15778 // to implement FCMP_OEQ.
15779 if (User->getOpcode() == ISD::BR) {
15780 SDValue FalseBB = User->getOperand(1);
15782 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15783 assert(NewBR == User);
15787 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15788 Chain, Dest, CC, Cmp);
15789 X86::CondCode CCode =
15790 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15791 CCode = X86::GetOppositeBranchCondition(CCode);
15792 CC = DAG.getConstant(CCode, dl, MVT::i8);
15798 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15799 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15800 // It should be transformed during dag combiner except when the condition
15801 // is set by a arithmetics with overflow node.
15802 X86::CondCode CCode =
15803 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15804 CCode = X86::GetOppositeBranchCondition(CCode);
15805 CC = DAG.getConstant(CCode, dl, MVT::i8);
15806 Cond = Cond.getOperand(0).getOperand(1);
15808 } else if (Cond.getOpcode() == ISD::SETCC &&
15809 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15810 // For FCMP_OEQ, we can emit
15811 // two branches instead of an explicit AND instruction with a
15812 // separate test. However, we only do this if this block doesn't
15813 // have a fall-through edge, because this requires an explicit
15814 // jmp when the condition is false.
15815 if (Op.getNode()->hasOneUse()) {
15816 SDNode *User = *Op.getNode()->use_begin();
15817 // Look for an unconditional branch following this conditional branch.
15818 // We need this because we need to reverse the successors in order
15819 // to implement FCMP_OEQ.
15820 if (User->getOpcode() == ISD::BR) {
15821 SDValue FalseBB = User->getOperand(1);
15823 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15824 assert(NewBR == User);
15828 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15829 Cond.getOperand(0), Cond.getOperand(1));
15830 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15831 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15832 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15833 Chain, Dest, CC, Cmp);
15834 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15839 } else if (Cond.getOpcode() == ISD::SETCC &&
15840 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15841 // For FCMP_UNE, we can emit
15842 // two branches instead of an explicit AND instruction with a
15843 // separate test. However, we only do this if this block doesn't
15844 // have a fall-through edge, because this requires an explicit
15845 // jmp when the condition is false.
15846 if (Op.getNode()->hasOneUse()) {
15847 SDNode *User = *Op.getNode()->use_begin();
15848 // Look for an unconditional branch following this conditional branch.
15849 // We need this because we need to reverse the successors in order
15850 // to implement FCMP_UNE.
15851 if (User->getOpcode() == ISD::BR) {
15852 SDValue FalseBB = User->getOperand(1);
15854 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15855 assert(NewBR == User);
15858 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15859 Cond.getOperand(0), Cond.getOperand(1));
15860 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15861 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15862 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15863 Chain, Dest, CC, Cmp);
15864 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15874 // Look pass the truncate if the high bits are known zero.
15875 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15876 Cond = Cond.getOperand(0);
15878 // We know the result of AND is compared against zero. Try to match
15880 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15881 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15882 CC = NewSetCC.getOperand(0);
15883 Cond = NewSetCC.getOperand(1);
15890 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15891 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15892 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15894 Cond = ConvertCmpIfNecessary(Cond, DAG);
15895 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15896 Chain, Dest, CC, Cond);
15899 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15900 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15901 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15902 // that the guard pages used by the OS virtual memory manager are allocated in
15903 // correct sequence.
15905 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15906 SelectionDAG &DAG) const {
15907 MachineFunction &MF = DAG.getMachineFunction();
15908 bool SplitStack = MF.shouldSplitStack();
15909 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15914 SDNode *Node = Op.getNode();
15915 SDValue Chain = Op.getOperand(0);
15916 SDValue Size = Op.getOperand(1);
15917 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15918 EVT VT = Node->getValueType(0);
15920 // Chain the dynamic stack allocation so that it doesn't modify the stack
15921 // pointer when other instructions are using the stack.
15922 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
15924 bool Is64Bit = Subtarget->is64Bit();
15925 MVT SPTy = getPointerTy(DAG.getDataLayout());
15929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15930 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15931 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15932 " not tell us which reg is the stack pointer!");
15933 EVT VT = Node->getValueType(0);
15934 SDValue Tmp3 = Node->getOperand(2);
15936 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15937 Chain = SP.getValue(1);
15938 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15939 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15940 unsigned StackAlign = TFI.getStackAlignment();
15941 Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15942 if (Align > StackAlign)
15943 Result = DAG.getNode(ISD::AND, dl, VT, Result,
15944 DAG.getConstant(-(uint64_t)Align, dl, VT));
15945 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain
15946 } else if (SplitStack) {
15947 MachineRegisterInfo &MRI = MF.getRegInfo();
15950 // The 64 bit implementation of segmented stacks needs to clobber both r10
15951 // r11. This makes it impossible to use it along with nested parameters.
15952 const Function *F = MF.getFunction();
15954 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15956 if (I->hasNestAttr())
15957 report_fatal_error("Cannot use segmented stacks with functions that "
15958 "have nested arguments.");
15961 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15962 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15963 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15964 Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15965 DAG.getRegister(Vreg, SPTy));
15968 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15970 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15971 Flag = Chain.getValue(1);
15972 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15974 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15976 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15977 unsigned SPReg = RegInfo->getStackRegister();
15978 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15979 Chain = SP.getValue(1);
15982 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15983 DAG.getConstant(-(uint64_t)Align, dl, VT));
15984 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15990 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15991 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
15993 SDValue Ops[2] = {Result, Chain};
15994 return DAG.getMergeValues(Ops, dl);
15997 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15998 MachineFunction &MF = DAG.getMachineFunction();
15999 auto PtrVT = getPointerTy(MF.getDataLayout());
16000 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16002 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16005 if (!Subtarget->is64Bit() ||
16006 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
16007 // vastart just stores the address of the VarArgsFrameIndex slot into the
16008 // memory location argument.
16009 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
16010 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
16011 MachinePointerInfo(SV), false, false, 0);
16015 // gp_offset (0 - 6 * 8)
16016 // fp_offset (48 - 48 + 8 * 16)
16017 // overflow_arg_area (point to parameters coming in memory).
16019 SmallVector<SDValue, 8> MemOps;
16020 SDValue FIN = Op.getOperand(1);
16022 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
16023 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
16025 FIN, MachinePointerInfo(SV), false, false, 0);
16026 MemOps.push_back(Store);
16029 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
16030 Store = DAG.getStore(Op.getOperand(0), DL,
16031 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
16033 FIN, MachinePointerInfo(SV, 4), false, false, 0);
16034 MemOps.push_back(Store);
16036 // Store ptr to overflow_arg_area
16037 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
16038 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
16039 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
16040 MachinePointerInfo(SV, 8),
16042 MemOps.push_back(Store);
16044 // Store ptr to reg_save_area.
16045 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
16046 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
16047 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
16048 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
16049 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
16050 MemOps.push_back(Store);
16051 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
16054 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
16055 assert(Subtarget->is64Bit() &&
16056 "LowerVAARG only handles 64-bit va_arg!");
16057 assert(Op.getNode()->getNumOperands() == 4);
16059 MachineFunction &MF = DAG.getMachineFunction();
16060 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
16061 // The Win64 ABI uses char* instead of a structure.
16062 return DAG.expandVAArg(Op.getNode());
16064 SDValue Chain = Op.getOperand(0);
16065 SDValue SrcPtr = Op.getOperand(1);
16066 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16067 unsigned Align = Op.getConstantOperandVal(3);
16070 EVT ArgVT = Op.getNode()->getValueType(0);
16071 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16072 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
16075 // Decide which area this value should be read from.
16076 // TODO: Implement the AMD64 ABI in its entirety. This simple
16077 // selection mechanism works only for the basic types.
16078 if (ArgVT == MVT::f80) {
16079 llvm_unreachable("va_arg for f80 not yet implemented");
16080 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
16081 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
16082 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
16083 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
16085 llvm_unreachable("Unhandled argument type in LowerVAARG");
16088 if (ArgMode == 2) {
16089 // Sanity Check: Make sure using fp_offset makes sense.
16090 assert(!Subtarget->useSoftFloat() &&
16091 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
16092 Subtarget->hasSSE1());
16095 // Insert VAARG_64 node into the DAG
16096 // VAARG_64 returns two values: Variable Argument Address, Chain
16097 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
16098 DAG.getConstant(ArgMode, dl, MVT::i8),
16099 DAG.getConstant(Align, dl, MVT::i32)};
16100 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
16101 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
16102 VTs, InstOps, MVT::i64,
16103 MachinePointerInfo(SV),
16105 /*Volatile=*/false,
16107 /*WriteMem=*/true);
16108 Chain = VAARG.getValue(1);
16110 // Load the next argument and return it
16111 return DAG.getLoad(ArgVT, dl,
16114 MachinePointerInfo(),
16115 false, false, false, 0);
16118 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
16119 SelectionDAG &DAG) {
16120 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
16121 // where a va_list is still an i8*.
16122 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
16123 if (Subtarget->isCallingConvWin64(
16124 DAG.getMachineFunction().getFunction()->getCallingConv()))
16125 // Probably a Win64 va_copy.
16126 return DAG.expandVACopy(Op.getNode());
16128 SDValue Chain = Op.getOperand(0);
16129 SDValue DstPtr = Op.getOperand(1);
16130 SDValue SrcPtr = Op.getOperand(2);
16131 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
16132 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16135 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
16136 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
16138 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
16141 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
16142 // amount is a constant. Takes immediate version of shift as input.
16143 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
16144 SDValue SrcOp, uint64_t ShiftAmt,
16145 SelectionDAG &DAG) {
16146 MVT ElementType = VT.getVectorElementType();
16148 // Fold this packed shift into its first operand if ShiftAmt is 0.
16152 // Check for ShiftAmt >= element width
16153 if (ShiftAmt >= ElementType.getSizeInBits()) {
16154 if (Opc == X86ISD::VSRAI)
16155 ShiftAmt = ElementType.getSizeInBits() - 1;
16157 return DAG.getConstant(0, dl, VT);
16160 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
16161 && "Unknown target vector shift-by-constant node");
16163 // Fold this packed vector shift into a build vector if SrcOp is a
16164 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
16165 if (VT == SrcOp.getSimpleValueType() &&
16166 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
16167 SmallVector<SDValue, 8> Elts;
16168 unsigned NumElts = SrcOp->getNumOperands();
16169 ConstantSDNode *ND;
16172 default: llvm_unreachable(nullptr);
16173 case X86ISD::VSHLI:
16174 for (unsigned i=0; i!=NumElts; ++i) {
16175 SDValue CurrentOp = SrcOp->getOperand(i);
16176 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16177 Elts.push_back(CurrentOp);
16180 ND = cast<ConstantSDNode>(CurrentOp);
16181 const APInt &C = ND->getAPIntValue();
16182 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
16185 case X86ISD::VSRLI:
16186 for (unsigned i=0; i!=NumElts; ++i) {
16187 SDValue CurrentOp = SrcOp->getOperand(i);
16188 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16189 Elts.push_back(CurrentOp);
16192 ND = cast<ConstantSDNode>(CurrentOp);
16193 const APInt &C = ND->getAPIntValue();
16194 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
16197 case X86ISD::VSRAI:
16198 for (unsigned i=0; i!=NumElts; ++i) {
16199 SDValue CurrentOp = SrcOp->getOperand(i);
16200 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16201 Elts.push_back(CurrentOp);
16204 ND = cast<ConstantSDNode>(CurrentOp);
16205 const APInt &C = ND->getAPIntValue();
16206 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
16211 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16214 return DAG.getNode(Opc, dl, VT, SrcOp,
16215 DAG.getConstant(ShiftAmt, dl, MVT::i8));
16218 // getTargetVShiftNode - Handle vector element shifts where the shift amount
16219 // may or may not be a constant. Takes immediate version of shift as input.
16220 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
16221 SDValue SrcOp, SDValue ShAmt,
16222 SelectionDAG &DAG) {
16223 MVT SVT = ShAmt.getSimpleValueType();
16224 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
16226 // Catch shift-by-constant.
16227 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16228 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16229 CShAmt->getZExtValue(), DAG);
16231 // Change opcode to non-immediate version
16233 default: llvm_unreachable("Unknown target vector shift node");
16234 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16235 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16236 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16239 const X86Subtarget &Subtarget =
16240 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16241 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16242 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16243 // Let the shuffle legalizer expand this shift amount node.
16244 SDValue Op0 = ShAmt.getOperand(0);
16245 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16246 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16248 // Need to build a vector containing shift amount.
16249 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16250 SmallVector<SDValue, 4> ShOps;
16251 ShOps.push_back(ShAmt);
16252 if (SVT == MVT::i32) {
16253 ShOps.push_back(DAG.getConstant(0, dl, SVT));
16254 ShOps.push_back(DAG.getUNDEF(SVT));
16256 ShOps.push_back(DAG.getUNDEF(SVT));
16258 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16259 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16262 // The return type has to be a 128-bit type with the same element
16263 // type as the input type.
16264 MVT EltVT = VT.getVectorElementType();
16265 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16267 ShAmt = DAG.getBitcast(ShVT, ShAmt);
16268 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16271 /// \brief Return Mask with the necessary casting or extending
16272 /// for \p Mask according to \p MaskVT when lowering masking intrinsics
16273 static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
16274 const X86Subtarget *Subtarget,
16275 SelectionDAG &DAG, SDLoc dl) {
16277 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
16278 // Mask should be extended
16279 Mask = DAG.getNode(ISD::ANY_EXTEND, dl,
16280 MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask);
16283 if (Mask.getSimpleValueType() == MVT::i64 && Subtarget->is32Bit()) {
16284 if (MaskVT == MVT::v64i1) {
16285 assert(Subtarget->hasBWI() && "Expected AVX512BW target!");
16286 // In case 32bit mode, bitcast i64 is illegal, extend/split it.
16288 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16289 DAG.getConstant(0, dl, MVT::i32));
16290 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16291 DAG.getConstant(1, dl, MVT::i32));
16293 Lo = DAG.getBitcast(MVT::v32i1, Lo);
16294 Hi = DAG.getBitcast(MVT::v32i1, Hi);
16296 return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
16298 // MaskVT require < 64bit. Truncate mask (should succeed in any case),
16300 MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
16301 return DAG.getBitcast(MaskVT,
16302 DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask));
16306 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16307 Mask.getSimpleValueType().getSizeInBits());
16308 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16309 // are extracted by EXTRACT_SUBVECTOR.
16310 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16311 DAG.getBitcast(BitcastVT, Mask),
16312 DAG.getIntPtrConstant(0, dl));
16316 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16317 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16318 /// necessary casting or extending for \p Mask when lowering masking intrinsics
16319 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16320 SDValue PreservedSrc,
16321 const X86Subtarget *Subtarget,
16322 SelectionDAG &DAG) {
16323 MVT VT = Op.getSimpleValueType();
16324 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16325 unsigned OpcodeSelect = ISD::VSELECT;
16328 if (isAllOnesConstant(Mask))
16331 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16333 switch (Op.getOpcode()) {
16335 case X86ISD::PCMPEQM:
16336 case X86ISD::PCMPGTM:
16338 case X86ISD::CMPMU:
16339 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16340 case X86ISD::VFPCLASS:
16341 case X86ISD::VFPCLASSS:
16342 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16343 case X86ISD::VTRUNC:
16344 case X86ISD::VTRUNCS:
16345 case X86ISD::VTRUNCUS:
16346 // We can't use ISD::VSELECT here because it is not always "Legal"
16347 // for the destination type. For example vpmovqb require only AVX512
16348 // and vselect that can operate on byte element type require BWI
16349 OpcodeSelect = X86ISD::SELECT;
16352 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16353 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16354 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16357 /// \brief Creates an SDNode for a predicated scalar operation.
16358 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16359 /// The mask is coming as MVT::i8 and it should be truncated
16360 /// to MVT::i1 while lowering masking intrinsics.
16361 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16362 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16363 /// for a scalar instruction.
16364 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16365 SDValue PreservedSrc,
16366 const X86Subtarget *Subtarget,
16367 SelectionDAG &DAG) {
16368 if (isAllOnesConstant(Mask))
16371 MVT VT = Op.getSimpleValueType();
16373 // The mask should be of type MVT::i1
16374 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16376 if (Op.getOpcode() == X86ISD::FSETCC)
16377 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16378 if (Op.getOpcode() == X86ISD::VFPCLASS ||
16379 Op.getOpcode() == X86ISD::VFPCLASSS)
16380 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16382 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16383 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16384 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16387 static int getSEHRegistrationNodeSize(const Function *Fn) {
16388 if (!Fn->hasPersonalityFn())
16389 report_fatal_error(
16390 "querying registration node size for function without personality");
16391 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16392 // WinEHStatePass for the full struct definition.
16393 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16394 case EHPersonality::MSVC_X86SEH: return 24;
16395 case EHPersonality::MSVC_CXX: return 16;
16398 report_fatal_error(
16399 "can only recover FP for 32-bit MSVC EH personality functions");
16402 /// When the MSVC runtime transfers control to us, either to an outlined
16403 /// function or when returning to a parent frame after catching an exception, we
16404 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16405 /// Here's the math:
16406 /// RegNodeBase = EntryEBP - RegNodeSize
16407 /// ParentFP = RegNodeBase - ParentFrameOffset
16408 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16409 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16410 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16411 SDValue EntryEBP) {
16412 MachineFunction &MF = DAG.getMachineFunction();
16415 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16416 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16418 // It's possible that the parent function no longer has a personality function
16419 // if the exceptional code was optimized away, in which case we just return
16420 // the incoming EBP.
16421 if (!Fn->hasPersonalityFn())
16424 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16425 // registration, or the .set_setframe offset.
16426 MCSymbol *OffsetSym =
16427 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16428 GlobalValue::getRealLinkageName(Fn->getName()));
16429 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16430 SDValue ParentFrameOffset =
16431 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16433 // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after
16434 // prologue to RBP in the parent function.
16435 const X86Subtarget &Subtarget =
16436 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16437 if (Subtarget.is64Bit())
16438 return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset);
16440 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16441 // RegNodeBase = EntryEBP - RegNodeSize
16442 // ParentFP = RegNodeBase - ParentFrameOffset
16443 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16444 DAG.getConstant(RegNodeSize, dl, PtrVT));
16445 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset);
16448 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16449 SelectionDAG &DAG) {
16451 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16452 MVT VT = Op.getSimpleValueType();
16453 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16455 switch(IntrData->Type) {
16456 case INTR_TYPE_1OP:
16457 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16458 case INTR_TYPE_2OP:
16459 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16461 case INTR_TYPE_2OP_IMM8:
16462 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16463 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16464 case INTR_TYPE_3OP:
16465 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16466 Op.getOperand(2), Op.getOperand(3));
16467 case INTR_TYPE_4OP:
16468 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16469 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16470 case INTR_TYPE_1OP_MASK_RM: {
16471 SDValue Src = Op.getOperand(1);
16472 SDValue PassThru = Op.getOperand(2);
16473 SDValue Mask = Op.getOperand(3);
16474 SDValue RoundingMode;
16475 // We allways add rounding mode to the Node.
16476 // If the rounding mode is not specified, we add the
16477 // "current direction" mode.
16478 if (Op.getNumOperands() == 4)
16480 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16482 RoundingMode = Op.getOperand(4);
16483 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16484 if (IntrWithRoundingModeOpcode != 0)
16485 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16486 X86::STATIC_ROUNDING::CUR_DIRECTION)
16487 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16488 dl, Op.getValueType(), Src, RoundingMode),
16489 Mask, PassThru, Subtarget, DAG);
16490 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16492 Mask, PassThru, Subtarget, DAG);
16494 case INTR_TYPE_1OP_MASK: {
16495 SDValue Src = Op.getOperand(1);
16496 SDValue PassThru = Op.getOperand(2);
16497 SDValue Mask = Op.getOperand(3);
16498 // We add rounding mode to the Node when
16499 // - RM Opcode is specified and
16500 // - RM is not "current direction".
16501 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16502 if (IntrWithRoundingModeOpcode != 0) {
16503 SDValue Rnd = Op.getOperand(4);
16504 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16505 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16506 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16507 dl, Op.getValueType(),
16509 Mask, PassThru, Subtarget, DAG);
16512 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16513 Mask, PassThru, Subtarget, DAG);
16515 case INTR_TYPE_SCALAR_MASK: {
16516 SDValue Src1 = Op.getOperand(1);
16517 SDValue Src2 = Op.getOperand(2);
16518 SDValue passThru = Op.getOperand(3);
16519 SDValue Mask = Op.getOperand(4);
16520 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16521 Mask, passThru, Subtarget, DAG);
16523 case INTR_TYPE_SCALAR_MASK_RM: {
16524 SDValue Src1 = Op.getOperand(1);
16525 SDValue Src2 = Op.getOperand(2);
16526 SDValue Src0 = Op.getOperand(3);
16527 SDValue Mask = Op.getOperand(4);
16528 // There are 2 kinds of intrinsics in this group:
16529 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16530 // (2) With rounding mode and sae - 7 operands.
16531 if (Op.getNumOperands() == 6) {
16532 SDValue Sae = Op.getOperand(5);
16533 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16534 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16536 Mask, Src0, Subtarget, DAG);
16538 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16539 SDValue RoundingMode = Op.getOperand(5);
16540 SDValue Sae = Op.getOperand(6);
16541 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16542 RoundingMode, Sae),
16543 Mask, Src0, Subtarget, DAG);
16545 case INTR_TYPE_2OP_MASK:
16546 case INTR_TYPE_2OP_IMM8_MASK: {
16547 SDValue Src1 = Op.getOperand(1);
16548 SDValue Src2 = Op.getOperand(2);
16549 SDValue PassThru = Op.getOperand(3);
16550 SDValue Mask = Op.getOperand(4);
16552 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16553 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16555 // We specify 2 possible opcodes for intrinsics with rounding modes.
16556 // First, we check if the intrinsic may have non-default rounding mode,
16557 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16558 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16559 if (IntrWithRoundingModeOpcode != 0) {
16560 SDValue Rnd = Op.getOperand(5);
16561 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16562 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16563 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16564 dl, Op.getValueType(),
16566 Mask, PassThru, Subtarget, DAG);
16569 // TODO: Intrinsics should have fast-math-flags to propagate.
16570 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16571 Mask, PassThru, Subtarget, DAG);
16573 case INTR_TYPE_2OP_MASK_RM: {
16574 SDValue Src1 = Op.getOperand(1);
16575 SDValue Src2 = Op.getOperand(2);
16576 SDValue PassThru = Op.getOperand(3);
16577 SDValue Mask = Op.getOperand(4);
16578 // We specify 2 possible modes for intrinsics, with/without rounding
16580 // First, we check if the intrinsic have rounding mode (6 operands),
16581 // if not, we set rounding mode to "current".
16583 if (Op.getNumOperands() == 6)
16584 Rnd = Op.getOperand(5);
16586 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16587 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16589 Mask, PassThru, Subtarget, DAG);
16591 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16592 SDValue Src1 = Op.getOperand(1);
16593 SDValue Src2 = Op.getOperand(2);
16594 SDValue Src3 = Op.getOperand(3);
16595 SDValue PassThru = Op.getOperand(4);
16596 SDValue Mask = Op.getOperand(5);
16597 SDValue Sae = Op.getOperand(6);
16599 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16601 Mask, PassThru, Subtarget, DAG);
16603 case INTR_TYPE_3OP_MASK_RM: {
16604 SDValue Src1 = Op.getOperand(1);
16605 SDValue Src2 = Op.getOperand(2);
16606 SDValue Imm = Op.getOperand(3);
16607 SDValue PassThru = Op.getOperand(4);
16608 SDValue Mask = Op.getOperand(5);
16609 // We specify 2 possible modes for intrinsics, with/without rounding
16611 // First, we check if the intrinsic have rounding mode (7 operands),
16612 // if not, we set rounding mode to "current".
16614 if (Op.getNumOperands() == 7)
16615 Rnd = Op.getOperand(6);
16617 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16618 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16619 Src1, Src2, Imm, Rnd),
16620 Mask, PassThru, Subtarget, DAG);
16622 case INTR_TYPE_3OP_IMM8_MASK:
16623 case INTR_TYPE_3OP_MASK:
16624 case INSERT_SUBVEC: {
16625 SDValue Src1 = Op.getOperand(1);
16626 SDValue Src2 = Op.getOperand(2);
16627 SDValue Src3 = Op.getOperand(3);
16628 SDValue PassThru = Op.getOperand(4);
16629 SDValue Mask = Op.getOperand(5);
16631 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16632 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16633 else if (IntrData->Type == INSERT_SUBVEC) {
16634 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16635 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16636 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16637 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16638 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16641 // We specify 2 possible opcodes for intrinsics with rounding modes.
16642 // First, we check if the intrinsic may have non-default rounding mode,
16643 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16644 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16645 if (IntrWithRoundingModeOpcode != 0) {
16646 SDValue Rnd = Op.getOperand(6);
16647 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16648 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16649 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16650 dl, Op.getValueType(),
16651 Src1, Src2, Src3, Rnd),
16652 Mask, PassThru, Subtarget, DAG);
16655 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16657 Mask, PassThru, Subtarget, DAG);
16659 case VPERM_3OP_MASKZ:
16660 case VPERM_3OP_MASK:{
16661 // Src2 is the PassThru
16662 SDValue Src1 = Op.getOperand(1);
16663 SDValue Src2 = Op.getOperand(2);
16664 SDValue Src3 = Op.getOperand(3);
16665 SDValue Mask = Op.getOperand(4);
16666 MVT VT = Op.getSimpleValueType();
16667 SDValue PassThru = SDValue();
16669 // set PassThru element
16670 if (IntrData->Type == VPERM_3OP_MASKZ)
16671 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16673 PassThru = DAG.getBitcast(VT, Src2);
16675 // Swap Src1 and Src2 in the node creation
16676 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16677 dl, Op.getValueType(),
16679 Mask, PassThru, Subtarget, DAG);
16683 case FMA_OP_MASK: {
16684 SDValue Src1 = Op.getOperand(1);
16685 SDValue Src2 = Op.getOperand(2);
16686 SDValue Src3 = Op.getOperand(3);
16687 SDValue Mask = Op.getOperand(4);
16688 MVT VT = Op.getSimpleValueType();
16689 SDValue PassThru = SDValue();
16691 // set PassThru element
16692 if (IntrData->Type == FMA_OP_MASKZ)
16693 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16694 else if (IntrData->Type == FMA_OP_MASK3)
16699 // We specify 2 possible opcodes for intrinsics with rounding modes.
16700 // First, we check if the intrinsic may have non-default rounding mode,
16701 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16702 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16703 if (IntrWithRoundingModeOpcode != 0) {
16704 SDValue Rnd = Op.getOperand(5);
16705 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16706 X86::STATIC_ROUNDING::CUR_DIRECTION)
16707 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16708 dl, Op.getValueType(),
16709 Src1, Src2, Src3, Rnd),
16710 Mask, PassThru, Subtarget, DAG);
16712 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16713 dl, Op.getValueType(),
16715 Mask, PassThru, Subtarget, DAG);
16717 case TERLOG_OP_MASK:
16718 case TERLOG_OP_MASKZ: {
16719 SDValue Src1 = Op.getOperand(1);
16720 SDValue Src2 = Op.getOperand(2);
16721 SDValue Src3 = Op.getOperand(3);
16722 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16723 SDValue Mask = Op.getOperand(5);
16724 MVT VT = Op.getSimpleValueType();
16725 SDValue PassThru = Src1;
16726 // Set PassThru element.
16727 if (IntrData->Type == TERLOG_OP_MASKZ)
16728 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16730 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16731 Src1, Src2, Src3, Src4),
16732 Mask, PassThru, Subtarget, DAG);
16735 // FPclass intrinsics with mask
16736 SDValue Src1 = Op.getOperand(1);
16737 MVT VT = Src1.getSimpleValueType();
16738 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16739 SDValue Imm = Op.getOperand(2);
16740 SDValue Mask = Op.getOperand(3);
16741 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16742 Mask.getSimpleValueType().getSizeInBits());
16743 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16744 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16745 DAG.getTargetConstant(0, dl, MaskVT),
16747 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16748 DAG.getUNDEF(BitcastVT), FPclassMask,
16749 DAG.getIntPtrConstant(0, dl));
16750 return DAG.getBitcast(Op.getValueType(), Res);
16753 SDValue Src1 = Op.getOperand(1);
16754 SDValue Imm = Op.getOperand(2);
16755 SDValue Mask = Op.getOperand(3);
16756 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16757 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16758 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16759 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16762 case CMP_MASK_CC: {
16763 // Comparison intrinsics with masks.
16764 // Example of transformation:
16765 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16766 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16768 // (v8i1 (insert_subvector undef,
16769 // (v2i1 (and (PCMPEQM %a, %b),
16770 // (extract_subvector
16771 // (v8i1 (bitcast %mask)), 0))), 0))))
16772 MVT VT = Op.getOperand(1).getSimpleValueType();
16773 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16774 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16775 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16776 Mask.getSimpleValueType().getSizeInBits());
16778 if (IntrData->Type == CMP_MASK_CC) {
16779 SDValue CC = Op.getOperand(3);
16780 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16781 // We specify 2 possible opcodes for intrinsics with rounding modes.
16782 // First, we check if the intrinsic may have non-default rounding mode,
16783 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16784 if (IntrData->Opc1 != 0) {
16785 SDValue Rnd = Op.getOperand(5);
16786 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16787 X86::STATIC_ROUNDING::CUR_DIRECTION)
16788 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16789 Op.getOperand(2), CC, Rnd);
16791 //default rounding mode
16793 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16794 Op.getOperand(2), CC);
16797 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16798 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16801 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16802 DAG.getTargetConstant(0, dl,
16805 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16806 DAG.getUNDEF(BitcastVT), CmpMask,
16807 DAG.getIntPtrConstant(0, dl));
16808 return DAG.getBitcast(Op.getValueType(), Res);
16810 case CMP_MASK_SCALAR_CC: {
16811 SDValue Src1 = Op.getOperand(1);
16812 SDValue Src2 = Op.getOperand(2);
16813 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16814 SDValue Mask = Op.getOperand(4);
16817 if (IntrData->Opc1 != 0) {
16818 SDValue Rnd = Op.getOperand(5);
16819 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16820 X86::STATIC_ROUNDING::CUR_DIRECTION)
16821 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16823 //default rounding mode
16825 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16827 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16828 DAG.getTargetConstant(0, dl,
16832 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16833 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16834 DAG.getValueType(MVT::i1));
16836 case COMI: { // Comparison intrinsics
16837 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16838 SDValue LHS = Op.getOperand(1);
16839 SDValue RHS = Op.getOperand(2);
16840 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16841 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16842 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16843 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16844 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16845 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16847 case COMI_RM: { // Comparison intrinsics with Sae
16848 SDValue LHS = Op.getOperand(1);
16849 SDValue RHS = Op.getOperand(2);
16850 SDValue CC = Op.getOperand(3);
16851 SDValue Sae = Op.getOperand(4);
16852 auto ComiType = TranslateX86ConstCondToX86CC(CC);
16853 // choose between ordered and unordered (comi/ucomi)
16854 unsigned comiOp = std::get<0>(ComiType) ? IntrData->Opc0 : IntrData->Opc1;
16856 if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
16857 X86::STATIC_ROUNDING::CUR_DIRECTION)
16858 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
16860 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
16861 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16862 DAG.getConstant(std::get<1>(ComiType), dl, MVT::i8), Cond);
16863 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16866 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16867 Op.getOperand(1), Op.getOperand(2), DAG);
16869 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16870 Op.getSimpleValueType(),
16872 Op.getOperand(2), DAG),
16873 Op.getOperand(4), Op.getOperand(3), Subtarget,
16875 case COMPRESS_EXPAND_IN_REG: {
16876 SDValue Mask = Op.getOperand(3);
16877 SDValue DataToCompress = Op.getOperand(1);
16878 SDValue PassThru = Op.getOperand(2);
16879 if (isAllOnesConstant(Mask)) // return data as is
16880 return Op.getOperand(1);
16882 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16884 Mask, PassThru, Subtarget, DAG);
16887 SDValue Mask = Op.getOperand(1);
16888 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16889 Mask.getSimpleValueType().getSizeInBits());
16890 Mask = DAG.getBitcast(MaskVT, Mask);
16891 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16894 SDValue Mask = Op.getOperand(3);
16895 MVT VT = Op.getSimpleValueType();
16896 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16897 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16898 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16902 MVT VT = Op.getSimpleValueType();
16903 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2);
16905 SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl);
16906 SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl);
16907 // Arguments should be swapped.
16908 SDValue Res = DAG.getNode(IntrData->Opc0, dl,
16909 MVT::getVectorVT(MVT::i1, VT.getSizeInBits()),
16911 return DAG.getBitcast(VT, Res);
16913 case CONVERT_TO_MASK: {
16914 MVT SrcVT = Op.getOperand(1).getSimpleValueType();
16915 MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements());
16916 MVT BitcastVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits());
16918 SDValue CvtMask = DAG.getNode(IntrData->Opc0, dl, MaskVT,
16920 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16921 DAG.getUNDEF(BitcastVT), CvtMask,
16922 DAG.getIntPtrConstant(0, dl));
16923 return DAG.getBitcast(Op.getValueType(), Res);
16925 case CONVERT_MASK_TO_VEC: {
16926 SDValue Mask = Op.getOperand(1);
16927 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16928 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16929 return DAG.getNode(IntrData->Opc0, dl, VT, VMask);
16931 case BRCST_SUBVEC_TO_VEC: {
16932 SDValue Src = Op.getOperand(1);
16933 SDValue Passthru = Op.getOperand(2);
16934 SDValue Mask = Op.getOperand(3);
16935 EVT resVT = Passthru.getValueType();
16936 SDValue subVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, resVT,
16937 DAG.getUNDEF(resVT), Src,
16938 DAG.getIntPtrConstant(0, dl));
16940 if (Src.getSimpleValueType().is256BitVector() && resVT.is512BitVector())
16941 immVal = DAG.getConstant(0x44, dl, MVT::i8);
16943 immVal = DAG.getConstant(0, dl, MVT::i8);
16944 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16945 subVec, subVec, immVal),
16946 Mask, Passthru, Subtarget, DAG);
16954 default: return SDValue(); // Don't custom lower most intrinsics.
16956 case Intrinsic::x86_avx2_permd:
16957 case Intrinsic::x86_avx2_permps:
16958 // Operands intentionally swapped. Mask is last operand to intrinsic,
16959 // but second operand for node/instruction.
16960 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16961 Op.getOperand(2), Op.getOperand(1));
16963 // ptest and testp intrinsics. The intrinsic these come from are designed to
16964 // return an integer value, not just an instruction so lower it to the ptest
16965 // or testp pattern and a setcc for the result.
16966 case Intrinsic::x86_sse41_ptestz:
16967 case Intrinsic::x86_sse41_ptestc:
16968 case Intrinsic::x86_sse41_ptestnzc:
16969 case Intrinsic::x86_avx_ptestz_256:
16970 case Intrinsic::x86_avx_ptestc_256:
16971 case Intrinsic::x86_avx_ptestnzc_256:
16972 case Intrinsic::x86_avx_vtestz_ps:
16973 case Intrinsic::x86_avx_vtestc_ps:
16974 case Intrinsic::x86_avx_vtestnzc_ps:
16975 case Intrinsic::x86_avx_vtestz_pd:
16976 case Intrinsic::x86_avx_vtestc_pd:
16977 case Intrinsic::x86_avx_vtestnzc_pd:
16978 case Intrinsic::x86_avx_vtestz_ps_256:
16979 case Intrinsic::x86_avx_vtestc_ps_256:
16980 case Intrinsic::x86_avx_vtestnzc_ps_256:
16981 case Intrinsic::x86_avx_vtestz_pd_256:
16982 case Intrinsic::x86_avx_vtestc_pd_256:
16983 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16984 bool IsTestPacked = false;
16987 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16988 case Intrinsic::x86_avx_vtestz_ps:
16989 case Intrinsic::x86_avx_vtestz_pd:
16990 case Intrinsic::x86_avx_vtestz_ps_256:
16991 case Intrinsic::x86_avx_vtestz_pd_256:
16992 IsTestPacked = true; // Fallthrough
16993 case Intrinsic::x86_sse41_ptestz:
16994 case Intrinsic::x86_avx_ptestz_256:
16996 X86CC = X86::COND_E;
16998 case Intrinsic::x86_avx_vtestc_ps:
16999 case Intrinsic::x86_avx_vtestc_pd:
17000 case Intrinsic::x86_avx_vtestc_ps_256:
17001 case Intrinsic::x86_avx_vtestc_pd_256:
17002 IsTestPacked = true; // Fallthrough
17003 case Intrinsic::x86_sse41_ptestc:
17004 case Intrinsic::x86_avx_ptestc_256:
17006 X86CC = X86::COND_B;
17008 case Intrinsic::x86_avx_vtestnzc_ps:
17009 case Intrinsic::x86_avx_vtestnzc_pd:
17010 case Intrinsic::x86_avx_vtestnzc_ps_256:
17011 case Intrinsic::x86_avx_vtestnzc_pd_256:
17012 IsTestPacked = true; // Fallthrough
17013 case Intrinsic::x86_sse41_ptestnzc:
17014 case Intrinsic::x86_avx_ptestnzc_256:
17016 X86CC = X86::COND_A;
17020 SDValue LHS = Op.getOperand(1);
17021 SDValue RHS = Op.getOperand(2);
17022 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
17023 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
17024 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
17025 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
17026 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17028 case Intrinsic::x86_avx512_kortestz_w:
17029 case Intrinsic::x86_avx512_kortestc_w: {
17030 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
17031 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
17032 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
17033 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
17034 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
17035 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17036 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17039 case Intrinsic::x86_sse42_pcmpistria128:
17040 case Intrinsic::x86_sse42_pcmpestria128:
17041 case Intrinsic::x86_sse42_pcmpistric128:
17042 case Intrinsic::x86_sse42_pcmpestric128:
17043 case Intrinsic::x86_sse42_pcmpistrio128:
17044 case Intrinsic::x86_sse42_pcmpestrio128:
17045 case Intrinsic::x86_sse42_pcmpistris128:
17046 case Intrinsic::x86_sse42_pcmpestris128:
17047 case Intrinsic::x86_sse42_pcmpistriz128:
17048 case Intrinsic::x86_sse42_pcmpestriz128: {
17052 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17053 case Intrinsic::x86_sse42_pcmpistria128:
17054 Opcode = X86ISD::PCMPISTRI;
17055 X86CC = X86::COND_A;
17057 case Intrinsic::x86_sse42_pcmpestria128:
17058 Opcode = X86ISD::PCMPESTRI;
17059 X86CC = X86::COND_A;
17061 case Intrinsic::x86_sse42_pcmpistric128:
17062 Opcode = X86ISD::PCMPISTRI;
17063 X86CC = X86::COND_B;
17065 case Intrinsic::x86_sse42_pcmpestric128:
17066 Opcode = X86ISD::PCMPESTRI;
17067 X86CC = X86::COND_B;
17069 case Intrinsic::x86_sse42_pcmpistrio128:
17070 Opcode = X86ISD::PCMPISTRI;
17071 X86CC = X86::COND_O;
17073 case Intrinsic::x86_sse42_pcmpestrio128:
17074 Opcode = X86ISD::PCMPESTRI;
17075 X86CC = X86::COND_O;
17077 case Intrinsic::x86_sse42_pcmpistris128:
17078 Opcode = X86ISD::PCMPISTRI;
17079 X86CC = X86::COND_S;
17081 case Intrinsic::x86_sse42_pcmpestris128:
17082 Opcode = X86ISD::PCMPESTRI;
17083 X86CC = X86::COND_S;
17085 case Intrinsic::x86_sse42_pcmpistriz128:
17086 Opcode = X86ISD::PCMPISTRI;
17087 X86CC = X86::COND_E;
17089 case Intrinsic::x86_sse42_pcmpestriz128:
17090 Opcode = X86ISD::PCMPESTRI;
17091 X86CC = X86::COND_E;
17094 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17095 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17096 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17097 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17098 DAG.getConstant(X86CC, dl, MVT::i8),
17099 SDValue(PCMP.getNode(), 1));
17100 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17103 case Intrinsic::x86_sse42_pcmpistri128:
17104 case Intrinsic::x86_sse42_pcmpestri128: {
17106 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17107 Opcode = X86ISD::PCMPISTRI;
17109 Opcode = X86ISD::PCMPESTRI;
17111 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17112 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17113 return DAG.getNode(Opcode, dl, VTs, NewOps);
17116 case Intrinsic::x86_seh_lsda: {
17117 // Compute the symbol for the LSDA. We know it'll get emitted later.
17118 MachineFunction &MF = DAG.getMachineFunction();
17119 SDValue Op1 = Op.getOperand(1);
17120 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
17121 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
17122 GlobalValue::getRealLinkageName(Fn->getName()));
17124 // Generate a simple absolute symbol reference. This intrinsic is only
17125 // supported on 32-bit Windows, which isn't PIC.
17126 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
17127 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
17130 case Intrinsic::x86_seh_recoverfp: {
17131 SDValue FnOp = Op.getOperand(1);
17132 SDValue IncomingFPOp = Op.getOperand(2);
17133 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
17134 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
17136 report_fatal_error(
17137 "llvm.x86.seh.recoverfp must take a function as the first argument");
17138 return recoverFramePointer(DAG, Fn, IncomingFPOp);
17141 case Intrinsic::localaddress: {
17142 // Returns one of the stack, base, or frame pointer registers, depending on
17143 // which is used to reference local variables.
17144 MachineFunction &MF = DAG.getMachineFunction();
17145 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17147 if (RegInfo->hasBasePointer(MF))
17148 Reg = RegInfo->getBaseRegister();
17149 else // This function handles the SP or FP case.
17150 Reg = RegInfo->getPtrSizedFrameRegister(MF);
17151 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
17156 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17157 SDValue Src, SDValue Mask, SDValue Base,
17158 SDValue Index, SDValue ScaleOp, SDValue Chain,
17159 const X86Subtarget * Subtarget) {
17161 auto *C = cast<ConstantSDNode>(ScaleOp);
17162 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17163 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17164 Index.getSimpleValueType().getVectorNumElements());
17166 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17168 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17170 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17171 Mask.getSimpleValueType().getSizeInBits());
17173 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17174 // are extracted by EXTRACT_SUBVECTOR.
17175 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17176 DAG.getBitcast(BitcastVT, Mask),
17177 DAG.getIntPtrConstant(0, dl));
17179 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17180 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17181 SDValue Segment = DAG.getRegister(0, MVT::i32);
17182 if (Src.getOpcode() == ISD::UNDEF)
17183 Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl);
17184 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17185 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17186 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17187 return DAG.getMergeValues(RetOps, dl);
17190 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17191 SDValue Src, SDValue Mask, SDValue Base,
17192 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17194 auto *C = cast<ConstantSDNode>(ScaleOp);
17195 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17196 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17197 SDValue Segment = DAG.getRegister(0, MVT::i32);
17198 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17199 Index.getSimpleValueType().getVectorNumElements());
17201 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17203 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17205 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17206 Mask.getSimpleValueType().getSizeInBits());
17208 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17209 // are extracted by EXTRACT_SUBVECTOR.
17210 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17211 DAG.getBitcast(BitcastVT, Mask),
17212 DAG.getIntPtrConstant(0, dl));
17214 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17215 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17216 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17217 return SDValue(Res, 1);
17220 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17221 SDValue Mask, SDValue Base, SDValue Index,
17222 SDValue ScaleOp, SDValue Chain) {
17224 auto *C = cast<ConstantSDNode>(ScaleOp);
17225 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17226 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17227 SDValue Segment = DAG.getRegister(0, MVT::i32);
17229 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17231 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17233 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17235 MaskInReg = DAG.getBitcast(MaskVT, Mask);
17236 //SDVTList VTs = DAG.getVTList(MVT::Other);
17237 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17238 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17239 return SDValue(Res, 0);
17242 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17243 // read performance monitor counters (x86_rdpmc).
17244 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17245 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17246 SmallVectorImpl<SDValue> &Results) {
17247 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17248 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17251 // The ECX register is used to select the index of the performance counter
17253 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17255 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17257 // Reads the content of a 64-bit performance counter and returns it in the
17258 // registers EDX:EAX.
17259 if (Subtarget->is64Bit()) {
17260 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17261 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17264 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17265 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17268 Chain = HI.getValue(1);
17270 if (Subtarget->is64Bit()) {
17271 // The EAX register is loaded with the low-order 32 bits. The EDX register
17272 // is loaded with the supported high-order bits of the counter.
17273 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17274 DAG.getConstant(32, DL, MVT::i8));
17275 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17276 Results.push_back(Chain);
17280 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17281 SDValue Ops[] = { LO, HI };
17282 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17283 Results.push_back(Pair);
17284 Results.push_back(Chain);
17287 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17288 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17289 // also used to custom lower READCYCLECOUNTER nodes.
17290 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17291 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17292 SmallVectorImpl<SDValue> &Results) {
17293 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17294 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17297 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17298 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17299 // and the EAX register is loaded with the low-order 32 bits.
17300 if (Subtarget->is64Bit()) {
17301 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17302 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17305 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17306 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17309 SDValue Chain = HI.getValue(1);
17311 if (Opcode == X86ISD::RDTSCP_DAG) {
17312 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17314 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17315 // the ECX register. Add 'ecx' explicitly to the chain.
17316 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17318 // Explicitly store the content of ECX at the location passed in input
17319 // to the 'rdtscp' intrinsic.
17320 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17321 MachinePointerInfo(), false, false, 0);
17324 if (Subtarget->is64Bit()) {
17325 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17326 // the EAX register is loaded with the low-order 32 bits.
17327 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17328 DAG.getConstant(32, DL, MVT::i8));
17329 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17330 Results.push_back(Chain);
17334 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17335 SDValue Ops[] = { LO, HI };
17336 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17337 Results.push_back(Pair);
17338 Results.push_back(Chain);
17341 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17342 SelectionDAG &DAG) {
17343 SmallVector<SDValue, 2> Results;
17345 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17347 return DAG.getMergeValues(Results, DL);
17350 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
17351 MachineFunction &MF = DAG.getMachineFunction();
17352 SDValue Chain = Op.getOperand(0);
17353 SDValue RegNode = Op.getOperand(2);
17354 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
17356 report_fatal_error("EH registrations only live in functions using WinEH");
17358 // Cast the operand to an alloca, and remember the frame index.
17359 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
17361 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
17362 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
17364 // Return the chain operand without making any DAG nodes.
17368 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
17369 /// return truncate Store/MaskedStore Node
17370 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17374 SDValue Mask = Op.getOperand(4);
17375 SDValue DataToTruncate = Op.getOperand(3);
17376 SDValue Addr = Op.getOperand(2);
17377 SDValue Chain = Op.getOperand(0);
17379 MVT VT = DataToTruncate.getSimpleValueType();
17380 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17382 if (isAllOnesConstant(Mask)) // return just a truncate store
17383 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17384 MachinePointerInfo(), SVT, false, false,
17385 SVT.getScalarSizeInBits()/8);
17387 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17388 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17389 Mask.getSimpleValueType().getSizeInBits());
17390 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17391 // are extracted by EXTRACT_SUBVECTOR.
17392 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17393 DAG.getBitcast(BitcastVT, Mask),
17394 DAG.getIntPtrConstant(0, dl));
17396 MachineMemOperand *MMO = DAG.getMachineFunction().
17397 getMachineMemOperand(MachinePointerInfo(),
17398 MachineMemOperand::MOStore, SVT.getStoreSize(),
17399 SVT.getScalarSizeInBits()/8);
17401 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17402 VMask, SVT, MMO, true);
17405 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17406 SelectionDAG &DAG) {
17407 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17409 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17411 if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17412 return MarkEHRegistrationNode(Op, DAG);
17413 if (IntNo == llvm::Intrinsic::x86_flags_read_u32 ||
17414 IntNo == llvm::Intrinsic::x86_flags_read_u64 ||
17415 IntNo == llvm::Intrinsic::x86_flags_write_u32 ||
17416 IntNo == llvm::Intrinsic::x86_flags_write_u64) {
17417 // We need a frame pointer because this will get lowered to a PUSH/POP
17419 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17420 MFI->setHasOpaqueSPAdjustment(true);
17421 // Don't do anything here, we will expand these intrinsics out later
17422 // during ExpandISelPseudos in EmitInstrWithCustomInserter.
17429 switch(IntrData->Type) {
17430 default: llvm_unreachable("Unknown Intrinsic Type");
17433 // Emit the node with the right value type.
17434 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17435 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17437 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17438 // Otherwise return the value from Rand, which is always 0, casted to i32.
17439 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17440 DAG.getConstant(1, dl, Op->getValueType(1)),
17441 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17442 SDValue(Result.getNode(), 1) };
17443 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17444 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17447 // Return { result, isValid, chain }.
17448 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17449 SDValue(Result.getNode(), 2));
17452 //gather(v1, mask, index, base, scale);
17453 SDValue Chain = Op.getOperand(0);
17454 SDValue Src = Op.getOperand(2);
17455 SDValue Base = Op.getOperand(3);
17456 SDValue Index = Op.getOperand(4);
17457 SDValue Mask = Op.getOperand(5);
17458 SDValue Scale = Op.getOperand(6);
17459 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17463 //scatter(base, mask, index, v1, scale);
17464 SDValue Chain = Op.getOperand(0);
17465 SDValue Base = Op.getOperand(2);
17466 SDValue Mask = Op.getOperand(3);
17467 SDValue Index = Op.getOperand(4);
17468 SDValue Src = Op.getOperand(5);
17469 SDValue Scale = Op.getOperand(6);
17470 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17474 SDValue Hint = Op.getOperand(6);
17475 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17476 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17477 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17478 SDValue Chain = Op.getOperand(0);
17479 SDValue Mask = Op.getOperand(2);
17480 SDValue Index = Op.getOperand(3);
17481 SDValue Base = Op.getOperand(4);
17482 SDValue Scale = Op.getOperand(5);
17483 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17485 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17487 SmallVector<SDValue, 2> Results;
17488 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17490 return DAG.getMergeValues(Results, dl);
17492 // Read Performance Monitoring Counters.
17494 SmallVector<SDValue, 2> Results;
17495 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17496 return DAG.getMergeValues(Results, dl);
17498 // XTEST intrinsics.
17500 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17501 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17502 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17503 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17505 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17506 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17507 Ret, SDValue(InTrans.getNode(), 1));
17511 SmallVector<SDValue, 2> Results;
17512 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17513 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17514 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17515 DAG.getConstant(-1, dl, MVT::i8));
17516 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17517 Op.getOperand(4), GenCF.getValue(1));
17518 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17519 Op.getOperand(5), MachinePointerInfo(),
17521 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17522 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17524 Results.push_back(SetCC);
17525 Results.push_back(Store);
17526 return DAG.getMergeValues(Results, dl);
17528 case COMPRESS_TO_MEM: {
17530 SDValue Mask = Op.getOperand(4);
17531 SDValue DataToCompress = Op.getOperand(3);
17532 SDValue Addr = Op.getOperand(2);
17533 SDValue Chain = Op.getOperand(0);
17535 MVT VT = DataToCompress.getSimpleValueType();
17536 if (isAllOnesConstant(Mask)) // return just a store
17537 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17538 MachinePointerInfo(), false, false,
17539 VT.getScalarSizeInBits()/8);
17541 SDValue Compressed =
17542 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17543 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17544 return DAG.getStore(Chain, dl, Compressed, Addr,
17545 MachinePointerInfo(), false, false,
17546 VT.getScalarSizeInBits()/8);
17548 case TRUNCATE_TO_MEM_VI8:
17549 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17550 case TRUNCATE_TO_MEM_VI16:
17551 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17552 case TRUNCATE_TO_MEM_VI32:
17553 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17554 case EXPAND_FROM_MEM: {
17556 SDValue Mask = Op.getOperand(4);
17557 SDValue PassThru = Op.getOperand(3);
17558 SDValue Addr = Op.getOperand(2);
17559 SDValue Chain = Op.getOperand(0);
17560 MVT VT = Op.getSimpleValueType();
17562 if (isAllOnesConstant(Mask)) // return just a load
17563 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17564 false, VT.getScalarSizeInBits()/8);
17566 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17567 false, false, false,
17568 VT.getScalarSizeInBits()/8);
17570 SDValue Results[] = {
17571 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17572 Mask, PassThru, Subtarget, DAG), Chain};
17573 return DAG.getMergeValues(Results, dl);
17578 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17579 SelectionDAG &DAG) const {
17580 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17581 MFI->setReturnAddressIsTaken(true);
17583 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17586 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17588 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17591 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17592 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17593 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17594 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17595 DAG.getNode(ISD::ADD, dl, PtrVT,
17596 FrameAddr, Offset),
17597 MachinePointerInfo(), false, false, false, 0);
17600 // Just load the return address.
17601 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17602 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17603 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17606 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17607 MachineFunction &MF = DAG.getMachineFunction();
17608 MachineFrameInfo *MFI = MF.getFrameInfo();
17609 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17610 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17611 EVT VT = Op.getValueType();
17613 MFI->setFrameAddressIsTaken(true);
17615 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17616 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17617 // is not possible to crawl up the stack without looking at the unwind codes
17619 int FrameAddrIndex = FuncInfo->getFAIndex();
17620 if (!FrameAddrIndex) {
17621 // Set up a frame object for the return address.
17622 unsigned SlotSize = RegInfo->getSlotSize();
17623 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17624 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17625 FuncInfo->setFAIndex(FrameAddrIndex);
17627 return DAG.getFrameIndex(FrameAddrIndex, VT);
17630 unsigned FrameReg =
17631 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17632 SDLoc dl(Op); // FIXME probably not meaningful
17633 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17634 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17635 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17636 "Invalid Frame Register!");
17637 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17639 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17640 MachinePointerInfo(),
17641 false, false, false, 0);
17645 // FIXME? Maybe this could be a TableGen attribute on some registers and
17646 // this table could be generated automatically from RegInfo.
17647 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17648 SelectionDAG &DAG) const {
17649 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17650 const MachineFunction &MF = DAG.getMachineFunction();
17652 unsigned Reg = StringSwitch<unsigned>(RegName)
17653 .Case("esp", X86::ESP)
17654 .Case("rsp", X86::RSP)
17655 .Case("ebp", X86::EBP)
17656 .Case("rbp", X86::RBP)
17659 if (Reg == X86::EBP || Reg == X86::RBP) {
17660 if (!TFI.hasFP(MF))
17661 report_fatal_error("register " + StringRef(RegName) +
17662 " is allocatable: function has no frame pointer");
17665 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17666 unsigned FrameReg =
17667 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17668 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17669 "Invalid Frame Register!");
17677 report_fatal_error("Invalid register name global variable");
17680 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17681 SelectionDAG &DAG) const {
17682 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17683 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17686 unsigned X86TargetLowering::getExceptionPointerRegister(
17687 const Constant *PersonalityFn) const {
17688 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17689 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17691 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17694 unsigned X86TargetLowering::getExceptionSelectorRegister(
17695 const Constant *PersonalityFn) const {
17696 // Funclet personalities don't use selectors (the runtime does the selection).
17697 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17698 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17701 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17702 SDValue Chain = Op.getOperand(0);
17703 SDValue Offset = Op.getOperand(1);
17704 SDValue Handler = Op.getOperand(2);
17707 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17708 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17709 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17710 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17711 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17712 "Invalid Frame Register!");
17713 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17714 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17716 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17717 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17719 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17720 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17722 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17724 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17725 DAG.getRegister(StoreAddrReg, PtrVT));
17728 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17729 SelectionDAG &DAG) const {
17731 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17732 DAG.getVTList(MVT::i32, MVT::Other),
17733 Op.getOperand(0), Op.getOperand(1));
17736 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17737 SelectionDAG &DAG) const {
17739 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17740 Op.getOperand(0), Op.getOperand(1));
17743 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17744 return Op.getOperand(0);
17747 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17748 SelectionDAG &DAG) const {
17749 SDValue Root = Op.getOperand(0);
17750 SDValue Trmp = Op.getOperand(1); // trampoline
17751 SDValue FPtr = Op.getOperand(2); // nested function
17752 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17755 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17756 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17758 if (Subtarget->is64Bit()) {
17759 SDValue OutChains[6];
17761 // Large code-model.
17762 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17763 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17765 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17766 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17768 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17770 // Load the pointer to the nested function into R11.
17771 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17772 SDValue Addr = Trmp;
17773 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17774 Addr, MachinePointerInfo(TrmpAddr),
17777 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17778 DAG.getConstant(2, dl, MVT::i64));
17779 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17780 MachinePointerInfo(TrmpAddr, 2),
17783 // Load the 'nest' parameter value into R10.
17784 // R10 is specified in X86CallingConv.td
17785 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17786 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17787 DAG.getConstant(10, dl, MVT::i64));
17788 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17789 Addr, MachinePointerInfo(TrmpAddr, 10),
17792 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17793 DAG.getConstant(12, dl, MVT::i64));
17794 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17795 MachinePointerInfo(TrmpAddr, 12),
17798 // Jump to the nested function.
17799 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17800 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17801 DAG.getConstant(20, dl, MVT::i64));
17802 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17803 Addr, MachinePointerInfo(TrmpAddr, 20),
17806 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17807 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17808 DAG.getConstant(22, dl, MVT::i64));
17809 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17810 Addr, MachinePointerInfo(TrmpAddr, 22),
17813 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17815 const Function *Func =
17816 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17817 CallingConv::ID CC = Func->getCallingConv();
17822 llvm_unreachable("Unsupported calling convention");
17823 case CallingConv::C:
17824 case CallingConv::X86_StdCall: {
17825 // Pass 'nest' parameter in ECX.
17826 // Must be kept in sync with X86CallingConv.td
17827 NestReg = X86::ECX;
17829 // Check that ECX wasn't needed by an 'inreg' parameter.
17830 FunctionType *FTy = Func->getFunctionType();
17831 const AttributeSet &Attrs = Func->getAttributes();
17833 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17834 unsigned InRegCount = 0;
17837 for (FunctionType::param_iterator I = FTy->param_begin(),
17838 E = FTy->param_end(); I != E; ++I, ++Idx)
17839 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17840 auto &DL = DAG.getDataLayout();
17841 // FIXME: should only count parameters that are lowered to integers.
17842 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17845 if (InRegCount > 2) {
17846 report_fatal_error("Nest register in use - reduce number of inreg"
17852 case CallingConv::X86_FastCall:
17853 case CallingConv::X86_ThisCall:
17854 case CallingConv::Fast:
17855 // Pass 'nest' parameter in EAX.
17856 // Must be kept in sync with X86CallingConv.td
17857 NestReg = X86::EAX;
17861 SDValue OutChains[4];
17862 SDValue Addr, Disp;
17864 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17865 DAG.getConstant(10, dl, MVT::i32));
17866 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17868 // This is storing the opcode for MOV32ri.
17869 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17870 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17871 OutChains[0] = DAG.getStore(Root, dl,
17872 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17873 Trmp, MachinePointerInfo(TrmpAddr),
17876 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17877 DAG.getConstant(1, dl, MVT::i32));
17878 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17879 MachinePointerInfo(TrmpAddr, 1),
17882 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17883 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17884 DAG.getConstant(5, dl, MVT::i32));
17885 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17886 Addr, MachinePointerInfo(TrmpAddr, 5),
17889 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17890 DAG.getConstant(6, dl, MVT::i32));
17891 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17892 MachinePointerInfo(TrmpAddr, 6),
17895 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17899 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17900 SelectionDAG &DAG) const {
17902 The rounding mode is in bits 11:10 of FPSR, and has the following
17904 00 Round to nearest
17909 FLT_ROUNDS, on the other hand, expects the following:
17916 To perform the conversion, we do:
17917 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17920 MachineFunction &MF = DAG.getMachineFunction();
17921 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17922 unsigned StackAlignment = TFI.getStackAlignment();
17923 MVT VT = Op.getSimpleValueType();
17926 // Save FP Control Word to stack slot
17927 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17928 SDValue StackSlot =
17929 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17931 MachineMemOperand *MMO =
17932 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17933 MachineMemOperand::MOStore, 2, 2);
17935 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17936 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17937 DAG.getVTList(MVT::Other),
17938 Ops, MVT::i16, MMO);
17940 // Load FP Control Word from stack slot
17941 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17942 MachinePointerInfo(), false, false, false, 0);
17944 // Transform as necessary
17946 DAG.getNode(ISD::SRL, DL, MVT::i16,
17947 DAG.getNode(ISD::AND, DL, MVT::i16,
17948 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17949 DAG.getConstant(11, DL, MVT::i8));
17951 DAG.getNode(ISD::SRL, DL, MVT::i16,
17952 DAG.getNode(ISD::AND, DL, MVT::i16,
17953 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17954 DAG.getConstant(9, DL, MVT::i8));
17957 DAG.getNode(ISD::AND, DL, MVT::i16,
17958 DAG.getNode(ISD::ADD, DL, MVT::i16,
17959 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17960 DAG.getConstant(1, DL, MVT::i16)),
17961 DAG.getConstant(3, DL, MVT::i16));
17963 return DAG.getNode((VT.getSizeInBits() < 16 ?
17964 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17967 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17969 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17970 // to 512-bit vector.
17971 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17972 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17973 // split the vector, perform operation on it's Lo a Hi part and
17974 // concatenate the results.
17975 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17977 MVT VT = Op.getSimpleValueType();
17978 MVT EltVT = VT.getVectorElementType();
17979 unsigned NumElems = VT.getVectorNumElements();
17981 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17982 // Extend to 512 bit vector.
17983 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17984 "Unsupported value type for operation");
17986 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17987 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17988 DAG.getUNDEF(NewVT),
17990 DAG.getIntPtrConstant(0, dl));
17991 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17993 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17994 DAG.getIntPtrConstant(0, dl));
17997 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17998 "Unsupported element type");
18000 if (16 < NumElems) {
18001 // Split vector, it's Lo and Hi parts will be handled in next iteration.
18003 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
18004 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
18006 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
18007 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
18009 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
18012 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
18014 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
18015 "Unsupported value type for operation");
18017 // Use native supported vector instruction vplzcntd.
18018 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
18019 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
18020 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
18021 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
18023 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
18026 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
18027 SelectionDAG &DAG) {
18028 MVT VT = Op.getSimpleValueType();
18030 unsigned NumBits = VT.getSizeInBits();
18033 if (VT.isVector() && Subtarget->hasAVX512())
18034 return LowerVectorCTLZ_AVX512(Op, DAG);
18036 Op = Op.getOperand(0);
18037 if (VT == MVT::i8) {
18038 // Zero extend to i32 since there is not an i8 bsr.
18040 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18043 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
18044 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18045 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18047 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18050 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
18051 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18054 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18056 // Finally xor with NumBits-1.
18057 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18058 DAG.getConstant(NumBits - 1, dl, OpVT));
18061 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18065 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
18066 SelectionDAG &DAG) {
18067 MVT VT = Op.getSimpleValueType();
18069 unsigned NumBits = VT.getSizeInBits();
18072 Op = Op.getOperand(0);
18073 if (VT == MVT::i8) {
18074 // Zero extend to i32 since there is not an i8 bsr.
18076 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18079 // Issue a bsr (scan bits in reverse).
18080 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18081 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18083 // And xor with NumBits-1.
18084 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18085 DAG.getConstant(NumBits - 1, dl, OpVT));
18088 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18092 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18093 MVT VT = Op.getSimpleValueType();
18094 unsigned NumBits = VT.getScalarSizeInBits();
18097 if (VT.isVector()) {
18098 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18100 SDValue N0 = Op.getOperand(0);
18101 SDValue Zero = DAG.getConstant(0, dl, VT);
18103 // lsb(x) = (x & -x)
18104 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
18105 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
18107 // cttz_undef(x) = (width - 1) - ctlz(lsb)
18108 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
18109 TLI.isOperationLegal(ISD::CTLZ, VT)) {
18110 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
18111 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
18112 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
18115 // cttz(x) = ctpop(lsb - 1)
18116 SDValue One = DAG.getConstant(1, dl, VT);
18117 return DAG.getNode(ISD::CTPOP, dl, VT,
18118 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
18121 assert(Op.getOpcode() == ISD::CTTZ &&
18122 "Only scalar CTTZ requires custom lowering");
18124 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18125 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18126 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
18128 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18131 DAG.getConstant(NumBits, dl, VT),
18132 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18135 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18138 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18139 // ones, and then concatenate the result back.
18140 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18141 MVT VT = Op.getSimpleValueType();
18143 assert(VT.is256BitVector() && VT.isInteger() &&
18144 "Unsupported value type for operation");
18146 unsigned NumElems = VT.getVectorNumElements();
18149 // Extract the LHS vectors
18150 SDValue LHS = Op.getOperand(0);
18151 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18152 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18154 // Extract the RHS vectors
18155 SDValue RHS = Op.getOperand(1);
18156 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18157 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18159 MVT EltVT = VT.getVectorElementType();
18160 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18162 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18163 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18164 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18167 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18168 if (Op.getValueType() == MVT::i1)
18169 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18170 Op.getOperand(0), Op.getOperand(1));
18171 assert(Op.getSimpleValueType().is256BitVector() &&
18172 Op.getSimpleValueType().isInteger() &&
18173 "Only handle AVX 256-bit vector integer operation");
18174 return Lower256IntArith(Op, DAG);
18177 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18178 if (Op.getValueType() == MVT::i1)
18179 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18180 Op.getOperand(0), Op.getOperand(1));
18181 assert(Op.getSimpleValueType().is256BitVector() &&
18182 Op.getSimpleValueType().isInteger() &&
18183 "Only handle AVX 256-bit vector integer operation");
18184 return Lower256IntArith(Op, DAG);
18187 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
18188 assert(Op.getSimpleValueType().is256BitVector() &&
18189 Op.getSimpleValueType().isInteger() &&
18190 "Only handle AVX 256-bit vector integer operation");
18191 return Lower256IntArith(Op, DAG);
18194 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18195 SelectionDAG &DAG) {
18197 MVT VT = Op.getSimpleValueType();
18200 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
18202 // Decompose 256-bit ops into smaller 128-bit ops.
18203 if (VT.is256BitVector() && !Subtarget->hasInt256())
18204 return Lower256IntArith(Op, DAG);
18206 SDValue A = Op.getOperand(0);
18207 SDValue B = Op.getOperand(1);
18209 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
18210 // pairs, multiply and truncate.
18211 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
18212 if (Subtarget->hasInt256()) {
18213 if (VT == MVT::v32i8) {
18214 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
18215 SDValue Lo = DAG.getIntPtrConstant(0, dl);
18216 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
18217 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
18218 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
18219 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
18220 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
18221 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18222 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
18223 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
18226 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
18227 return DAG.getNode(
18228 ISD::TRUNCATE, dl, VT,
18229 DAG.getNode(ISD::MUL, dl, ExVT,
18230 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
18231 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
18234 assert(VT == MVT::v16i8 &&
18235 "Pre-AVX2 support only supports v16i8 multiplication");
18236 MVT ExVT = MVT::v8i16;
18238 // Extract the lo parts and sign extend to i16
18240 if (Subtarget->hasSSE41()) {
18241 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
18242 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
18244 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
18245 -1, 4, -1, 5, -1, 6, -1, 7};
18246 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18247 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18248 ALo = DAG.getBitcast(ExVT, ALo);
18249 BLo = DAG.getBitcast(ExVT, BLo);
18250 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
18251 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
18254 // Extract the hi parts and sign extend to i16
18256 if (Subtarget->hasSSE41()) {
18257 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
18258 -1, -1, -1, -1, -1, -1, -1, -1};
18259 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18260 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18261 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
18262 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
18264 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
18265 -1, 12, -1, 13, -1, 14, -1, 15};
18266 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18267 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18268 AHi = DAG.getBitcast(ExVT, AHi);
18269 BHi = DAG.getBitcast(ExVT, BHi);
18270 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
18271 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
18274 // Multiply, mask the lower 8bits of the lo/hi results and pack
18275 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
18276 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
18277 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
18278 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
18279 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18282 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18283 if (VT == MVT::v4i32) {
18284 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18285 "Should not custom lower when pmuldq is available!");
18287 // Extract the odd parts.
18288 static const int UnpackMask[] = { 1, -1, 3, -1 };
18289 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18290 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18292 // Multiply the even parts.
18293 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18294 // Now multiply odd parts.
18295 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18297 Evens = DAG.getBitcast(VT, Evens);
18298 Odds = DAG.getBitcast(VT, Odds);
18300 // Merge the two vectors back together with a shuffle. This expands into 2
18302 static const int ShufMask[] = { 0, 4, 2, 6 };
18303 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18306 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18307 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18309 // Ahi = psrlqi(a, 32);
18310 // Bhi = psrlqi(b, 32);
18312 // AloBlo = pmuludq(a, b);
18313 // AloBhi = pmuludq(a, Bhi);
18314 // AhiBlo = pmuludq(Ahi, b);
18316 // AloBhi = psllqi(AloBhi, 32);
18317 // AhiBlo = psllqi(AhiBlo, 32);
18318 // return AloBlo + AloBhi + AhiBlo;
18320 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18321 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18323 SDValue AhiBlo = Ahi;
18324 SDValue AloBhi = Bhi;
18325 // Bit cast to 32-bit vectors for MULUDQ
18326 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18327 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18328 A = DAG.getBitcast(MulVT, A);
18329 B = DAG.getBitcast(MulVT, B);
18330 Ahi = DAG.getBitcast(MulVT, Ahi);
18331 Bhi = DAG.getBitcast(MulVT, Bhi);
18333 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18334 // After shifting right const values the result may be all-zero.
18335 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
18336 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18337 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18339 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
18340 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18341 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18344 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18345 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18348 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18349 assert(Subtarget->isTargetWin64() && "Unexpected target");
18350 EVT VT = Op.getValueType();
18351 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18352 "Unexpected return type for lowering");
18356 switch (Op->getOpcode()) {
18357 default: llvm_unreachable("Unexpected request for libcall!");
18358 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18359 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18360 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18361 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18362 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18363 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18367 SDValue InChain = DAG.getEntryNode();
18369 TargetLowering::ArgListTy Args;
18370 TargetLowering::ArgListEntry Entry;
18371 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18372 EVT ArgVT = Op->getOperand(i).getValueType();
18373 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18374 "Unexpected argument type for lowering");
18375 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18376 Entry.Node = StackPtr;
18377 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18379 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18380 Entry.Ty = PointerType::get(ArgTy,0);
18381 Entry.isSExt = false;
18382 Entry.isZExt = false;
18383 Args.push_back(Entry);
18386 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18387 getPointerTy(DAG.getDataLayout()));
18389 TargetLowering::CallLoweringInfo CLI(DAG);
18390 CLI.setDebugLoc(dl).setChain(InChain)
18391 .setCallee(getLibcallCallingConv(LC),
18392 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18393 Callee, std::move(Args), 0)
18394 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18396 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18397 return DAG.getBitcast(VT, CallInfo.first);
18400 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18401 SelectionDAG &DAG) {
18402 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18403 MVT VT = Op0.getSimpleValueType();
18406 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18407 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18409 // PMULxD operations multiply each even value (starting at 0) of LHS with
18410 // the related value of RHS and produce a widen result.
18411 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18412 // => <2 x i64> <ae|cg>
18414 // In other word, to have all the results, we need to perform two PMULxD:
18415 // 1. one with the even values.
18416 // 2. one with the odd values.
18417 // To achieve #2, with need to place the odd values at an even position.
18419 // Place the odd value at an even position (basically, shift all values 1
18420 // step to the left):
18421 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18422 // <a|b|c|d> => <b|undef|d|undef>
18423 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18424 // <e|f|g|h> => <f|undef|h|undef>
18425 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18427 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18429 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18430 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18432 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18433 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18434 // => <2 x i64> <ae|cg>
18435 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18436 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18437 // => <2 x i64> <bf|dh>
18438 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18440 // Shuffle it back into the right order.
18441 SDValue Highs, Lows;
18442 if (VT == MVT::v8i32) {
18443 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18444 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18445 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18446 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18448 const int HighMask[] = {1, 5, 3, 7};
18449 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18450 const int LowMask[] = {0, 4, 2, 6};
18451 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18454 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18455 // unsigned multiply.
18456 if (IsSigned && !Subtarget->hasSSE41()) {
18457 SDValue ShAmt = DAG.getConstant(
18459 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18460 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18461 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18462 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18463 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18465 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18466 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18469 // The first result of MUL_LOHI is actually the low value, followed by the
18471 SDValue Ops[] = {Lows, Highs};
18472 return DAG.getMergeValues(Ops, dl);
18475 // Return true if the required (according to Opcode) shift-imm form is natively
18476 // supported by the Subtarget
18477 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18479 if (VT.getScalarSizeInBits() < 16)
18482 if (VT.is512BitVector() &&
18483 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18486 bool LShift = VT.is128BitVector() ||
18487 (VT.is256BitVector() && Subtarget->hasInt256());
18489 bool AShift = LShift && (Subtarget->hasVLX() ||
18490 (VT != MVT::v2i64 && VT != MVT::v4i64));
18491 return (Opcode == ISD::SRA) ? AShift : LShift;
18494 // The shift amount is a variable, but it is the same for all vector lanes.
18495 // These instructions are defined together with shift-immediate.
18497 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18499 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18502 // Return true if the required (according to Opcode) variable-shift form is
18503 // natively supported by the Subtarget
18504 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18507 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18510 // vXi16 supported only on AVX-512, BWI
18511 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18514 if (VT.is512BitVector() || Subtarget->hasVLX())
18517 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18518 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18519 return (Opcode == ISD::SRA) ? AShift : LShift;
18522 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18523 const X86Subtarget *Subtarget) {
18524 MVT VT = Op.getSimpleValueType();
18526 SDValue R = Op.getOperand(0);
18527 SDValue Amt = Op.getOperand(1);
18529 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18530 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18532 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18533 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18534 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18535 SDValue Ex = DAG.getBitcast(ExVT, R);
18537 if (ShiftAmt >= 32) {
18538 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18540 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18541 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18542 ShiftAmt - 32, DAG);
18543 if (VT == MVT::v2i64)
18544 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18545 if (VT == MVT::v4i64)
18546 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18547 {9, 1, 11, 3, 13, 5, 15, 7});
18549 // SRA upper i32, SHL whole i64 and select lower i32.
18550 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18553 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18554 Lower = DAG.getBitcast(ExVT, Lower);
18555 if (VT == MVT::v2i64)
18556 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18557 if (VT == MVT::v4i64)
18558 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18559 {8, 1, 10, 3, 12, 5, 14, 7});
18561 return DAG.getBitcast(VT, Ex);
18564 // Optimize shl/srl/sra with constant shift amount.
18565 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18566 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18567 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18569 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18570 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18572 // i64 SRA needs to be performed as partial shifts.
18573 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18574 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18575 return ArithmeticShiftRight64(ShiftAmt);
18577 if (VT == MVT::v16i8 ||
18578 (Subtarget->hasInt256() && VT == MVT::v32i8) ||
18579 VT == MVT::v64i8) {
18580 unsigned NumElts = VT.getVectorNumElements();
18581 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18583 // Simple i8 add case
18584 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18585 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18587 // ashr(R, 7) === cmp_slt(R, 0)
18588 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18589 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18590 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18593 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18594 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18597 if (Op.getOpcode() == ISD::SHL) {
18598 // Make a large shift.
18599 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18601 SHL = DAG.getBitcast(VT, SHL);
18602 // Zero out the rightmost bits.
18603 return DAG.getNode(ISD::AND, dl, VT, SHL,
18604 DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, VT));
18606 if (Op.getOpcode() == ISD::SRL) {
18607 // Make a large shift.
18608 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18610 SRL = DAG.getBitcast(VT, SRL);
18611 // Zero out the leftmost bits.
18612 return DAG.getNode(ISD::AND, dl, VT, SRL,
18613 DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, VT));
18615 if (Op.getOpcode() == ISD::SRA) {
18616 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18617 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18619 SDValue Mask = DAG.getConstant(128 >> ShiftAmt, dl, VT);
18620 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18621 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18624 llvm_unreachable("Unknown shift opcode.");
18629 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18630 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18631 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18633 // Peek through any splat that was introduced for i64 shift vectorization.
18634 int SplatIndex = -1;
18635 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18636 if (SVN->isSplat()) {
18637 SplatIndex = SVN->getSplatIndex();
18638 Amt = Amt.getOperand(0);
18639 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18640 "Splat shuffle referencing second operand");
18643 if (Amt.getOpcode() != ISD::BITCAST ||
18644 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18647 Amt = Amt.getOperand(0);
18648 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18649 VT.getVectorNumElements();
18650 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18651 uint64_t ShiftAmt = 0;
18652 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18653 for (unsigned i = 0; i != Ratio; ++i) {
18654 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18658 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18661 // Check remaining shift amounts (if not a splat).
18662 if (SplatIndex < 0) {
18663 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18664 uint64_t ShAmt = 0;
18665 for (unsigned j = 0; j != Ratio; ++j) {
18666 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18670 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18672 if (ShAmt != ShiftAmt)
18677 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18678 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18680 if (Op.getOpcode() == ISD::SRA)
18681 return ArithmeticShiftRight64(ShiftAmt);
18687 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18688 const X86Subtarget* Subtarget) {
18689 MVT VT = Op.getSimpleValueType();
18691 SDValue R = Op.getOperand(0);
18692 SDValue Amt = Op.getOperand(1);
18694 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18695 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18697 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18698 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18700 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18702 MVT EltVT = VT.getVectorElementType();
18704 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18705 // Check if this build_vector node is doing a splat.
18706 // If so, then set BaseShAmt equal to the splat value.
18707 BaseShAmt = BV->getSplatValue();
18708 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18709 BaseShAmt = SDValue();
18711 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18712 Amt = Amt.getOperand(0);
18714 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18715 if (SVN && SVN->isSplat()) {
18716 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18717 SDValue InVec = Amt.getOperand(0);
18718 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18719 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18720 "Unexpected shuffle index found!");
18721 BaseShAmt = InVec.getOperand(SplatIdx);
18722 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18723 if (ConstantSDNode *C =
18724 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18725 if (C->getZExtValue() == SplatIdx)
18726 BaseShAmt = InVec.getOperand(1);
18731 // Avoid introducing an extract element from a shuffle.
18732 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18733 DAG.getIntPtrConstant(SplatIdx, dl));
18737 if (BaseShAmt.getNode()) {
18738 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18739 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18740 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18741 else if (EltVT.bitsLT(MVT::i32))
18742 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18744 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18748 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18749 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18750 Amt.getOpcode() == ISD::BITCAST &&
18751 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18752 Amt = Amt.getOperand(0);
18753 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18754 VT.getVectorNumElements();
18755 std::vector<SDValue> Vals(Ratio);
18756 for (unsigned i = 0; i != Ratio; ++i)
18757 Vals[i] = Amt.getOperand(i);
18758 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18759 for (unsigned j = 0; j != Ratio; ++j)
18760 if (Vals[j] != Amt.getOperand(i + j))
18764 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18765 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18770 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18771 SelectionDAG &DAG) {
18772 MVT VT = Op.getSimpleValueType();
18774 SDValue R = Op.getOperand(0);
18775 SDValue Amt = Op.getOperand(1);
18777 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18778 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18780 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18783 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18786 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18789 // XOP has 128-bit variable logical/arithmetic shifts.
18790 // +ve/-ve Amt = shift left/right.
18791 if (Subtarget->hasXOP() &&
18792 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18793 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18794 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18795 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18796 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18798 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18799 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18800 if (Op.getOpcode() == ISD::SRA)
18801 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18804 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18805 // shifts per-lane and then shuffle the partial results back together.
18806 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18807 // Splat the shift amounts so the scalar shifts above will catch it.
18808 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18809 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18810 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18811 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18812 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18815 // i64 vector arithmetic shift can be emulated with the transform:
18816 // M = lshr(SIGN_BIT, Amt)
18817 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18818 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18819 Op.getOpcode() == ISD::SRA) {
18820 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18821 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18822 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18823 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18824 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18828 // If possible, lower this packed shift into a vector multiply instead of
18829 // expanding it into a sequence of scalar shifts.
18830 // Do this only if the vector shift count is a constant build_vector.
18831 if (Op.getOpcode() == ISD::SHL &&
18832 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18833 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18834 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18835 SmallVector<SDValue, 8> Elts;
18836 MVT SVT = VT.getVectorElementType();
18837 unsigned SVTBits = SVT.getSizeInBits();
18838 APInt One(SVTBits, 1);
18839 unsigned NumElems = VT.getVectorNumElements();
18841 for (unsigned i=0; i !=NumElems; ++i) {
18842 SDValue Op = Amt->getOperand(i);
18843 if (Op->getOpcode() == ISD::UNDEF) {
18844 Elts.push_back(Op);
18848 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18849 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18850 uint64_t ShAmt = C.getZExtValue();
18851 if (ShAmt >= SVTBits) {
18852 Elts.push_back(DAG.getUNDEF(SVT));
18855 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18857 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18858 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18861 // Lower SHL with variable shift amount.
18862 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18863 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18865 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18866 DAG.getConstant(0x3f800000U, dl, VT));
18867 Op = DAG.getBitcast(MVT::v4f32, Op);
18868 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18869 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18872 // If possible, lower this shift as a sequence of two shifts by
18873 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18875 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18877 // Could be rewritten as:
18878 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18880 // The advantage is that the two shifts from the example would be
18881 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18882 // the vector shift into four scalar shifts plus four pairs of vector
18884 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18885 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18886 unsigned TargetOpcode = X86ISD::MOVSS;
18887 bool CanBeSimplified;
18888 // The splat value for the first packed shift (the 'X' from the example).
18889 SDValue Amt1 = Amt->getOperand(0);
18890 // The splat value for the second packed shift (the 'Y' from the example).
18891 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18892 Amt->getOperand(2);
18894 // See if it is possible to replace this node with a sequence of
18895 // two shifts followed by a MOVSS/MOVSD
18896 if (VT == MVT::v4i32) {
18897 // Check if it is legal to use a MOVSS.
18898 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18899 Amt2 == Amt->getOperand(3);
18900 if (!CanBeSimplified) {
18901 // Otherwise, check if we can still simplify this node using a MOVSD.
18902 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18903 Amt->getOperand(2) == Amt->getOperand(3);
18904 TargetOpcode = X86ISD::MOVSD;
18905 Amt2 = Amt->getOperand(2);
18908 // Do similar checks for the case where the machine value type
18910 CanBeSimplified = Amt1 == Amt->getOperand(1);
18911 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18912 CanBeSimplified = Amt2 == Amt->getOperand(i);
18914 if (!CanBeSimplified) {
18915 TargetOpcode = X86ISD::MOVSD;
18916 CanBeSimplified = true;
18917 Amt2 = Amt->getOperand(4);
18918 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18919 CanBeSimplified = Amt1 == Amt->getOperand(i);
18920 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18921 CanBeSimplified = Amt2 == Amt->getOperand(j);
18925 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18926 isa<ConstantSDNode>(Amt2)) {
18927 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18928 MVT CastVT = MVT::v4i32;
18930 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18931 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18933 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18934 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18935 if (TargetOpcode == X86ISD::MOVSD)
18936 CastVT = MVT::v2i64;
18937 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18938 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18939 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18941 return DAG.getBitcast(VT, Result);
18945 // v4i32 Non Uniform Shifts.
18946 // If the shift amount is constant we can shift each lane using the SSE2
18947 // immediate shifts, else we need to zero-extend each lane to the lower i64
18948 // and shift using the SSE2 variable shifts.
18949 // The separate results can then be blended together.
18950 if (VT == MVT::v4i32) {
18951 unsigned Opc = Op.getOpcode();
18952 SDValue Amt0, Amt1, Amt2, Amt3;
18953 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18954 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18955 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18956 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18957 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18959 // ISD::SHL is handled above but we include it here for completeness.
18962 llvm_unreachable("Unknown target vector shift node");
18964 Opc = X86ISD::VSHL;
18967 Opc = X86ISD::VSRL;
18970 Opc = X86ISD::VSRA;
18973 // The SSE2 shifts use the lower i64 as the same shift amount for
18974 // all lanes and the upper i64 is ignored. These shuffle masks
18975 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18976 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18977 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18978 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18979 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18980 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18983 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18984 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18985 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18986 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18987 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18988 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18989 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18992 if (VT == MVT::v16i8 ||
18993 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18994 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18995 unsigned ShiftOpcode = Op->getOpcode();
18997 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18998 // On SSE41 targets we make use of the fact that VSELECT lowers
18999 // to PBLENDVB which selects bytes based just on the sign bit.
19000 if (Subtarget->hasSSE41()) {
19001 V0 = DAG.getBitcast(VT, V0);
19002 V1 = DAG.getBitcast(VT, V1);
19003 Sel = DAG.getBitcast(VT, Sel);
19004 return DAG.getBitcast(SelVT,
19005 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
19007 // On pre-SSE41 targets we test for the sign bit by comparing to
19008 // zero - a negative value will set all bits of the lanes to true
19009 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
19010 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
19011 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
19012 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
19015 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
19016 // We can safely do this using i16 shifts as we're only interested in
19017 // the 3 lower bits of each byte.
19018 Amt = DAG.getBitcast(ExtVT, Amt);
19019 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
19020 Amt = DAG.getBitcast(VT, Amt);
19022 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
19023 // r = VSELECT(r, shift(r, 4), a);
19025 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
19026 R = SignBitSelect(VT, Amt, M, R);
19029 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19031 // r = VSELECT(r, shift(r, 2), a);
19032 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
19033 R = SignBitSelect(VT, Amt, M, R);
19036 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19038 // return VSELECT(r, shift(r, 1), a);
19039 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
19040 R = SignBitSelect(VT, Amt, M, R);
19044 if (Op->getOpcode() == ISD::SRA) {
19045 // For SRA we need to unpack each byte to the higher byte of a i16 vector
19046 // so we can correctly sign extend. We don't care what happens to the
19048 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
19049 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
19050 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
19051 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
19052 ALo = DAG.getBitcast(ExtVT, ALo);
19053 AHi = DAG.getBitcast(ExtVT, AHi);
19054 RLo = DAG.getBitcast(ExtVT, RLo);
19055 RHi = DAG.getBitcast(ExtVT, RHi);
19057 // r = VSELECT(r, shift(r, 4), a);
19058 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19059 DAG.getConstant(4, dl, ExtVT));
19060 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19061 DAG.getConstant(4, dl, ExtVT));
19062 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19063 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19066 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19067 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19069 // r = VSELECT(r, shift(r, 2), a);
19070 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19071 DAG.getConstant(2, dl, ExtVT));
19072 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19073 DAG.getConstant(2, dl, ExtVT));
19074 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19075 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19078 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19079 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19081 // r = VSELECT(r, shift(r, 1), a);
19082 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19083 DAG.getConstant(1, dl, ExtVT));
19084 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19085 DAG.getConstant(1, dl, ExtVT));
19086 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19087 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19089 // Logical shift the result back to the lower byte, leaving a zero upper
19091 // meaning that we can safely pack with PACKUSWB.
19093 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
19095 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
19096 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
19100 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
19101 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
19102 // solution better.
19103 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
19104 MVT ExtVT = MVT::v8i32;
19106 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
19107 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
19108 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
19109 return DAG.getNode(ISD::TRUNCATE, dl, VT,
19110 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
19113 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
19114 MVT ExtVT = MVT::v8i32;
19115 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
19116 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
19117 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
19118 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
19119 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
19120 ALo = DAG.getBitcast(ExtVT, ALo);
19121 AHi = DAG.getBitcast(ExtVT, AHi);
19122 RLo = DAG.getBitcast(ExtVT, RLo);
19123 RHi = DAG.getBitcast(ExtVT, RHi);
19124 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
19125 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
19126 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
19127 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
19128 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
19131 if (VT == MVT::v8i16) {
19132 unsigned ShiftOpcode = Op->getOpcode();
19134 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
19135 // On SSE41 targets we make use of the fact that VSELECT lowers
19136 // to PBLENDVB which selects bytes based just on the sign bit.
19137 if (Subtarget->hasSSE41()) {
19138 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
19139 V0 = DAG.getBitcast(ExtVT, V0);
19140 V1 = DAG.getBitcast(ExtVT, V1);
19141 Sel = DAG.getBitcast(ExtVT, Sel);
19142 return DAG.getBitcast(
19143 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
19145 // On pre-SSE41 targets we splat the sign bit - a negative value will
19146 // set all bits of the lanes to true and VSELECT uses that in
19147 // its OR(AND(V0,C),AND(V1,~C)) lowering.
19149 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
19150 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
19153 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
19154 if (Subtarget->hasSSE41()) {
19155 // On SSE41 targets we need to replicate the shift mask in both
19156 // bytes for PBLENDVB.
19159 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
19160 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
19162 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
19165 // r = VSELECT(r, shift(r, 8), a);
19166 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
19167 R = SignBitSelect(Amt, M, R);
19170 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19172 // r = VSELECT(r, shift(r, 4), a);
19173 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
19174 R = SignBitSelect(Amt, M, R);
19177 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19179 // r = VSELECT(r, shift(r, 2), a);
19180 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
19181 R = SignBitSelect(Amt, M, R);
19184 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19186 // return VSELECT(r, shift(r, 1), a);
19187 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
19188 R = SignBitSelect(Amt, M, R);
19192 // Decompose 256-bit shifts into smaller 128-bit shifts.
19193 if (VT.is256BitVector()) {
19194 unsigned NumElems = VT.getVectorNumElements();
19195 MVT EltVT = VT.getVectorElementType();
19196 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19198 // Extract the two vectors
19199 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
19200 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
19202 // Recreate the shift amount vectors
19203 SDValue Amt1, Amt2;
19204 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
19205 // Constant shift amount
19206 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
19207 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
19208 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
19210 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
19211 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19213 // Variable shift amount
19214 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19215 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19218 // Issue new vector shifts for the smaller types
19219 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19220 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19222 // Concatenate the result back
19223 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19229 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
19230 SelectionDAG &DAG) {
19231 MVT VT = Op.getSimpleValueType();
19233 SDValue R = Op.getOperand(0);
19234 SDValue Amt = Op.getOperand(1);
19236 assert(VT.isVector() && "Custom lowering only for vector rotates!");
19237 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
19238 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
19240 // XOP has 128-bit vector variable + immediate rotates.
19241 // +ve/-ve Amt = rotate left/right.
19243 // Split 256-bit integers.
19244 if (VT.is256BitVector())
19245 return Lower256IntArith(Op, DAG);
19247 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
19249 // Attempt to rotate by immediate.
19250 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
19251 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
19252 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
19253 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
19254 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
19255 DAG.getConstant(RotateAmt, DL, MVT::i8));
19259 // Use general rotate by variable (per-element).
19260 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
19263 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19264 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19265 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19266 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19267 // has only one use.
19268 SDNode *N = Op.getNode();
19269 SDValue LHS = N->getOperand(0);
19270 SDValue RHS = N->getOperand(1);
19271 unsigned BaseOp = 0;
19274 switch (Op.getOpcode()) {
19275 default: llvm_unreachable("Unknown ovf instruction!");
19277 // A subtract of one will be selected as a INC. Note that INC doesn't
19278 // set CF, so we can't do this for UADDO.
19279 if (isOneConstant(RHS)) {
19280 BaseOp = X86ISD::INC;
19281 Cond = X86::COND_O;
19284 BaseOp = X86ISD::ADD;
19285 Cond = X86::COND_O;
19288 BaseOp = X86ISD::ADD;
19289 Cond = X86::COND_B;
19292 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19293 // set CF, so we can't do this for USUBO.
19294 if (isOneConstant(RHS)) {
19295 BaseOp = X86ISD::DEC;
19296 Cond = X86::COND_O;
19299 BaseOp = X86ISD::SUB;
19300 Cond = X86::COND_O;
19303 BaseOp = X86ISD::SUB;
19304 Cond = X86::COND_B;
19307 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19308 Cond = X86::COND_O;
19310 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19311 if (N->getValueType(0) == MVT::i8) {
19312 BaseOp = X86ISD::UMUL8;
19313 Cond = X86::COND_O;
19316 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19318 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19321 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19322 DAG.getConstant(X86::COND_O, DL, MVT::i32),
19323 SDValue(Sum.getNode(), 2));
19325 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19329 // Also sets EFLAGS.
19330 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19331 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19334 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19335 DAG.getConstant(Cond, DL, MVT::i32),
19336 SDValue(Sum.getNode(), 1));
19338 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19341 /// Returns true if the operand type is exactly twice the native width, and
19342 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19343 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19344 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19345 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
19346 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19349 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19350 else if (OpWidth == 128)
19351 return Subtarget->hasCmpxchg16b();
19356 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19357 return needsCmpXchgNb(SI->getValueOperand()->getType());
19360 // Note: this turns large loads into lock cmpxchg8b/16b.
19361 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19362 TargetLowering::AtomicExpansionKind
19363 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19364 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19365 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
19366 : AtomicExpansionKind::None;
19369 TargetLowering::AtomicExpansionKind
19370 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19371 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19372 Type *MemType = AI->getType();
19374 // If the operand is too big, we must see if cmpxchg8/16b is available
19375 // and default to library calls otherwise.
19376 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
19377 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19378 : AtomicExpansionKind::None;
19381 AtomicRMWInst::BinOp Op = AI->getOperation();
19384 llvm_unreachable("Unknown atomic operation");
19385 case AtomicRMWInst::Xchg:
19386 case AtomicRMWInst::Add:
19387 case AtomicRMWInst::Sub:
19388 // It's better to use xadd, xsub or xchg for these in all cases.
19389 return AtomicExpansionKind::None;
19390 case AtomicRMWInst::Or:
19391 case AtomicRMWInst::And:
19392 case AtomicRMWInst::Xor:
19393 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19394 // prefix to a normal instruction for these operations.
19395 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19396 : AtomicExpansionKind::None;
19397 case AtomicRMWInst::Nand:
19398 case AtomicRMWInst::Max:
19399 case AtomicRMWInst::Min:
19400 case AtomicRMWInst::UMax:
19401 case AtomicRMWInst::UMin:
19402 // These always require a non-trivial set of data operations on x86. We must
19403 // use a cmpxchg loop.
19404 return AtomicExpansionKind::CmpXChg;
19408 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19409 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19410 // no-sse2). There isn't any reason to disable it if the target processor
19412 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19416 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19417 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19418 Type *MemType = AI->getType();
19419 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19420 // there is no benefit in turning such RMWs into loads, and it is actually
19421 // harmful as it introduces a mfence.
19422 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19425 auto Builder = IRBuilder<>(AI);
19426 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19427 auto SynchScope = AI->getSynchScope();
19428 // We must restrict the ordering to avoid generating loads with Release or
19429 // ReleaseAcquire orderings.
19430 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19431 auto Ptr = AI->getPointerOperand();
19433 // Before the load we need a fence. Here is an example lifted from
19434 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19437 // x.store(1, relaxed);
19438 // r1 = y.fetch_add(0, release);
19440 // y.fetch_add(42, acquire);
19441 // r2 = x.load(relaxed);
19442 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19443 // lowered to just a load without a fence. A mfence flushes the store buffer,
19444 // making the optimization clearly correct.
19445 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19446 // otherwise, we might be able to be more aggressive on relaxed idempotent
19447 // rmw. In practice, they do not look useful, so we don't try to be
19448 // especially clever.
19449 if (SynchScope == SingleThread)
19450 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19451 // the IR level, so we must wrap it in an intrinsic.
19454 if (!hasMFENCE(*Subtarget))
19455 // FIXME: it might make sense to use a locked operation here but on a
19456 // different cache-line to prevent cache-line bouncing. In practice it
19457 // is probably a small win, and x86 processors without mfence are rare
19458 // enough that we do not bother.
19462 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19463 Builder.CreateCall(MFence, {});
19465 // Finally we can emit the atomic load.
19466 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19467 AI->getType()->getPrimitiveSizeInBits());
19468 Loaded->setAtomic(Order, SynchScope);
19469 AI->replaceAllUsesWith(Loaded);
19470 AI->eraseFromParent();
19474 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19475 SelectionDAG &DAG) {
19477 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19478 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19479 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19480 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19482 // The only fence that needs an instruction is a sequentially-consistent
19483 // cross-thread fence.
19484 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19485 if (hasMFENCE(*Subtarget))
19486 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19488 SDValue Chain = Op.getOperand(0);
19489 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19491 DAG.getRegister(X86::ESP, MVT::i32), // Base
19492 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19493 DAG.getRegister(0, MVT::i32), // Index
19494 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19495 DAG.getRegister(0, MVT::i32), // Segment.
19499 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19500 return SDValue(Res, 0);
19503 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19504 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19507 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19508 SelectionDAG &DAG) {
19509 MVT T = Op.getSimpleValueType();
19513 switch(T.SimpleTy) {
19514 default: llvm_unreachable("Invalid value type!");
19515 case MVT::i8: Reg = X86::AL; size = 1; break;
19516 case MVT::i16: Reg = X86::AX; size = 2; break;
19517 case MVT::i32: Reg = X86::EAX; size = 4; break;
19519 assert(Subtarget->is64Bit() && "Node not type legal!");
19520 Reg = X86::RAX; size = 8;
19523 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19524 Op.getOperand(2), SDValue());
19525 SDValue Ops[] = { cpIn.getValue(0),
19528 DAG.getTargetConstant(size, DL, MVT::i8),
19529 cpIn.getValue(1) };
19530 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19531 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19532 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19536 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19537 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19538 MVT::i32, cpOut.getValue(2));
19539 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19540 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19543 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19544 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19545 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19549 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19550 SelectionDAG &DAG) {
19551 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19552 MVT DstVT = Op.getSimpleValueType();
19554 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8 ||
19555 SrcVT == MVT::i64) {
19556 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19557 if (DstVT != MVT::f64)
19558 // This conversion needs to be expanded.
19561 SDValue Op0 = Op->getOperand(0);
19562 SmallVector<SDValue, 16> Elts;
19566 if (SrcVT.isVector()) {
19567 NumElts = SrcVT.getVectorNumElements();
19568 SVT = SrcVT.getVectorElementType();
19570 // Widen the vector in input in the case of MVT::v2i32.
19571 // Example: from MVT::v2i32 to MVT::v4i32.
19572 for (unsigned i = 0, e = NumElts; i != e; ++i)
19573 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, Op0,
19574 DAG.getIntPtrConstant(i, dl)));
19576 assert(SrcVT == MVT::i64 && !Subtarget->is64Bit() &&
19577 "Unexpected source type in LowerBITCAST");
19578 Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
19579 DAG.getIntPtrConstant(0, dl)));
19580 Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
19581 DAG.getIntPtrConstant(1, dl)));
19585 // Explicitly mark the extra elements as Undef.
19586 Elts.append(NumElts, DAG.getUNDEF(SVT));
19588 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19589 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19590 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19591 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19592 DAG.getIntPtrConstant(0, dl));
19595 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19596 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19597 assert((DstVT == MVT::i64 ||
19598 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19599 "Unexpected custom BITCAST");
19600 // i64 <=> MMX conversions are Legal.
19601 if (SrcVT==MVT::i64 && DstVT.isVector())
19603 if (DstVT==MVT::i64 && SrcVT.isVector())
19605 // MMX <=> MMX conversions are Legal.
19606 if (SrcVT.isVector() && DstVT.isVector())
19608 // All other conversions need to be expanded.
19612 /// Compute the horizontal sum of bytes in V for the elements of VT.
19614 /// Requires V to be a byte vector and VT to be an integer vector type with
19615 /// wider elements than V's type. The width of the elements of VT determines
19616 /// how many bytes of V are summed horizontally to produce each element of the
19618 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19619 const X86Subtarget *Subtarget,
19620 SelectionDAG &DAG) {
19622 MVT ByteVecVT = V.getSimpleValueType();
19623 MVT EltVT = VT.getVectorElementType();
19624 int NumElts = VT.getVectorNumElements();
19625 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19626 "Expected value to have byte element type.");
19627 assert(EltVT != MVT::i8 &&
19628 "Horizontal byte sum only makes sense for wider elements!");
19629 unsigned VecSize = VT.getSizeInBits();
19630 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19632 // PSADBW instruction horizontally add all bytes and leave the result in i64
19633 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19634 if (EltVT == MVT::i64) {
19635 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19636 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19637 V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
19638 return DAG.getBitcast(VT, V);
19641 if (EltVT == MVT::i32) {
19642 // We unpack the low half and high half into i32s interleaved with zeros so
19643 // that we can use PSADBW to horizontally sum them. The most useful part of
19644 // this is that it lines up the results of two PSADBW instructions to be
19645 // two v2i64 vectors which concatenated are the 4 population counts. We can
19646 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19647 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19648 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19649 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19651 // Do the horizontal sums into two v2i64s.
19652 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19653 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19654 Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19655 DAG.getBitcast(ByteVecVT, Low), Zeros);
19656 High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19657 DAG.getBitcast(ByteVecVT, High), Zeros);
19659 // Merge them together.
19660 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19661 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19662 DAG.getBitcast(ShortVecVT, Low),
19663 DAG.getBitcast(ShortVecVT, High));
19665 return DAG.getBitcast(VT, V);
19668 // The only element type left is i16.
19669 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19671 // To obtain pop count for each i16 element starting from the pop count for
19672 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19673 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19674 // directly supported.
19675 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19676 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19677 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19678 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19679 DAG.getBitcast(ByteVecVT, V));
19680 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19683 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19684 const X86Subtarget *Subtarget,
19685 SelectionDAG &DAG) {
19686 MVT VT = Op.getSimpleValueType();
19687 MVT EltVT = VT.getVectorElementType();
19688 unsigned VecSize = VT.getSizeInBits();
19690 // Implement a lookup table in register by using an algorithm based on:
19691 // http://wm.ite.pl/articles/sse-popcount.html
19693 // The general idea is that every lower byte nibble in the input vector is an
19694 // index into a in-register pre-computed pop count table. We then split up the
19695 // input vector in two new ones: (1) a vector with only the shifted-right
19696 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19697 // masked out higher ones) for each byte. PSHUB is used separately with both
19698 // to index the in-register table. Next, both are added and the result is a
19699 // i8 vector where each element contains the pop count for input byte.
19701 // To obtain the pop count for elements != i8, we follow up with the same
19702 // approach and use additional tricks as described below.
19704 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19705 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19706 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19707 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19709 int NumByteElts = VecSize / 8;
19710 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19711 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19712 SmallVector<SDValue, 16> LUTVec;
19713 for (int i = 0; i < NumByteElts; ++i)
19714 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19715 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19716 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19717 DAG.getConstant(0x0F, DL, MVT::i8));
19718 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19721 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19722 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19723 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19726 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19728 // The input vector is used as the shuffle mask that index elements into the
19729 // LUT. After counting low and high nibbles, add the vector to obtain the
19730 // final pop count per i8 element.
19731 SDValue HighPopCnt =
19732 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19733 SDValue LowPopCnt =
19734 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19735 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19737 if (EltVT == MVT::i8)
19740 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19743 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19744 const X86Subtarget *Subtarget,
19745 SelectionDAG &DAG) {
19746 MVT VT = Op.getSimpleValueType();
19747 assert(VT.is128BitVector() &&
19748 "Only 128-bit vector bitmath lowering supported.");
19750 int VecSize = VT.getSizeInBits();
19751 MVT EltVT = VT.getVectorElementType();
19752 int Len = EltVT.getSizeInBits();
19754 // This is the vectorized version of the "best" algorithm from
19755 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19756 // with a minor tweak to use a series of adds + shifts instead of vector
19757 // multiplications. Implemented for all integer vector types. We only use
19758 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19759 // much faster, even faster than using native popcnt instructions.
19761 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19762 MVT VT = V.getSimpleValueType();
19763 SmallVector<SDValue, 32> Shifters(
19764 VT.getVectorNumElements(),
19765 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19766 return DAG.getNode(OpCode, DL, VT, V,
19767 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19769 auto GetMask = [&](SDValue V, APInt Mask) {
19770 MVT VT = V.getSimpleValueType();
19771 SmallVector<SDValue, 32> Masks(
19772 VT.getVectorNumElements(),
19773 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19774 return DAG.getNode(ISD::AND, DL, VT, V,
19775 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19778 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19779 // x86, so set the SRL type to have elements at least i16 wide. This is
19780 // correct because all of our SRLs are followed immediately by a mask anyways
19781 // that handles any bits that sneak into the high bits of the byte elements.
19782 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19786 // v = v - ((v >> 1) & 0x55555555...)
19788 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19789 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19790 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19792 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19793 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19794 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19795 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19796 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19798 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19799 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19800 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19801 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19803 // At this point, V contains the byte-wise population count, and we are
19804 // merely doing a horizontal sum if necessary to get the wider element
19806 if (EltVT == MVT::i8)
19809 return LowerHorizontalByteSum(
19810 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19814 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19815 SelectionDAG &DAG) {
19816 MVT VT = Op.getSimpleValueType();
19817 // FIXME: Need to add AVX-512 support here!
19818 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19819 "Unknown CTPOP type to handle");
19820 SDLoc DL(Op.getNode());
19821 SDValue Op0 = Op.getOperand(0);
19823 if (!Subtarget->hasSSSE3()) {
19824 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19825 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19826 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19829 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19830 unsigned NumElems = VT.getVectorNumElements();
19832 // Extract each 128-bit vector, compute pop count and concat the result.
19833 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19834 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19836 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19837 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19838 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19841 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19844 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19845 SelectionDAG &DAG) {
19846 assert(Op.getSimpleValueType().isVector() &&
19847 "We only do custom lowering for vector population count.");
19848 return LowerVectorCTPOP(Op, Subtarget, DAG);
19851 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19852 SDNode *Node = Op.getNode();
19854 EVT T = Node->getValueType(0);
19855 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19856 DAG.getConstant(0, dl, T), Node->getOperand(2));
19857 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19858 cast<AtomicSDNode>(Node)->getMemoryVT(),
19859 Node->getOperand(0),
19860 Node->getOperand(1), negOp,
19861 cast<AtomicSDNode>(Node)->getMemOperand(),
19862 cast<AtomicSDNode>(Node)->getOrdering(),
19863 cast<AtomicSDNode>(Node)->getSynchScope());
19866 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19867 SDNode *Node = Op.getNode();
19869 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19871 // Convert seq_cst store -> xchg
19872 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19873 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19874 // (The only way to get a 16-byte store is cmpxchg16b)
19875 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19876 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19877 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19878 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19879 cast<AtomicSDNode>(Node)->getMemoryVT(),
19880 Node->getOperand(0),
19881 Node->getOperand(1), Node->getOperand(2),
19882 cast<AtomicSDNode>(Node)->getMemOperand(),
19883 cast<AtomicSDNode>(Node)->getOrdering(),
19884 cast<AtomicSDNode>(Node)->getSynchScope());
19885 return Swap.getValue(1);
19887 // Other atomic stores have a simple pattern.
19891 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19892 MVT VT = Op.getNode()->getSimpleValueType(0);
19894 // Let legalize expand this if it isn't a legal type yet.
19895 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19898 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19901 bool ExtraOp = false;
19902 switch (Op.getOpcode()) {
19903 default: llvm_unreachable("Invalid code");
19904 case ISD::ADDC: Opc = X86ISD::ADD; break;
19905 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19906 case ISD::SUBC: Opc = X86ISD::SUB; break;
19907 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19911 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19913 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19914 Op.getOperand(1), Op.getOperand(2));
19917 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19918 SelectionDAG &DAG) {
19919 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19921 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19922 // which returns the values as { float, float } (in XMM0) or
19923 // { double, double } (which is returned in XMM0, XMM1).
19925 SDValue Arg = Op.getOperand(0);
19926 EVT ArgVT = Arg.getValueType();
19927 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19929 TargetLowering::ArgListTy Args;
19930 TargetLowering::ArgListEntry Entry;
19934 Entry.isSExt = false;
19935 Entry.isZExt = false;
19936 Args.push_back(Entry);
19938 bool isF64 = ArgVT == MVT::f64;
19939 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19940 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19941 // the results are returned via SRet in memory.
19942 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19943 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19945 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19947 Type *RetTy = isF64
19948 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19949 : (Type*)VectorType::get(ArgTy, 4);
19951 TargetLowering::CallLoweringInfo CLI(DAG);
19952 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19953 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19955 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19958 // Returned in xmm0 and xmm1.
19959 return CallResult.first;
19961 // Returned in bits 0:31 and 32:64 xmm0.
19962 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19963 CallResult.first, DAG.getIntPtrConstant(0, dl));
19964 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19965 CallResult.first, DAG.getIntPtrConstant(1, dl));
19966 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19967 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19970 /// Widen a vector input to a vector of NVT. The
19971 /// input vector must have the same element type as NVT.
19972 static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG,
19973 bool FillWithZeroes = false) {
19974 // Check if InOp already has the right width.
19975 MVT InVT = InOp.getSimpleValueType();
19979 if (InOp.isUndef())
19980 return DAG.getUNDEF(NVT);
19982 assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
19983 "input and widen element type must match");
19985 unsigned InNumElts = InVT.getVectorNumElements();
19986 unsigned WidenNumElts = NVT.getVectorNumElements();
19987 assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&
19988 "Unexpected request for vector widening");
19990 EVT EltVT = NVT.getVectorElementType();
19993 if (InOp.getOpcode() == ISD::CONCAT_VECTORS &&
19994 InOp.getNumOperands() == 2) {
19995 SDValue N1 = InOp.getOperand(1);
19996 if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) ||
19998 InOp = InOp.getOperand(0);
19999 InVT = InOp.getSimpleValueType();
20000 InNumElts = InVT.getVectorNumElements();
20003 if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) ||
20004 ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) {
20005 SmallVector<SDValue, 16> Ops;
20006 for (unsigned i = 0; i < InNumElts; ++i)
20007 Ops.push_back(InOp.getOperand(i));
20009 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) :
20010 DAG.getUNDEF(EltVT);
20011 for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i)
20012 Ops.push_back(FillVal);
20013 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
20015 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) :
20017 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal,
20018 InOp, DAG.getIntPtrConstant(0, dl));
20021 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
20022 SelectionDAG &DAG) {
20023 assert(Subtarget->hasAVX512() &&
20024 "MGATHER/MSCATTER are supported on AVX-512 arch only");
20026 // X86 scatter kills mask register, so its type should be added to
20027 // the list of return values.
20028 // If the "scatter" has 2 return values, it is already handled.
20029 if (Op.getNode()->getNumValues() == 2)
20032 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
20033 SDValue Src = N->getValue();
20034 MVT VT = Src.getSimpleValueType();
20035 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
20038 SDValue NewScatter;
20039 SDValue Index = N->getIndex();
20040 SDValue Mask = N->getMask();
20041 SDValue Chain = N->getChain();
20042 SDValue BasePtr = N->getBasePtr();
20043 MVT MemVT = N->getMemoryVT().getSimpleVT();
20044 MVT IndexVT = Index.getSimpleValueType();
20045 MVT MaskVT = Mask.getSimpleValueType();
20047 if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) {
20048 // The v2i32 value was promoted to v2i64.
20049 // Now we "redo" the type legalizer's work and widen the original
20050 // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64
20052 assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&
20053 "Unexpected memory type");
20054 int ShuffleMask[] = {0, 2, -1, -1};
20055 Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src),
20056 DAG.getUNDEF(MVT::v4i32), ShuffleMask);
20057 // Now we have 4 elements instead of 2.
20058 // Expand the index.
20059 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4);
20060 Index = ExtendToType(Index, NewIndexVT, DAG);
20062 // Expand the mask with zeroes
20063 // Mask may be <2 x i64> or <2 x i1> at this moment
20064 assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&
20065 "Unexpected mask type");
20066 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4);
20067 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20071 unsigned NumElts = VT.getVectorNumElements();
20072 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20073 !Index.getSimpleValueType().is512BitVector()) {
20074 // AVX512F supports only 512-bit vectors. Or data or index should
20075 // be 512 bit wide. If now the both index and data are 256-bit, but
20076 // the vector contains 8 elements, we just sign-extend the index
20077 if (IndexVT == MVT::v8i32)
20078 // Just extend index
20079 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20081 // The minimal number of elts in scatter is 8
20084 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20085 // Use original index here, do not modify the index twice
20086 Index = ExtendToType(N->getIndex(), NewIndexVT, DAG);
20087 if (IndexVT.getScalarType() == MVT::i32)
20088 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20091 // At this point we have promoted mask operand
20092 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20093 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20094 // Use the original mask here, do not modify the mask twice
20095 Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true);
20097 // The value that should be stored
20098 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20099 Src = ExtendToType(Src, NewVT, DAG);
20102 // If the mask is "wide" at this point - truncate it to i1 vector
20103 MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts);
20104 Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask);
20106 // The mask is killed by scatter, add it to the values
20107 SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other);
20108 SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
20109 NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops,
20110 N->getMemOperand());
20111 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
20112 return SDValue(NewScatter.getNode(), 0);
20115 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget *Subtarget,
20116 SelectionDAG &DAG) {
20118 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
20119 MVT VT = Op.getSimpleValueType();
20120 SDValue Mask = N->getMask();
20123 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20124 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20125 // This operation is legal for targets with VLX, but without
20126 // VLX the vector should be widened to 512 bit
20127 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20128 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20129 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20130 SDValue Src0 = N->getSrc0();
20131 Src0 = ExtendToType(Src0, WideDataVT, DAG);
20132 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20133 SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(),
20134 N->getBasePtr(), Mask, Src0,
20135 N->getMemoryVT(), N->getMemOperand(),
20136 N->getExtensionType());
20138 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20139 NewLoad.getValue(0),
20140 DAG.getIntPtrConstant(0, dl));
20141 SDValue RetOps[] = {Exract, NewLoad.getValue(1)};
20142 return DAG.getMergeValues(RetOps, dl);
20147 static SDValue LowerMSTORE(SDValue Op, const X86Subtarget *Subtarget,
20148 SelectionDAG &DAG) {
20149 MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode());
20150 SDValue DataToStore = N->getValue();
20151 MVT VT = DataToStore.getSimpleValueType();
20152 SDValue Mask = N->getMask();
20155 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20156 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20157 // This operation is legal for targets with VLX, but without
20158 // VLX the vector should be widened to 512 bit
20159 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20160 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20161 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20162 DataToStore = ExtendToType(DataToStore, WideDataVT, DAG);
20163 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20164 return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(),
20165 Mask, N->getMemoryVT(), N->getMemOperand(),
20166 N->isTruncatingStore());
20171 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
20172 SelectionDAG &DAG) {
20173 assert(Subtarget->hasAVX512() &&
20174 "MGATHER/MSCATTER are supported on AVX-512 arch only");
20176 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
20178 MVT VT = Op.getSimpleValueType();
20179 SDValue Index = N->getIndex();
20180 SDValue Mask = N->getMask();
20181 SDValue Src0 = N->getValue();
20182 MVT IndexVT = Index.getSimpleValueType();
20183 MVT MaskVT = Mask.getSimpleValueType();
20185 unsigned NumElts = VT.getVectorNumElements();
20186 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
20188 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20189 !Index.getSimpleValueType().is512BitVector()) {
20190 // AVX512F supports only 512-bit vectors. Or data or index should
20191 // be 512 bit wide. If now the both index and data are 256-bit, but
20192 // the vector contains 8 elements, we just sign-extend the index
20193 if (NumElts == 8) {
20194 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20195 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
20196 N->getOperand(3), Index };
20197 DAG.UpdateNodeOperands(N, Ops);
20201 // Minimal number of elements in Gather
20204 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20205 Index = ExtendToType(Index, NewIndexVT, DAG);
20206 if (IndexVT.getScalarType() == MVT::i32)
20207 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20210 MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts);
20211 // At this point we have promoted mask operand
20212 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20213 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20214 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20215 Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask);
20217 // The pass-thru value
20218 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20219 Src0 = ExtendToType(Src0, NewVT, DAG);
20221 SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
20222 SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other),
20223 N->getMemoryVT(), dl, Ops,
20224 N->getMemOperand());
20225 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20226 NewGather.getValue(0),
20227 DAG.getIntPtrConstant(0, dl));
20228 SDValue RetOps[] = {Exract, NewGather.getValue(1)};
20229 return DAG.getMergeValues(RetOps, dl);
20234 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
20235 SelectionDAG &DAG) const {
20236 // TODO: Eventually, the lowering of these nodes should be informed by or
20237 // deferred to the GC strategy for the function in which they appear. For
20238 // now, however, they must be lowered to something. Since they are logically
20239 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20240 // require special handling for these nodes), lower them as literal NOOPs for
20242 SmallVector<SDValue, 2> Ops;
20244 Ops.push_back(Op.getOperand(0));
20245 if (Op->getGluedNode())
20246 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20249 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20250 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20255 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
20256 SelectionDAG &DAG) const {
20257 // TODO: Eventually, the lowering of these nodes should be informed by or
20258 // deferred to the GC strategy for the function in which they appear. For
20259 // now, however, they must be lowered to something. Since they are logically
20260 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20261 // require special handling for these nodes), lower them as literal NOOPs for
20263 SmallVector<SDValue, 2> Ops;
20265 Ops.push_back(Op.getOperand(0));
20266 if (Op->getGluedNode())
20267 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20270 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20271 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20276 /// LowerOperation - Provide custom lowering hooks for some operations.
20278 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20279 switch (Op.getOpcode()) {
20280 default: llvm_unreachable("Should not custom lower this!");
20281 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20282 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20283 return LowerCMP_SWAP(Op, Subtarget, DAG);
20284 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20285 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20286 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20287 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20288 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
20289 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
20290 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20291 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20292 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20293 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20294 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20295 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20296 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20297 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20298 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20299 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20300 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20301 case ISD::SHL_PARTS:
20302 case ISD::SRA_PARTS:
20303 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20304 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20305 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20306 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20307 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20308 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20309 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20310 case ISD::SIGN_EXTEND_VECTOR_INREG:
20311 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
20312 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20313 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20314 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20315 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20317 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20318 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20319 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20320 case ISD::SETCC: return LowerSETCC(Op, DAG);
20321 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
20322 case ISD::SELECT: return LowerSELECT(Op, DAG);
20323 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20324 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20325 case ISD::VASTART: return LowerVASTART(Op, DAG);
20326 case ISD::VAARG: return LowerVAARG(Op, DAG);
20327 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20328 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20329 case ISD::INTRINSIC_VOID:
20330 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20331 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20332 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20333 case ISD::FRAME_TO_ARGS_OFFSET:
20334 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20335 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20336 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20337 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20338 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20339 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20340 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20341 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20342 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
20343 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
20345 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
20346 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20347 case ISD::UMUL_LOHI:
20348 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20349 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
20352 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20358 case ISD::UMULO: return LowerXALUO(Op, DAG);
20359 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20360 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20364 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20365 case ISD::ADD: return LowerADD(Op, DAG);
20366 case ISD::SUB: return LowerSUB(Op, DAG);
20370 case ISD::UMIN: return LowerMINMAX(Op, DAG);
20371 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20372 case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
20373 case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);
20374 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
20375 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
20376 case ISD::GC_TRANSITION_START:
20377 return LowerGC_TRANSITION_START(Op, DAG);
20378 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
20382 /// ReplaceNodeResults - Replace a node with an illegal result type
20383 /// with a new node built out of custom code.
20384 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20385 SmallVectorImpl<SDValue>&Results,
20386 SelectionDAG &DAG) const {
20388 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20389 switch (N->getOpcode()) {
20391 llvm_unreachable("Do not know how to custom type legalize this operation!");
20392 case X86ISD::AVG: {
20393 // Legalize types for X86ISD::AVG by expanding vectors.
20394 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20396 auto InVT = N->getValueType(0);
20397 auto InVTSize = InVT.getSizeInBits();
20398 const unsigned RegSize =
20399 (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
20400 assert((!Subtarget->hasAVX512() || RegSize < 512) &&
20401 "512-bit vector requires AVX512");
20402 assert((!Subtarget->hasAVX2() || RegSize < 256) &&
20403 "256-bit vector requires AVX2");
20405 auto ElemVT = InVT.getVectorElementType();
20406 auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
20407 RegSize / ElemVT.getSizeInBits());
20408 assert(RegSize % InVT.getSizeInBits() == 0);
20409 unsigned NumConcat = RegSize / InVT.getSizeInBits();
20411 SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
20412 Ops[0] = N->getOperand(0);
20413 SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20414 Ops[0] = N->getOperand(1);
20415 SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20417 SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
20418 Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
20419 DAG.getIntPtrConstant(0, dl)));
20422 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20423 case X86ISD::FMINC:
20425 case X86ISD::FMAXC:
20426 case X86ISD::FMAX: {
20427 EVT VT = N->getValueType(0);
20428 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
20429 SDValue UNDEF = DAG.getUNDEF(VT);
20430 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20431 N->getOperand(0), UNDEF);
20432 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20433 N->getOperand(1), UNDEF);
20434 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20437 case ISD::SIGN_EXTEND_INREG:
20442 // We don't want to expand or promote these.
20449 case ISD::UDIVREM: {
20450 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20451 Results.push_back(V);
20454 case ISD::FP_TO_SINT:
20455 case ISD::FP_TO_UINT: {
20456 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20458 std::pair<SDValue,SDValue> Vals =
20459 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20460 SDValue FIST = Vals.first, StackSlot = Vals.second;
20461 if (FIST.getNode()) {
20462 EVT VT = N->getValueType(0);
20463 // Return a load from the stack slot.
20464 if (StackSlot.getNode())
20465 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20466 MachinePointerInfo(),
20467 false, false, false, 0));
20469 Results.push_back(FIST);
20473 case ISD::UINT_TO_FP: {
20474 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20475 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20476 N->getValueType(0) != MVT::v2f32)
20478 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20480 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
20482 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20483 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20484 DAG.getBitcast(MVT::v2i64, VBias));
20485 Or = DAG.getBitcast(MVT::v2f64, Or);
20486 // TODO: Are there any fast-math-flags to propagate here?
20487 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20488 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20491 case ISD::FP_ROUND: {
20492 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20494 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20495 Results.push_back(V);
20498 case ISD::FP_EXTEND: {
20499 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
20500 // No other ValueType for FP_EXTEND should reach this point.
20501 assert(N->getValueType(0) == MVT::v2f32 &&
20502 "Do not know how to legalize this Node");
20505 case ISD::INTRINSIC_W_CHAIN: {
20506 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20508 default : llvm_unreachable("Do not know how to custom type "
20509 "legalize this intrinsic operation!");
20510 case Intrinsic::x86_rdtsc:
20511 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20513 case Intrinsic::x86_rdtscp:
20514 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20516 case Intrinsic::x86_rdpmc:
20517 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20520 case ISD::INTRINSIC_WO_CHAIN: {
20521 if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG))
20522 Results.push_back(V);
20525 case ISD::READCYCLECOUNTER: {
20526 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20529 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20530 EVT T = N->getValueType(0);
20531 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20532 bool Regs64bit = T == MVT::i128;
20533 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20534 SDValue cpInL, cpInH;
20535 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20536 DAG.getConstant(0, dl, HalfT));
20537 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20538 DAG.getConstant(1, dl, HalfT));
20539 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20540 Regs64bit ? X86::RAX : X86::EAX,
20542 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20543 Regs64bit ? X86::RDX : X86::EDX,
20544 cpInH, cpInL.getValue(1));
20545 SDValue swapInL, swapInH;
20546 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20547 DAG.getConstant(0, dl, HalfT));
20548 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20549 DAG.getConstant(1, dl, HalfT));
20550 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20551 Regs64bit ? X86::RBX : X86::EBX,
20552 swapInL, cpInH.getValue(1));
20553 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20554 Regs64bit ? X86::RCX : X86::ECX,
20555 swapInH, swapInL.getValue(1));
20556 SDValue Ops[] = { swapInH.getValue(0),
20558 swapInH.getValue(1) };
20559 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20560 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20561 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20562 X86ISD::LCMPXCHG8_DAG;
20563 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20564 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20565 Regs64bit ? X86::RAX : X86::EAX,
20566 HalfT, Result.getValue(1));
20567 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20568 Regs64bit ? X86::RDX : X86::EDX,
20569 HalfT, cpOutL.getValue(2));
20570 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20572 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20573 MVT::i32, cpOutH.getValue(2));
20575 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20576 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
20577 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20579 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20580 Results.push_back(Success);
20581 Results.push_back(EFLAGS.getValue(1));
20584 case ISD::ATOMIC_SWAP:
20585 case ISD::ATOMIC_LOAD_ADD:
20586 case ISD::ATOMIC_LOAD_SUB:
20587 case ISD::ATOMIC_LOAD_AND:
20588 case ISD::ATOMIC_LOAD_OR:
20589 case ISD::ATOMIC_LOAD_XOR:
20590 case ISD::ATOMIC_LOAD_NAND:
20591 case ISD::ATOMIC_LOAD_MIN:
20592 case ISD::ATOMIC_LOAD_MAX:
20593 case ISD::ATOMIC_LOAD_UMIN:
20594 case ISD::ATOMIC_LOAD_UMAX:
20595 case ISD::ATOMIC_LOAD: {
20596 // Delegate to generic TypeLegalization. Situations we can really handle
20597 // should have already been dealt with by AtomicExpandPass.cpp.
20600 case ISD::BITCAST: {
20601 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20602 EVT DstVT = N->getValueType(0);
20603 EVT SrcVT = N->getOperand(0)->getValueType(0);
20605 if (SrcVT != MVT::f64 ||
20606 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20609 unsigned NumElts = DstVT.getVectorNumElements();
20610 EVT SVT = DstVT.getVectorElementType();
20611 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20612 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20613 MVT::v2f64, N->getOperand(0));
20614 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
20616 if (ExperimentalVectorWideningLegalization) {
20617 // If we are legalizing vectors by widening, we already have the desired
20618 // legal vector type, just return it.
20619 Results.push_back(ToVecInt);
20623 SmallVector<SDValue, 8> Elts;
20624 for (unsigned i = 0, e = NumElts; i != e; ++i)
20625 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20626 ToVecInt, DAG.getIntPtrConstant(i, dl)));
20628 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20633 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20634 switch ((X86ISD::NodeType)Opcode) {
20635 case X86ISD::FIRST_NUMBER: break;
20636 case X86ISD::BSF: return "X86ISD::BSF";
20637 case X86ISD::BSR: return "X86ISD::BSR";
20638 case X86ISD::SHLD: return "X86ISD::SHLD";
20639 case X86ISD::SHRD: return "X86ISD::SHRD";
20640 case X86ISD::FAND: return "X86ISD::FAND";
20641 case X86ISD::FANDN: return "X86ISD::FANDN";
20642 case X86ISD::FOR: return "X86ISD::FOR";
20643 case X86ISD::FXOR: return "X86ISD::FXOR";
20644 case X86ISD::FILD: return "X86ISD::FILD";
20645 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20646 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20647 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20648 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20649 case X86ISD::FLD: return "X86ISD::FLD";
20650 case X86ISD::FST: return "X86ISD::FST";
20651 case X86ISD::CALL: return "X86ISD::CALL";
20652 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20653 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20654 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20655 case X86ISD::BT: return "X86ISD::BT";
20656 case X86ISD::CMP: return "X86ISD::CMP";
20657 case X86ISD::COMI: return "X86ISD::COMI";
20658 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20659 case X86ISD::CMPM: return "X86ISD::CMPM";
20660 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20661 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20662 case X86ISD::SETCC: return "X86ISD::SETCC";
20663 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20664 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20665 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20666 case X86ISD::CMOV: return "X86ISD::CMOV";
20667 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20668 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20669 case X86ISD::IRET: return "X86ISD::IRET";
20670 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20671 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20672 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20673 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20674 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20675 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20676 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20677 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20678 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20679 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20680 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20681 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20682 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20683 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20684 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20685 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20686 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20687 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20688 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20689 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20690 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20691 case X86ISD::HADD: return "X86ISD::HADD";
20692 case X86ISD::HSUB: return "X86ISD::HSUB";
20693 case X86ISD::FHADD: return "X86ISD::FHADD";
20694 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20695 case X86ISD::ABS: return "X86ISD::ABS";
20696 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20697 case X86ISD::FMAX: return "X86ISD::FMAX";
20698 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20699 case X86ISD::FMIN: return "X86ISD::FMIN";
20700 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20701 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20702 case X86ISD::FMINC: return "X86ISD::FMINC";
20703 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20704 case X86ISD::FRCP: return "X86ISD::FRCP";
20705 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20706 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20707 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20708 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20709 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20710 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20711 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20712 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20713 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20714 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20715 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20716 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20717 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20718 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20719 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20720 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20721 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20722 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20723 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20724 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20725 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20726 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20727 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20728 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20729 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20730 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20731 case X86ISD::CVT2MASK: return "X86ISD::CVT2MASK";
20732 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20733 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20734 case X86ISD::VSHL: return "X86ISD::VSHL";
20735 case X86ISD::VSRL: return "X86ISD::VSRL";
20736 case X86ISD::VSRA: return "X86ISD::VSRA";
20737 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20738 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20739 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20740 case X86ISD::CMPP: return "X86ISD::CMPP";
20741 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20742 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20743 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20744 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20745 case X86ISD::ADD: return "X86ISD::ADD";
20746 case X86ISD::SUB: return "X86ISD::SUB";
20747 case X86ISD::ADC: return "X86ISD::ADC";
20748 case X86ISD::SBB: return "X86ISD::SBB";
20749 case X86ISD::SMUL: return "X86ISD::SMUL";
20750 case X86ISD::UMUL: return "X86ISD::UMUL";
20751 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20752 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20753 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20754 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20755 case X86ISD::INC: return "X86ISD::INC";
20756 case X86ISD::DEC: return "X86ISD::DEC";
20757 case X86ISD::OR: return "X86ISD::OR";
20758 case X86ISD::XOR: return "X86ISD::XOR";
20759 case X86ISD::AND: return "X86ISD::AND";
20760 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20761 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20762 case X86ISD::PTEST: return "X86ISD::PTEST";
20763 case X86ISD::TESTP: return "X86ISD::TESTP";
20764 case X86ISD::TESTM: return "X86ISD::TESTM";
20765 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20766 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20767 case X86ISD::KTEST: return "X86ISD::KTEST";
20768 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20769 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20770 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20771 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20772 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20773 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20774 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20775 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20776 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20777 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20778 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20779 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20780 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20781 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20782 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20783 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20784 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20785 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20786 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20787 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20788 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20789 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20790 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20791 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20792 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20793 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20794 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20795 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20796 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20797 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20798 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20799 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20800 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20801 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20802 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20803 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20804 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20805 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20806 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20807 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20808 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20809 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20810 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20811 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20812 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20813 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20814 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20815 case X86ISD::SAHF: return "X86ISD::SAHF";
20816 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20817 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20818 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20819 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20820 case X86ISD::VPROT: return "X86ISD::VPROT";
20821 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20822 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20823 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20824 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20825 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20826 case X86ISD::FMADD: return "X86ISD::FMADD";
20827 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20828 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20829 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20830 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20831 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20832 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20833 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20834 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20835 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20836 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20837 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20838 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20839 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20840 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20841 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20842 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20843 case X86ISD::XTEST: return "X86ISD::XTEST";
20844 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20845 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20846 case X86ISD::SELECT: return "X86ISD::SELECT";
20847 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20848 case X86ISD::RCP28: return "X86ISD::RCP28";
20849 case X86ISD::EXP2: return "X86ISD::EXP2";
20850 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20851 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20852 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20853 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20854 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20855 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20856 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20857 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20858 case X86ISD::ADDS: return "X86ISD::ADDS";
20859 case X86ISD::SUBS: return "X86ISD::SUBS";
20860 case X86ISD::AVG: return "X86ISD::AVG";
20861 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20862 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20863 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20864 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20865 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20866 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20867 case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
20872 // isLegalAddressingMode - Return true if the addressing mode represented
20873 // by AM is legal for this target, for a load/store of the specified type.
20874 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20875 const AddrMode &AM, Type *Ty,
20876 unsigned AS) const {
20877 // X86 supports extremely general addressing modes.
20878 CodeModel::Model M = getTargetMachine().getCodeModel();
20879 Reloc::Model R = getTargetMachine().getRelocationModel();
20881 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20882 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20887 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20889 // If a reference to this global requires an extra load, we can't fold it.
20890 if (isGlobalStubReference(GVFlags))
20893 // If BaseGV requires a register for the PIC base, we cannot also have a
20894 // BaseReg specified.
20895 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20898 // If lower 4G is not available, then we must use rip-relative addressing.
20899 if ((M != CodeModel::Small || R != Reloc::Static) &&
20900 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20904 switch (AM.Scale) {
20910 // These scales always work.
20915 // These scales are formed with basereg+scalereg. Only accept if there is
20920 default: // Other stuff never works.
20927 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20928 unsigned Bits = Ty->getScalarSizeInBits();
20930 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20931 // particularly cheaper than those without.
20935 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20936 // variable shifts just as cheap as scalar ones.
20937 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20940 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20941 // fully general vector.
20945 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20946 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20948 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20949 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20950 return NumBits1 > NumBits2;
20953 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20954 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20957 if (!isTypeLegal(EVT::getEVT(Ty1)))
20960 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20962 // Assuming the caller doesn't have a zeroext or signext return parameter,
20963 // truncation all the way down to i1 is valid.
20967 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20968 return isInt<32>(Imm);
20971 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20972 // Can also use sub to handle negated immediates.
20973 return isInt<32>(Imm);
20976 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20977 if (!VT1.isInteger() || !VT2.isInteger())
20979 unsigned NumBits1 = VT1.getSizeInBits();
20980 unsigned NumBits2 = VT2.getSizeInBits();
20981 return NumBits1 > NumBits2;
20984 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20985 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20986 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20989 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20990 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20991 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20994 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20995 EVT VT1 = Val.getValueType();
20996 if (isZExtFree(VT1, VT2))
20999 if (Val.getOpcode() != ISD::LOAD)
21002 if (!VT1.isSimple() || !VT1.isInteger() ||
21003 !VT2.isSimple() || !VT2.isInteger())
21006 switch (VT1.getSimpleVT().SimpleTy) {
21011 // X86 has 8, 16, and 32-bit zero-extending loads.
21018 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
21021 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
21022 if (!Subtarget->hasAnyFMA())
21025 VT = VT.getScalarType();
21027 if (!VT.isSimple())
21030 switch (VT.getSimpleVT().SimpleTy) {
21041 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
21042 // i16 instructions are longer (0x66 prefix) and potentially slower.
21043 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
21046 /// isShuffleMaskLegal - Targets can use this to indicate that they only
21047 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
21048 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
21049 /// are assumed to be legal.
21051 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
21053 if (!VT.isSimple())
21056 // Not for i1 vectors
21057 if (VT.getSimpleVT().getScalarType() == MVT::i1)
21060 // Very little shuffling can be done for 64-bit vectors right now.
21061 if (VT.getSimpleVT().getSizeInBits() == 64)
21064 // We only care that the types being shuffled are legal. The lowering can
21065 // handle any possible shuffle mask that results.
21066 return isTypeLegal(VT.getSimpleVT());
21070 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
21072 // Just delegate to the generic legality, clear masks aren't special.
21073 return isShuffleMaskLegal(Mask, VT);
21076 //===----------------------------------------------------------------------===//
21077 // X86 Scheduler Hooks
21078 //===----------------------------------------------------------------------===//
21080 /// Utility function to emit xbegin specifying the start of an RTM region.
21081 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
21082 const TargetInstrInfo *TII) {
21083 DebugLoc DL = MI->getDebugLoc();
21085 const BasicBlock *BB = MBB->getBasicBlock();
21086 MachineFunction::iterator I = ++MBB->getIterator();
21088 // For the v = xbegin(), we generate
21099 MachineBasicBlock *thisMBB = MBB;
21100 MachineFunction *MF = MBB->getParent();
21101 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21102 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21103 MF->insert(I, mainMBB);
21104 MF->insert(I, sinkMBB);
21106 // Transfer the remainder of BB and its successor edges to sinkMBB.
21107 sinkMBB->splice(sinkMBB->begin(), MBB,
21108 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21109 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21113 // # fallthrough to mainMBB
21114 // # abortion to sinkMBB
21115 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
21116 thisMBB->addSuccessor(mainMBB);
21117 thisMBB->addSuccessor(sinkMBB);
21121 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
21122 mainMBB->addSuccessor(sinkMBB);
21125 // EAX is live into the sinkMBB
21126 sinkMBB->addLiveIn(X86::EAX);
21127 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21128 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21131 MI->eraseFromParent();
21135 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
21136 // or XMM0_V32I8 in AVX all of this code can be replaced with that
21137 // in the .td file.
21138 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
21139 const TargetInstrInfo *TII) {
21141 switch (MI->getOpcode()) {
21142 default: llvm_unreachable("illegal opcode!");
21143 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
21144 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
21145 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
21146 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
21147 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
21148 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
21149 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
21150 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
21153 DebugLoc dl = MI->getDebugLoc();
21154 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21156 unsigned NumArgs = MI->getNumOperands();
21157 for (unsigned i = 1; i < NumArgs; ++i) {
21158 MachineOperand &Op = MI->getOperand(i);
21159 if (!(Op.isReg() && Op.isImplicit()))
21160 MIB.addOperand(Op);
21162 if (MI->hasOneMemOperand())
21163 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21165 BuildMI(*BB, MI, dl,
21166 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21167 .addReg(X86::XMM0);
21169 MI->eraseFromParent();
21173 // FIXME: Custom handling because TableGen doesn't support multiple implicit
21174 // defs in an instruction pattern
21175 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
21176 const TargetInstrInfo *TII) {
21178 switch (MI->getOpcode()) {
21179 default: llvm_unreachable("illegal opcode!");
21180 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
21181 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
21182 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
21183 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
21184 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
21185 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
21186 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
21187 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
21190 DebugLoc dl = MI->getDebugLoc();
21191 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21193 unsigned NumArgs = MI->getNumOperands(); // remove the results
21194 for (unsigned i = 1; i < NumArgs; ++i) {
21195 MachineOperand &Op = MI->getOperand(i);
21196 if (!(Op.isReg() && Op.isImplicit()))
21197 MIB.addOperand(Op);
21199 if (MI->hasOneMemOperand())
21200 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21202 BuildMI(*BB, MI, dl,
21203 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21206 MI->eraseFromParent();
21210 static MachineBasicBlock *EmitWRPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21211 const X86Subtarget *Subtarget) {
21212 DebugLoc dl = MI->getDebugLoc();
21213 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21215 // insert input VAL into EAX
21216 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
21217 .addReg(MI->getOperand(0).getReg());
21218 // insert zero to ECX
21219 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21222 // insert zero to EDX
21223 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::EDX)
21226 // insert WRPKRU instruction
21227 BuildMI(*BB, MI, dl, TII->get(X86::WRPKRUr));
21229 MI->eraseFromParent(); // The pseudo is gone now.
21233 static MachineBasicBlock *EmitRDPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21234 const X86Subtarget *Subtarget) {
21235 DebugLoc dl = MI->getDebugLoc();
21236 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21238 // insert zero to ECX
21239 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21242 // insert RDPKRU instruction
21243 BuildMI(*BB, MI, dl, TII->get(X86::RDPKRUr));
21244 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21247 MI->eraseFromParent(); // The pseudo is gone now.
21251 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
21252 const X86Subtarget *Subtarget) {
21253 DebugLoc dl = MI->getDebugLoc();
21254 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21255 // Address into RAX/EAX, other two args into ECX, EDX.
21256 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
21257 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
21258 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
21259 for (int i = 0; i < X86::AddrNumOperands; ++i)
21260 MIB.addOperand(MI->getOperand(i));
21262 unsigned ValOps = X86::AddrNumOperands;
21263 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
21264 .addReg(MI->getOperand(ValOps).getReg());
21265 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
21266 .addReg(MI->getOperand(ValOps+1).getReg());
21268 // The instruction doesn't actually take any operands though.
21269 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
21271 MI->eraseFromParent(); // The pseudo is gone now.
21275 MachineBasicBlock *
21276 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
21277 MachineBasicBlock *MBB) const {
21278 // Emit va_arg instruction on X86-64.
21280 // Operands to this pseudo-instruction:
21281 // 0 ) Output : destination address (reg)
21282 // 1-5) Input : va_list address (addr, i64mem)
21283 // 6 ) ArgSize : Size (in bytes) of vararg type
21284 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
21285 // 8 ) Align : Alignment of type
21286 // 9 ) EFLAGS (implicit-def)
21288 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
21289 static_assert(X86::AddrNumOperands == 5,
21290 "VAARG_64 assumes 5 address operands");
21292 unsigned DestReg = MI->getOperand(0).getReg();
21293 MachineOperand &Base = MI->getOperand(1);
21294 MachineOperand &Scale = MI->getOperand(2);
21295 MachineOperand &Index = MI->getOperand(3);
21296 MachineOperand &Disp = MI->getOperand(4);
21297 MachineOperand &Segment = MI->getOperand(5);
21298 unsigned ArgSize = MI->getOperand(6).getImm();
21299 unsigned ArgMode = MI->getOperand(7).getImm();
21300 unsigned Align = MI->getOperand(8).getImm();
21302 // Memory Reference
21303 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
21304 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21305 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21307 // Machine Information
21308 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21309 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
21310 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
21311 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
21312 DebugLoc DL = MI->getDebugLoc();
21314 // struct va_list {
21317 // i64 overflow_area (address)
21318 // i64 reg_save_area (address)
21320 // sizeof(va_list) = 24
21321 // alignment(va_list) = 8
21323 unsigned TotalNumIntRegs = 6;
21324 unsigned TotalNumXMMRegs = 8;
21325 bool UseGPOffset = (ArgMode == 1);
21326 bool UseFPOffset = (ArgMode == 2);
21327 unsigned MaxOffset = TotalNumIntRegs * 8 +
21328 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21330 /* Align ArgSize to a multiple of 8 */
21331 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21332 bool NeedsAlign = (Align > 8);
21334 MachineBasicBlock *thisMBB = MBB;
21335 MachineBasicBlock *overflowMBB;
21336 MachineBasicBlock *offsetMBB;
21337 MachineBasicBlock *endMBB;
21339 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21340 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21341 unsigned OffsetReg = 0;
21343 if (!UseGPOffset && !UseFPOffset) {
21344 // If we only pull from the overflow region, we don't create a branch.
21345 // We don't need to alter control flow.
21346 OffsetDestReg = 0; // unused
21347 OverflowDestReg = DestReg;
21349 offsetMBB = nullptr;
21350 overflowMBB = thisMBB;
21353 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21354 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21355 // If not, pull from overflow_area. (branch to overflowMBB)
21360 // offsetMBB overflowMBB
21365 // Registers for the PHI in endMBB
21366 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21367 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21369 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21370 MachineFunction *MF = MBB->getParent();
21371 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21372 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21373 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21375 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21377 // Insert the new basic blocks
21378 MF->insert(MBBIter, offsetMBB);
21379 MF->insert(MBBIter, overflowMBB);
21380 MF->insert(MBBIter, endMBB);
21382 // Transfer the remainder of MBB and its successor edges to endMBB.
21383 endMBB->splice(endMBB->begin(), thisMBB,
21384 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21385 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21387 // Make offsetMBB and overflowMBB successors of thisMBB
21388 thisMBB->addSuccessor(offsetMBB);
21389 thisMBB->addSuccessor(overflowMBB);
21391 // endMBB is a successor of both offsetMBB and overflowMBB
21392 offsetMBB->addSuccessor(endMBB);
21393 overflowMBB->addSuccessor(endMBB);
21395 // Load the offset value into a register
21396 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21397 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21401 .addDisp(Disp, UseFPOffset ? 4 : 0)
21402 .addOperand(Segment)
21403 .setMemRefs(MMOBegin, MMOEnd);
21405 // Check if there is enough room left to pull this argument.
21406 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21408 .addImm(MaxOffset + 8 - ArgSizeA8);
21410 // Branch to "overflowMBB" if offset >= max
21411 // Fall through to "offsetMBB" otherwise
21412 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21413 .addMBB(overflowMBB);
21416 // In offsetMBB, emit code to use the reg_save_area.
21418 assert(OffsetReg != 0);
21420 // Read the reg_save_area address.
21421 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21422 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21427 .addOperand(Segment)
21428 .setMemRefs(MMOBegin, MMOEnd);
21430 // Zero-extend the offset
21431 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21432 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21435 .addImm(X86::sub_32bit);
21437 // Add the offset to the reg_save_area to get the final address.
21438 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21439 .addReg(OffsetReg64)
21440 .addReg(RegSaveReg);
21442 // Compute the offset for the next argument
21443 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21444 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21446 .addImm(UseFPOffset ? 16 : 8);
21448 // Store it back into the va_list.
21449 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21453 .addDisp(Disp, UseFPOffset ? 4 : 0)
21454 .addOperand(Segment)
21455 .addReg(NextOffsetReg)
21456 .setMemRefs(MMOBegin, MMOEnd);
21459 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21464 // Emit code to use overflow area
21467 // Load the overflow_area address into a register.
21468 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21469 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21474 .addOperand(Segment)
21475 .setMemRefs(MMOBegin, MMOEnd);
21477 // If we need to align it, do so. Otherwise, just copy the address
21478 // to OverflowDestReg.
21480 // Align the overflow address
21481 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21482 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21484 // aligned_addr = (addr + (align-1)) & ~(align-1)
21485 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21486 .addReg(OverflowAddrReg)
21489 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21491 .addImm(~(uint64_t)(Align-1));
21493 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21494 .addReg(OverflowAddrReg);
21497 // Compute the next overflow address after this argument.
21498 // (the overflow address should be kept 8-byte aligned)
21499 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21500 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21501 .addReg(OverflowDestReg)
21502 .addImm(ArgSizeA8);
21504 // Store the new overflow address.
21505 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21510 .addOperand(Segment)
21511 .addReg(NextAddrReg)
21512 .setMemRefs(MMOBegin, MMOEnd);
21514 // If we branched, emit the PHI to the front of endMBB.
21516 BuildMI(*endMBB, endMBB->begin(), DL,
21517 TII->get(X86::PHI), DestReg)
21518 .addReg(OffsetDestReg).addMBB(offsetMBB)
21519 .addReg(OverflowDestReg).addMBB(overflowMBB);
21522 // Erase the pseudo instruction
21523 MI->eraseFromParent();
21528 MachineBasicBlock *
21529 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21531 MachineBasicBlock *MBB) const {
21532 // Emit code to save XMM registers to the stack. The ABI says that the
21533 // number of registers to save is given in %al, so it's theoretically
21534 // possible to do an indirect jump trick to avoid saving all of them,
21535 // however this code takes a simpler approach and just executes all
21536 // of the stores if %al is non-zero. It's less code, and it's probably
21537 // easier on the hardware branch predictor, and stores aren't all that
21538 // expensive anyway.
21540 // Create the new basic blocks. One block contains all the XMM stores,
21541 // and one block is the final destination regardless of whether any
21542 // stores were performed.
21543 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21544 MachineFunction *F = MBB->getParent();
21545 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21546 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21547 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21548 F->insert(MBBIter, XMMSaveMBB);
21549 F->insert(MBBIter, EndMBB);
21551 // Transfer the remainder of MBB and its successor edges to EndMBB.
21552 EndMBB->splice(EndMBB->begin(), MBB,
21553 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21554 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21556 // The original block will now fall through to the XMM save block.
21557 MBB->addSuccessor(XMMSaveMBB);
21558 // The XMMSaveMBB will fall through to the end block.
21559 XMMSaveMBB->addSuccessor(EndMBB);
21561 // Now add the instructions.
21562 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21563 DebugLoc DL = MI->getDebugLoc();
21565 unsigned CountReg = MI->getOperand(0).getReg();
21566 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21567 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21569 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
21570 // If %al is 0, branch around the XMM save block.
21571 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21572 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21573 MBB->addSuccessor(EndMBB);
21576 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21577 // that was just emitted, but clearly shouldn't be "saved".
21578 assert((MI->getNumOperands() <= 3 ||
21579 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21580 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21581 && "Expected last argument to be EFLAGS");
21582 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21583 // In the XMM save block, save all the XMM argument registers.
21584 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21585 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21586 MachineMemOperand *MMO = F->getMachineMemOperand(
21587 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
21588 MachineMemOperand::MOStore,
21589 /*Size=*/16, /*Align=*/16);
21590 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21591 .addFrameIndex(RegSaveFrameIndex)
21592 .addImm(/*Scale=*/1)
21593 .addReg(/*IndexReg=*/0)
21594 .addImm(/*Disp=*/Offset)
21595 .addReg(/*Segment=*/0)
21596 .addReg(MI->getOperand(i).getReg())
21597 .addMemOperand(MMO);
21600 MI->eraseFromParent(); // The pseudo instruction is gone now.
21605 // The EFLAGS operand of SelectItr might be missing a kill marker
21606 // because there were multiple uses of EFLAGS, and ISel didn't know
21607 // which to mark. Figure out whether SelectItr should have had a
21608 // kill marker, and set it if it should. Returns the correct kill
21610 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21611 MachineBasicBlock* BB,
21612 const TargetRegisterInfo* TRI) {
21613 // Scan forward through BB for a use/def of EFLAGS.
21614 MachineBasicBlock::iterator miI(std::next(SelectItr));
21615 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21616 const MachineInstr& mi = *miI;
21617 if (mi.readsRegister(X86::EFLAGS))
21619 if (mi.definesRegister(X86::EFLAGS))
21620 break; // Should have kill-flag - update below.
21623 // If we hit the end of the block, check whether EFLAGS is live into a
21625 if (miI == BB->end()) {
21626 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21627 sEnd = BB->succ_end();
21628 sItr != sEnd; ++sItr) {
21629 MachineBasicBlock* succ = *sItr;
21630 if (succ->isLiveIn(X86::EFLAGS))
21635 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21636 // out. SelectMI should have a kill flag on EFLAGS.
21637 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21641 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
21642 // together with other CMOV pseudo-opcodes into a single basic-block with
21643 // conditional jump around it.
21644 static bool isCMOVPseudo(MachineInstr *MI) {
21645 switch (MI->getOpcode()) {
21646 case X86::CMOV_FR32:
21647 case X86::CMOV_FR64:
21648 case X86::CMOV_GR8:
21649 case X86::CMOV_GR16:
21650 case X86::CMOV_GR32:
21651 case X86::CMOV_RFP32:
21652 case X86::CMOV_RFP64:
21653 case X86::CMOV_RFP80:
21654 case X86::CMOV_V2F64:
21655 case X86::CMOV_V2I64:
21656 case X86::CMOV_V4F32:
21657 case X86::CMOV_V4F64:
21658 case X86::CMOV_V4I64:
21659 case X86::CMOV_V16F32:
21660 case X86::CMOV_V8F32:
21661 case X86::CMOV_V8F64:
21662 case X86::CMOV_V8I64:
21663 case X86::CMOV_V8I1:
21664 case X86::CMOV_V16I1:
21665 case X86::CMOV_V32I1:
21666 case X86::CMOV_V64I1:
21674 MachineBasicBlock *
21675 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21676 MachineBasicBlock *BB) const {
21677 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21678 DebugLoc DL = MI->getDebugLoc();
21680 // To "insert" a SELECT_CC instruction, we actually have to insert the
21681 // diamond control-flow pattern. The incoming instruction knows the
21682 // destination vreg to set, the condition code register to branch on, the
21683 // true/false values to select between, and a branch opcode to use.
21684 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21685 MachineFunction::iterator It = ++BB->getIterator();
21690 // cmpTY ccX, r1, r2
21692 // fallthrough --> copy0MBB
21693 MachineBasicBlock *thisMBB = BB;
21694 MachineFunction *F = BB->getParent();
21696 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21697 // as described above, by inserting a BB, and then making a PHI at the join
21698 // point to select the true and false operands of the CMOV in the PHI.
21700 // The code also handles two different cases of multiple CMOV opcodes
21704 // In this case, there are multiple CMOVs in a row, all which are based on
21705 // the same condition setting (or the exact opposite condition setting).
21706 // In this case we can lower all the CMOVs using a single inserted BB, and
21707 // then make a number of PHIs at the join point to model the CMOVs. The only
21708 // trickiness here, is that in a case like:
21710 // t2 = CMOV cond1 t1, f1
21711 // t3 = CMOV cond1 t2, f2
21713 // when rewriting this into PHIs, we have to perform some renaming on the
21714 // temps since you cannot have a PHI operand refer to a PHI result earlier
21715 // in the same block. The "simple" but wrong lowering would be:
21717 // t2 = PHI t1(BB1), f1(BB2)
21718 // t3 = PHI t2(BB1), f2(BB2)
21720 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21721 // renaming is to note that on the path through BB1, t2 is really just a
21722 // copy of t1, and do that renaming, properly generating:
21724 // t2 = PHI t1(BB1), f1(BB2)
21725 // t3 = PHI t1(BB1), f2(BB2)
21727 // Case 2, we lower cascaded CMOVs such as
21729 // (CMOV (CMOV F, T, cc1), T, cc2)
21731 // to two successives branches. For that, we look for another CMOV as the
21732 // following instruction.
21734 // Without this, we would add a PHI between the two jumps, which ends up
21735 // creating a few copies all around. For instance, for
21737 // (sitofp (zext (fcmp une)))
21739 // we would generate:
21741 // ucomiss %xmm1, %xmm0
21742 // movss <1.0f>, %xmm0
21743 // movaps %xmm0, %xmm1
21745 // xorps %xmm1, %xmm1
21748 // movaps %xmm1, %xmm0
21752 // because this custom-inserter would have generated:
21764 // A: X = ...; Y = ...
21766 // C: Z = PHI [X, A], [Y, B]
21768 // E: PHI [X, C], [Z, D]
21770 // If we lower both CMOVs in a single step, we can instead generate:
21782 // A: X = ...; Y = ...
21784 // E: PHI [X, A], [X, C], [Y, D]
21786 // Which, in our sitofp/fcmp example, gives us something like:
21788 // ucomiss %xmm1, %xmm0
21789 // movss <1.0f>, %xmm0
21792 // xorps %xmm0, %xmm0
21796 MachineInstr *CascadedCMOV = nullptr;
21797 MachineInstr *LastCMOV = MI;
21798 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21799 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21800 MachineBasicBlock::iterator NextMIIt =
21801 std::next(MachineBasicBlock::iterator(MI));
21803 // Check for case 1, where there are multiple CMOVs with the same condition
21804 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21805 // number of jumps the most.
21807 if (isCMOVPseudo(MI)) {
21808 // See if we have a string of CMOVS with the same condition.
21809 while (NextMIIt != BB->end() &&
21810 isCMOVPseudo(NextMIIt) &&
21811 (NextMIIt->getOperand(3).getImm() == CC ||
21812 NextMIIt->getOperand(3).getImm() == OppCC)) {
21813 LastCMOV = &*NextMIIt;
21818 // This checks for case 2, but only do this if we didn't already find
21819 // case 1, as indicated by LastCMOV == MI.
21820 if (LastCMOV == MI &&
21821 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21822 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21823 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21824 CascadedCMOV = &*NextMIIt;
21827 MachineBasicBlock *jcc1MBB = nullptr;
21829 // If we have a cascaded CMOV, we lower it to two successive branches to
21830 // the same block. EFLAGS is used by both, so mark it as live in the second.
21831 if (CascadedCMOV) {
21832 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21833 F->insert(It, jcc1MBB);
21834 jcc1MBB->addLiveIn(X86::EFLAGS);
21837 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21838 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21839 F->insert(It, copy0MBB);
21840 F->insert(It, sinkMBB);
21842 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21843 // live into the sink and copy blocks.
21844 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21846 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21847 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21848 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21849 copy0MBB->addLiveIn(X86::EFLAGS);
21850 sinkMBB->addLiveIn(X86::EFLAGS);
21853 // Transfer the remainder of BB and its successor edges to sinkMBB.
21854 sinkMBB->splice(sinkMBB->begin(), BB,
21855 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21856 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21858 // Add the true and fallthrough blocks as its successors.
21859 if (CascadedCMOV) {
21860 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21861 BB->addSuccessor(jcc1MBB);
21863 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21864 // jump to the sinkMBB.
21865 jcc1MBB->addSuccessor(copy0MBB);
21866 jcc1MBB->addSuccessor(sinkMBB);
21868 BB->addSuccessor(copy0MBB);
21871 // The true block target of the first (or only) branch is always sinkMBB.
21872 BB->addSuccessor(sinkMBB);
21874 // Create the conditional branch instruction.
21875 unsigned Opc = X86::GetCondBranchFromCond(CC);
21876 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21878 if (CascadedCMOV) {
21879 unsigned Opc2 = X86::GetCondBranchFromCond(
21880 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21881 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21885 // %FalseValue = ...
21886 // # fallthrough to sinkMBB
21887 copy0MBB->addSuccessor(sinkMBB);
21890 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21892 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21893 MachineBasicBlock::iterator MIItEnd =
21894 std::next(MachineBasicBlock::iterator(LastCMOV));
21895 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21896 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21897 MachineInstrBuilder MIB;
21899 // As we are creating the PHIs, we have to be careful if there is more than
21900 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21901 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21902 // That also means that PHI construction must work forward from earlier to
21903 // later, and that the code must maintain a mapping from earlier PHI's
21904 // destination registers, and the registers that went into the PHI.
21906 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21907 unsigned DestReg = MIIt->getOperand(0).getReg();
21908 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21909 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21911 // If this CMOV we are generating is the opposite condition from
21912 // the jump we generated, then we have to swap the operands for the
21913 // PHI that is going to be generated.
21914 if (MIIt->getOperand(3).getImm() == OppCC)
21915 std::swap(Op1Reg, Op2Reg);
21917 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21918 Op1Reg = RegRewriteTable[Op1Reg].first;
21920 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21921 Op2Reg = RegRewriteTable[Op2Reg].second;
21923 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21924 TII->get(X86::PHI), DestReg)
21925 .addReg(Op1Reg).addMBB(copy0MBB)
21926 .addReg(Op2Reg).addMBB(thisMBB);
21928 // Add this PHI to the rewrite table.
21929 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21932 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21933 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21934 if (CascadedCMOV) {
21935 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21936 // Copy the PHI result to the register defined by the second CMOV.
21937 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21938 DL, TII->get(TargetOpcode::COPY),
21939 CascadedCMOV->getOperand(0).getReg())
21940 .addReg(MI->getOperand(0).getReg());
21941 CascadedCMOV->eraseFromParent();
21944 // Now remove the CMOV(s).
21945 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21946 (MIIt++)->eraseFromParent();
21951 MachineBasicBlock *
21952 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21953 MachineBasicBlock *BB) const {
21954 // Combine the following atomic floating-point modification pattern:
21955 // a.store(reg OP a.load(acquire), release)
21956 // Transform them into:
21957 // OPss (%gpr), %xmm
21958 // movss %xmm, (%gpr)
21959 // Or sd equivalent for 64-bit operations.
21961 switch (MI->getOpcode()) {
21962 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21963 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21964 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21966 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21967 DebugLoc DL = MI->getDebugLoc();
21968 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21969 MachineOperand MSrc = MI->getOperand(0);
21970 unsigned VSrc = MI->getOperand(5).getReg();
21971 const MachineOperand &Disp = MI->getOperand(3);
21972 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21973 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21974 if (hasDisp && MSrc.isReg())
21975 MSrc.setIsKill(false);
21976 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21977 .addOperand(/*Base=*/MSrc)
21978 .addImm(/*Scale=*/1)
21979 .addReg(/*Index=*/0)
21980 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21982 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21983 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21985 .addOperand(/*Base=*/MSrc)
21986 .addImm(/*Scale=*/1)
21987 .addReg(/*Index=*/0)
21988 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21989 .addReg(/*Segment=*/0);
21990 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21991 MI->eraseFromParent(); // The pseudo instruction is gone now.
21995 MachineBasicBlock *
21996 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21997 MachineBasicBlock *BB) const {
21998 MachineFunction *MF = BB->getParent();
21999 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22000 DebugLoc DL = MI->getDebugLoc();
22001 const BasicBlock *LLVM_BB = BB->getBasicBlock();
22003 assert(MF->shouldSplitStack());
22005 const bool Is64Bit = Subtarget->is64Bit();
22006 const bool IsLP64 = Subtarget->isTarget64BitLP64();
22008 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
22009 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
22012 // ... [Till the alloca]
22013 // If stacklet is not large enough, jump to mallocMBB
22016 // Allocate by subtracting from RSP
22017 // Jump to continueMBB
22020 // Allocate by call to runtime
22024 // [rest of original BB]
22027 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
22028 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
22029 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
22031 MachineRegisterInfo &MRI = MF->getRegInfo();
22032 const TargetRegisterClass *AddrRegClass =
22033 getRegClassFor(getPointerTy(MF->getDataLayout()));
22035 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
22036 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
22037 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
22038 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
22039 sizeVReg = MI->getOperand(1).getReg(),
22040 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
22042 MachineFunction::iterator MBBIter = ++BB->getIterator();
22044 MF->insert(MBBIter, bumpMBB);
22045 MF->insert(MBBIter, mallocMBB);
22046 MF->insert(MBBIter, continueMBB);
22048 continueMBB->splice(continueMBB->begin(), BB,
22049 std::next(MachineBasicBlock::iterator(MI)), BB->end());
22050 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
22052 // Add code to the main basic block to check if the stack limit has been hit,
22053 // and if so, jump to mallocMBB otherwise to bumpMBB.
22054 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
22055 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
22056 .addReg(tmpSPVReg).addReg(sizeVReg);
22057 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
22058 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
22059 .addReg(SPLimitVReg);
22060 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
22062 // bumpMBB simply decreases the stack pointer, since we know the current
22063 // stacklet has enough space.
22064 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
22065 .addReg(SPLimitVReg);
22066 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
22067 .addReg(SPLimitVReg);
22068 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22070 // Calls into a routine in libgcc to allocate more space from the heap.
22071 const uint32_t *RegMask =
22072 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
22074 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
22076 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22077 .addExternalSymbol("__morestack_allocate_stack_space")
22078 .addRegMask(RegMask)
22079 .addReg(X86::RDI, RegState::Implicit)
22080 .addReg(X86::RAX, RegState::ImplicitDefine);
22081 } else if (Is64Bit) {
22082 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
22084 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22085 .addExternalSymbol("__morestack_allocate_stack_space")
22086 .addRegMask(RegMask)
22087 .addReg(X86::EDI, RegState::Implicit)
22088 .addReg(X86::EAX, RegState::ImplicitDefine);
22090 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
22092 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
22093 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
22094 .addExternalSymbol("__morestack_allocate_stack_space")
22095 .addRegMask(RegMask)
22096 .addReg(X86::EAX, RegState::ImplicitDefine);
22100 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
22103 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
22104 .addReg(IsLP64 ? X86::RAX : X86::EAX);
22105 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22107 // Set up the CFG correctly.
22108 BB->addSuccessor(bumpMBB);
22109 BB->addSuccessor(mallocMBB);
22110 mallocMBB->addSuccessor(continueMBB);
22111 bumpMBB->addSuccessor(continueMBB);
22113 // Take care of the PHI nodes.
22114 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
22115 MI->getOperand(0).getReg())
22116 .addReg(mallocPtrVReg).addMBB(mallocMBB)
22117 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
22119 // Delete the original pseudo instruction.
22120 MI->eraseFromParent();
22123 return continueMBB;
22126 MachineBasicBlock *
22127 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
22128 MachineBasicBlock *BB) const {
22129 assert(!Subtarget->isTargetMachO());
22130 DebugLoc DL = MI->getDebugLoc();
22131 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
22132 *BB->getParent(), *BB, MI, DL, false);
22133 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
22134 MI->eraseFromParent(); // The pseudo instruction is gone now.
22138 MachineBasicBlock *
22139 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
22140 MachineBasicBlock *BB) const {
22141 MachineFunction *MF = BB->getParent();
22142 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22143 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
22144 DebugLoc DL = MI->getDebugLoc();
22146 assert(!isAsynchronousEHPersonality(
22147 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
22148 "SEH does not use catchret!");
22150 // Only 32-bit EH needs to worry about manually restoring stack pointers.
22151 if (!Subtarget->is32Bit())
22154 // C++ EH creates a new target block to hold the restore code, and wires up
22155 // the new block to the return destination with a normal JMP_4.
22156 MachineBasicBlock *RestoreMBB =
22157 MF->CreateMachineBasicBlock(BB->getBasicBlock());
22158 assert(BB->succ_size() == 1);
22159 MF->insert(std::next(BB->getIterator()), RestoreMBB);
22160 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
22161 BB->addSuccessor(RestoreMBB);
22162 MI->getOperand(0).setMBB(RestoreMBB);
22164 auto RestoreMBBI = RestoreMBB->begin();
22165 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
22166 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
22170 MachineBasicBlock *
22171 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
22172 MachineBasicBlock *BB) const {
22173 MachineFunction *MF = BB->getParent();
22174 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
22175 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
22176 // Only 32-bit SEH requires special handling for catchpad.
22177 if (IsSEH && Subtarget->is32Bit()) {
22178 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22179 DebugLoc DL = MI->getDebugLoc();
22180 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
22182 MI->eraseFromParent();
22186 MachineBasicBlock *
22187 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
22188 MachineBasicBlock *BB) const {
22189 // This is pretty easy. We're taking the value that we received from
22190 // our load from the relocation, sticking it in either RDI (x86-64)
22191 // or EAX and doing an indirect call. The return value will then
22192 // be in the normal return register.
22193 MachineFunction *F = BB->getParent();
22194 const X86InstrInfo *TII = Subtarget->getInstrInfo();
22195 DebugLoc DL = MI->getDebugLoc();
22197 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
22198 assert(MI->getOperand(3).isGlobal() && "This should be a global");
22200 // Get a register mask for the lowered call.
22201 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
22202 // proper register mask.
22203 const uint32_t *RegMask =
22204 Subtarget->is64Bit() ?
22205 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
22206 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
22207 if (Subtarget->is64Bit()) {
22208 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22209 TII->get(X86::MOV64rm), X86::RDI)
22211 .addImm(0).addReg(0)
22212 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22213 MI->getOperand(3).getTargetFlags())
22215 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
22216 addDirectMem(MIB, X86::RDI);
22217 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
22218 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
22219 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22220 TII->get(X86::MOV32rm), X86::EAX)
22222 .addImm(0).addReg(0)
22223 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22224 MI->getOperand(3).getTargetFlags())
22226 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22227 addDirectMem(MIB, X86::EAX);
22228 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22230 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22231 TII->get(X86::MOV32rm), X86::EAX)
22232 .addReg(TII->getGlobalBaseReg(F))
22233 .addImm(0).addReg(0)
22234 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22235 MI->getOperand(3).getTargetFlags())
22237 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22238 addDirectMem(MIB, X86::EAX);
22239 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22242 MI->eraseFromParent(); // The pseudo instruction is gone now.
22246 MachineBasicBlock *
22247 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
22248 MachineBasicBlock *MBB) const {
22249 DebugLoc DL = MI->getDebugLoc();
22250 MachineFunction *MF = MBB->getParent();
22251 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22252 MachineRegisterInfo &MRI = MF->getRegInfo();
22254 const BasicBlock *BB = MBB->getBasicBlock();
22255 MachineFunction::iterator I = ++MBB->getIterator();
22257 // Memory Reference
22258 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22259 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22262 unsigned MemOpndSlot = 0;
22264 unsigned CurOp = 0;
22266 DstReg = MI->getOperand(CurOp++).getReg();
22267 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
22268 assert(RC->hasType(MVT::i32) && "Invalid destination!");
22269 unsigned mainDstReg = MRI.createVirtualRegister(RC);
22270 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
22272 MemOpndSlot = CurOp;
22274 MVT PVT = getPointerTy(MF->getDataLayout());
22275 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22276 "Invalid Pointer Size!");
22278 // For v = setjmp(buf), we generate
22281 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
22282 // SjLjSetup restoreMBB
22288 // v = phi(main, restore)
22291 // if base pointer being used, load it from frame
22294 MachineBasicBlock *thisMBB = MBB;
22295 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
22296 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
22297 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
22298 MF->insert(I, mainMBB);
22299 MF->insert(I, sinkMBB);
22300 MF->push_back(restoreMBB);
22301 restoreMBB->setHasAddressTaken();
22303 MachineInstrBuilder MIB;
22305 // Transfer the remainder of BB and its successor edges to sinkMBB.
22306 sinkMBB->splice(sinkMBB->begin(), MBB,
22307 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
22308 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
22311 unsigned PtrStoreOpc = 0;
22312 unsigned LabelReg = 0;
22313 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22314 Reloc::Model RM = MF->getTarget().getRelocationModel();
22315 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
22316 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
22318 // Prepare IP either in reg or imm.
22319 if (!UseImmLabel) {
22320 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
22321 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
22322 LabelReg = MRI.createVirtualRegister(PtrRC);
22323 if (Subtarget->is64Bit()) {
22324 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
22328 .addMBB(restoreMBB)
22331 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
22332 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
22333 .addReg(XII->getGlobalBaseReg(MF))
22336 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
22340 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
22342 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
22343 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22344 if (i == X86::AddrDisp)
22345 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
22347 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
22350 MIB.addReg(LabelReg);
22352 MIB.addMBB(restoreMBB);
22353 MIB.setMemRefs(MMOBegin, MMOEnd);
22355 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
22356 .addMBB(restoreMBB);
22358 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22359 MIB.addRegMask(RegInfo->getNoPreservedMask());
22360 thisMBB->addSuccessor(mainMBB);
22361 thisMBB->addSuccessor(restoreMBB);
22365 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
22366 mainMBB->addSuccessor(sinkMBB);
22369 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
22370 TII->get(X86::PHI), DstReg)
22371 .addReg(mainDstReg).addMBB(mainMBB)
22372 .addReg(restoreDstReg).addMBB(restoreMBB);
22375 if (RegInfo->hasBasePointer(*MF)) {
22376 const bool Uses64BitFramePtr =
22377 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
22378 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
22379 X86FI->setRestoreBasePointer(MF);
22380 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
22381 unsigned BasePtr = RegInfo->getBaseRegister();
22382 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
22383 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
22384 FramePtr, true, X86FI->getRestoreBasePointerOffset())
22385 .setMIFlag(MachineInstr::FrameSetup);
22387 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
22388 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
22389 restoreMBB->addSuccessor(sinkMBB);
22391 MI->eraseFromParent();
22395 MachineBasicBlock *
22396 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
22397 MachineBasicBlock *MBB) const {
22398 DebugLoc DL = MI->getDebugLoc();
22399 MachineFunction *MF = MBB->getParent();
22400 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22401 MachineRegisterInfo &MRI = MF->getRegInfo();
22403 // Memory Reference
22404 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22405 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22407 MVT PVT = getPointerTy(MF->getDataLayout());
22408 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22409 "Invalid Pointer Size!");
22411 const TargetRegisterClass *RC =
22412 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
22413 unsigned Tmp = MRI.createVirtualRegister(RC);
22414 // Since FP is only updated here but NOT referenced, it's treated as GPR.
22415 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22416 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
22417 unsigned SP = RegInfo->getStackRegister();
22419 MachineInstrBuilder MIB;
22421 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22422 const int64_t SPOffset = 2 * PVT.getStoreSize();
22424 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
22425 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
22428 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
22429 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
22430 MIB.addOperand(MI->getOperand(i));
22431 MIB.setMemRefs(MMOBegin, MMOEnd);
22433 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
22434 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22435 if (i == X86::AddrDisp)
22436 MIB.addDisp(MI->getOperand(i), LabelOffset);
22438 MIB.addOperand(MI->getOperand(i));
22440 MIB.setMemRefs(MMOBegin, MMOEnd);
22442 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
22443 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22444 if (i == X86::AddrDisp)
22445 MIB.addDisp(MI->getOperand(i), SPOffset);
22447 MIB.addOperand(MI->getOperand(i));
22449 MIB.setMemRefs(MMOBegin, MMOEnd);
22451 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
22453 MI->eraseFromParent();
22457 // Replace 213-type (isel default) FMA3 instructions with 231-type for
22458 // accumulator loops. Writing back to the accumulator allows the coalescer
22459 // to remove extra copies in the loop.
22460 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
22461 MachineBasicBlock *
22462 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
22463 MachineBasicBlock *MBB) const {
22464 MachineOperand &AddendOp = MI->getOperand(3);
22466 // Bail out early if the addend isn't a register - we can't switch these.
22467 if (!AddendOp.isReg())
22470 MachineFunction &MF = *MBB->getParent();
22471 MachineRegisterInfo &MRI = MF.getRegInfo();
22473 // Check whether the addend is defined by a PHI:
22474 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
22475 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
22476 if (!AddendDef.isPHI())
22479 // Look for the following pattern:
22481 // %addend = phi [%entry, 0], [%loop, %result]
22483 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
22487 // %addend = phi [%entry, 0], [%loop, %result]
22489 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
22491 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
22492 assert(AddendDef.getOperand(i).isReg());
22493 MachineOperand PHISrcOp = AddendDef.getOperand(i);
22494 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
22495 if (&PHISrcInst == MI) {
22496 // Found a matching instruction.
22497 unsigned NewFMAOpc = 0;
22498 switch (MI->getOpcode()) {
22499 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
22500 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
22501 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
22502 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
22503 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
22504 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
22505 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
22506 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
22507 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
22508 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
22509 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
22510 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
22511 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
22512 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
22513 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
22514 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
22515 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
22516 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
22517 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
22518 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
22520 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
22521 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
22522 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
22523 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
22524 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
22525 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
22526 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
22527 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
22528 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
22529 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
22530 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
22531 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
22532 default: llvm_unreachable("Unrecognized FMA variant.");
22535 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22536 MachineInstrBuilder MIB =
22537 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
22538 .addOperand(MI->getOperand(0))
22539 .addOperand(MI->getOperand(3))
22540 .addOperand(MI->getOperand(2))
22541 .addOperand(MI->getOperand(1));
22542 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
22543 MI->eraseFromParent();
22550 MachineBasicBlock *
22551 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
22552 MachineBasicBlock *BB) const {
22553 switch (MI->getOpcode()) {
22554 default: llvm_unreachable("Unexpected instr type to insert");
22555 case X86::TAILJMPd64:
22556 case X86::TAILJMPr64:
22557 case X86::TAILJMPm64:
22558 case X86::TAILJMPd64_REX:
22559 case X86::TAILJMPr64_REX:
22560 case X86::TAILJMPm64_REX:
22561 llvm_unreachable("TAILJMP64 would not be touched here.");
22562 case X86::TCRETURNdi64:
22563 case X86::TCRETURNri64:
22564 case X86::TCRETURNmi64:
22566 case X86::WIN_ALLOCA:
22567 return EmitLoweredWinAlloca(MI, BB);
22568 case X86::CATCHRET:
22569 return EmitLoweredCatchRet(MI, BB);
22570 case X86::CATCHPAD:
22571 return EmitLoweredCatchPad(MI, BB);
22572 case X86::SEG_ALLOCA_32:
22573 case X86::SEG_ALLOCA_64:
22574 return EmitLoweredSegAlloca(MI, BB);
22575 case X86::TLSCall_32:
22576 case X86::TLSCall_64:
22577 return EmitLoweredTLSCall(MI, BB);
22578 case X86::CMOV_FR32:
22579 case X86::CMOV_FR64:
22580 case X86::CMOV_FR128:
22581 case X86::CMOV_GR8:
22582 case X86::CMOV_GR16:
22583 case X86::CMOV_GR32:
22584 case X86::CMOV_RFP32:
22585 case X86::CMOV_RFP64:
22586 case X86::CMOV_RFP80:
22587 case X86::CMOV_V2F64:
22588 case X86::CMOV_V2I64:
22589 case X86::CMOV_V4F32:
22590 case X86::CMOV_V4F64:
22591 case X86::CMOV_V4I64:
22592 case X86::CMOV_V16F32:
22593 case X86::CMOV_V8F32:
22594 case X86::CMOV_V8F64:
22595 case X86::CMOV_V8I64:
22596 case X86::CMOV_V8I1:
22597 case X86::CMOV_V16I1:
22598 case X86::CMOV_V32I1:
22599 case X86::CMOV_V64I1:
22600 return EmitLoweredSelect(MI, BB);
22602 case X86::RDFLAGS32:
22603 case X86::RDFLAGS64: {
22604 DebugLoc DL = MI->getDebugLoc();
22605 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22607 MI->getOpcode() == X86::RDFLAGS32 ? X86::PUSHF32 : X86::PUSHF64;
22609 MI->getOpcode() == X86::RDFLAGS32 ? X86::POP32r : X86::POP64r;
22610 BuildMI(*BB, MI, DL, TII->get(PushF));
22611 BuildMI(*BB, MI, DL, TII->get(Pop), MI->getOperand(0).getReg());
22613 MI->eraseFromParent(); // The pseudo is gone now.
22617 case X86::WRFLAGS32:
22618 case X86::WRFLAGS64: {
22619 DebugLoc DL = MI->getDebugLoc();
22620 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22622 MI->getOpcode() == X86::WRFLAGS32 ? X86::PUSH32r : X86::PUSH64r;
22624 MI->getOpcode() == X86::WRFLAGS32 ? X86::POPF32 : X86::POPF64;
22625 BuildMI(*BB, MI, DL, TII->get(Push)).addReg(MI->getOperand(0).getReg());
22626 BuildMI(*BB, MI, DL, TII->get(PopF));
22628 MI->eraseFromParent(); // The pseudo is gone now.
22632 case X86::RELEASE_FADD32mr:
22633 case X86::RELEASE_FADD64mr:
22634 return EmitLoweredAtomicFP(MI, BB);
22636 case X86::FP32_TO_INT16_IN_MEM:
22637 case X86::FP32_TO_INT32_IN_MEM:
22638 case X86::FP32_TO_INT64_IN_MEM:
22639 case X86::FP64_TO_INT16_IN_MEM:
22640 case X86::FP64_TO_INT32_IN_MEM:
22641 case X86::FP64_TO_INT64_IN_MEM:
22642 case X86::FP80_TO_INT16_IN_MEM:
22643 case X86::FP80_TO_INT32_IN_MEM:
22644 case X86::FP80_TO_INT64_IN_MEM: {
22645 MachineFunction *F = BB->getParent();
22646 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22647 DebugLoc DL = MI->getDebugLoc();
22649 // Change the floating point control register to use "round towards zero"
22650 // mode when truncating to an integer value.
22651 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
22652 addFrameReference(BuildMI(*BB, MI, DL,
22653 TII->get(X86::FNSTCW16m)), CWFrameIdx);
22655 // Load the old value of the high byte of the control word...
22657 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
22658 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
22661 // Set the high part to be round to zero...
22662 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
22665 // Reload the modified control word now...
22666 addFrameReference(BuildMI(*BB, MI, DL,
22667 TII->get(X86::FLDCW16m)), CWFrameIdx);
22669 // Restore the memory image of control word to original value
22670 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
22673 // Get the X86 opcode to use.
22675 switch (MI->getOpcode()) {
22676 default: llvm_unreachable("illegal opcode!");
22677 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
22678 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22679 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22680 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22681 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22682 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22683 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22684 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22685 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22689 MachineOperand &Op = MI->getOperand(0);
22691 AM.BaseType = X86AddressMode::RegBase;
22692 AM.Base.Reg = Op.getReg();
22694 AM.BaseType = X86AddressMode::FrameIndexBase;
22695 AM.Base.FrameIndex = Op.getIndex();
22697 Op = MI->getOperand(1);
22699 AM.Scale = Op.getImm();
22700 Op = MI->getOperand(2);
22702 AM.IndexReg = Op.getImm();
22703 Op = MI->getOperand(3);
22704 if (Op.isGlobal()) {
22705 AM.GV = Op.getGlobal();
22707 AM.Disp = Op.getImm();
22709 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22710 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22712 // Reload the original control word now.
22713 addFrameReference(BuildMI(*BB, MI, DL,
22714 TII->get(X86::FLDCW16m)), CWFrameIdx);
22716 MI->eraseFromParent(); // The pseudo instruction is gone now.
22719 // String/text processing lowering.
22720 case X86::PCMPISTRM128REG:
22721 case X86::VPCMPISTRM128REG:
22722 case X86::PCMPISTRM128MEM:
22723 case X86::VPCMPISTRM128MEM:
22724 case X86::PCMPESTRM128REG:
22725 case X86::VPCMPESTRM128REG:
22726 case X86::PCMPESTRM128MEM:
22727 case X86::VPCMPESTRM128MEM:
22728 assert(Subtarget->hasSSE42() &&
22729 "Target must have SSE4.2 or AVX features enabled");
22730 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22732 // String/text processing lowering.
22733 case X86::PCMPISTRIREG:
22734 case X86::VPCMPISTRIREG:
22735 case X86::PCMPISTRIMEM:
22736 case X86::VPCMPISTRIMEM:
22737 case X86::PCMPESTRIREG:
22738 case X86::VPCMPESTRIREG:
22739 case X86::PCMPESTRIMEM:
22740 case X86::VPCMPESTRIMEM:
22741 assert(Subtarget->hasSSE42() &&
22742 "Target must have SSE4.2 or AVX features enabled");
22743 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22745 // Thread synchronization.
22747 return EmitMonitor(MI, BB, Subtarget);
22750 return EmitWRPKRU(MI, BB, Subtarget);
22752 return EmitRDPKRU(MI, BB, Subtarget);
22755 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22757 case X86::VASTART_SAVE_XMM_REGS:
22758 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22760 case X86::VAARG_64:
22761 return EmitVAARG64WithCustomInserter(MI, BB);
22763 case X86::EH_SjLj_SetJmp32:
22764 case X86::EH_SjLj_SetJmp64:
22765 return emitEHSjLjSetJmp(MI, BB);
22767 case X86::EH_SjLj_LongJmp32:
22768 case X86::EH_SjLj_LongJmp64:
22769 return emitEHSjLjLongJmp(MI, BB);
22771 case TargetOpcode::STATEPOINT:
22772 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22773 // this point in the process. We diverge later.
22774 return emitPatchPoint(MI, BB);
22776 case TargetOpcode::STACKMAP:
22777 case TargetOpcode::PATCHPOINT:
22778 return emitPatchPoint(MI, BB);
22780 case X86::VFMADDPDr213r:
22781 case X86::VFMADDPSr213r:
22782 case X86::VFMADDSDr213r:
22783 case X86::VFMADDSSr213r:
22784 case X86::VFMSUBPDr213r:
22785 case X86::VFMSUBPSr213r:
22786 case X86::VFMSUBSDr213r:
22787 case X86::VFMSUBSSr213r:
22788 case X86::VFNMADDPDr213r:
22789 case X86::VFNMADDPSr213r:
22790 case X86::VFNMADDSDr213r:
22791 case X86::VFNMADDSSr213r:
22792 case X86::VFNMSUBPDr213r:
22793 case X86::VFNMSUBPSr213r:
22794 case X86::VFNMSUBSDr213r:
22795 case X86::VFNMSUBSSr213r:
22796 case X86::VFMADDSUBPDr213r:
22797 case X86::VFMADDSUBPSr213r:
22798 case X86::VFMSUBADDPDr213r:
22799 case X86::VFMSUBADDPSr213r:
22800 case X86::VFMADDPDr213rY:
22801 case X86::VFMADDPSr213rY:
22802 case X86::VFMSUBPDr213rY:
22803 case X86::VFMSUBPSr213rY:
22804 case X86::VFNMADDPDr213rY:
22805 case X86::VFNMADDPSr213rY:
22806 case X86::VFNMSUBPDr213rY:
22807 case X86::VFNMSUBPSr213rY:
22808 case X86::VFMADDSUBPDr213rY:
22809 case X86::VFMADDSUBPSr213rY:
22810 case X86::VFMSUBADDPDr213rY:
22811 case X86::VFMSUBADDPSr213rY:
22812 return emitFMA3Instr(MI, BB);
22816 //===----------------------------------------------------------------------===//
22817 // X86 Optimization Hooks
22818 //===----------------------------------------------------------------------===//
22820 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22823 const SelectionDAG &DAG,
22824 unsigned Depth) const {
22825 unsigned BitWidth = KnownZero.getBitWidth();
22826 unsigned Opc = Op.getOpcode();
22827 assert((Opc >= ISD::BUILTIN_OP_END ||
22828 Opc == ISD::INTRINSIC_WO_CHAIN ||
22829 Opc == ISD::INTRINSIC_W_CHAIN ||
22830 Opc == ISD::INTRINSIC_VOID) &&
22831 "Should use MaskedValueIsZero if you don't know whether Op"
22832 " is a target node!");
22834 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22848 // These nodes' second result is a boolean.
22849 if (Op.getResNo() == 0)
22852 case X86ISD::SETCC:
22853 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22855 case ISD::INTRINSIC_WO_CHAIN: {
22856 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22857 unsigned NumLoBits = 0;
22860 case Intrinsic::x86_sse_movmsk_ps:
22861 case Intrinsic::x86_avx_movmsk_ps_256:
22862 case Intrinsic::x86_sse2_movmsk_pd:
22863 case Intrinsic::x86_avx_movmsk_pd_256:
22864 case Intrinsic::x86_mmx_pmovmskb:
22865 case Intrinsic::x86_sse2_pmovmskb_128:
22866 case Intrinsic::x86_avx2_pmovmskb: {
22867 // High bits of movmskp{s|d}, pmovmskb are known zero.
22869 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22870 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22871 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22872 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22873 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22874 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22875 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22876 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22878 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22887 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22889 const SelectionDAG &,
22890 unsigned Depth) const {
22891 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22892 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22893 return Op.getValueType().getScalarSizeInBits();
22899 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22900 /// node is a GlobalAddress + offset.
22901 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22902 const GlobalValue* &GA,
22903 int64_t &Offset) const {
22904 if (N->getOpcode() == X86ISD::Wrapper) {
22905 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22906 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22907 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22911 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22914 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22915 /// FIXME: This could be expanded to support 512 bit vectors as well.
22916 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22917 TargetLowering::DAGCombinerInfo &DCI,
22918 const X86Subtarget* Subtarget) {
22920 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22921 SDValue V1 = SVOp->getOperand(0);
22922 SDValue V2 = SVOp->getOperand(1);
22923 MVT VT = SVOp->getSimpleValueType(0);
22924 unsigned NumElems = VT.getVectorNumElements();
22926 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22927 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22931 // V UNDEF BUILD_VECTOR UNDEF
22933 // CONCAT_VECTOR CONCAT_VECTOR
22936 // RESULT: V + zero extended
22938 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22939 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22940 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22943 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22946 // To match the shuffle mask, the first half of the mask should
22947 // be exactly the first vector, and all the rest a splat with the
22948 // first element of the second one.
22949 for (unsigned i = 0; i != NumElems/2; ++i)
22950 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22951 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22954 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22955 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22956 if (Ld->hasNUsesOfValue(1, 0)) {
22957 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22958 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22960 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22962 Ld->getPointerInfo(),
22963 Ld->getAlignment(),
22964 false/*isVolatile*/, true/*ReadMem*/,
22965 false/*WriteMem*/);
22967 // Make sure the newly-created LOAD is in the same position as Ld in
22968 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22969 // and update uses of Ld's output chain to use the TokenFactor.
22970 if (Ld->hasAnyUseOfValue(1)) {
22971 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22972 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22973 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22974 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22975 SDValue(ResNode.getNode(), 1));
22978 return DAG.getBitcast(VT, ResNode);
22982 // Emit a zeroed vector and insert the desired subvector on its
22984 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22985 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22986 return DCI.CombineTo(N, InsV);
22992 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22995 /// This is the leaf of the recursive combinine below. When we have found some
22996 /// chain of single-use x86 shuffle instructions and accumulated the combined
22997 /// shuffle mask represented by them, this will try to pattern match that mask
22998 /// into either a single instruction if there is a special purpose instruction
22999 /// for this operation, or into a PSHUFB instruction which is a fully general
23000 /// instruction but should only be used to replace chains over a certain depth.
23001 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
23002 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
23003 TargetLowering::DAGCombinerInfo &DCI,
23004 const X86Subtarget *Subtarget) {
23005 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
23007 // Find the operand that enters the chain. Note that multiple uses are OK
23008 // here, we're not going to remove the operand we find.
23009 SDValue Input = Op.getOperand(0);
23010 while (Input.getOpcode() == ISD::BITCAST)
23011 Input = Input.getOperand(0);
23013 MVT VT = Input.getSimpleValueType();
23014 MVT RootVT = Root.getSimpleValueType();
23017 if (Mask.size() == 1) {
23018 int Index = Mask[0];
23019 assert((Index >= 0 || Index == SM_SentinelUndef ||
23020 Index == SM_SentinelZero) &&
23021 "Invalid shuffle index found!");
23023 // We may end up with an accumulated mask of size 1 as a result of
23024 // widening of shuffle operands (see function canWidenShuffleElements).
23025 // If the only shuffle index is equal to SM_SentinelZero then propagate
23026 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
23027 // mask, and therefore the entire chain of shuffles can be folded away.
23028 if (Index == SM_SentinelZero)
23029 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
23031 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
23036 // Use the float domain if the operand type is a floating point type.
23037 bool FloatDomain = VT.isFloatingPoint();
23039 // For floating point shuffles, we don't have free copies in the shuffle
23040 // instructions or the ability to load as part of the instruction, so
23041 // canonicalize their shuffles to UNPCK or MOV variants.
23043 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
23044 // vectors because it can have a load folded into it that UNPCK cannot. This
23045 // doesn't preclude something switching to the shorter encoding post-RA.
23047 // FIXME: Should teach these routines about AVX vector widths.
23048 if (FloatDomain && VT.is128BitVector()) {
23049 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
23050 bool Lo = Mask.equals({0, 0});
23053 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
23054 // is no slower than UNPCKLPD but has the option to fold the input operand
23055 // into even an unaligned memory load.
23056 if (Lo && Subtarget->hasSSE3()) {
23057 Shuffle = X86ISD::MOVDDUP;
23058 ShuffleVT = MVT::v2f64;
23060 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
23061 // than the UNPCK variants.
23062 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
23063 ShuffleVT = MVT::v4f32;
23065 if (Depth == 1 && Root->getOpcode() == Shuffle)
23066 return false; // Nothing to do!
23067 Op = DAG.getBitcast(ShuffleVT, Input);
23068 DCI.AddToWorklist(Op.getNode());
23069 if (Shuffle == X86ISD::MOVDDUP)
23070 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23072 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23073 DCI.AddToWorklist(Op.getNode());
23074 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23078 if (Subtarget->hasSSE3() &&
23079 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
23080 bool Lo = Mask.equals({0, 0, 2, 2});
23081 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
23082 MVT ShuffleVT = MVT::v4f32;
23083 if (Depth == 1 && Root->getOpcode() == Shuffle)
23084 return false; // Nothing to do!
23085 Op = DAG.getBitcast(ShuffleVT, Input);
23086 DCI.AddToWorklist(Op.getNode());
23087 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23088 DCI.AddToWorklist(Op.getNode());
23089 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23093 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
23094 bool Lo = Mask.equals({0, 0, 1, 1});
23095 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23096 MVT ShuffleVT = MVT::v4f32;
23097 if (Depth == 1 && Root->getOpcode() == Shuffle)
23098 return false; // Nothing to do!
23099 Op = DAG.getBitcast(ShuffleVT, Input);
23100 DCI.AddToWorklist(Op.getNode());
23101 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23102 DCI.AddToWorklist(Op.getNode());
23103 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23109 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
23110 // variants as none of these have single-instruction variants that are
23111 // superior to the UNPCK formulation.
23112 if (!FloatDomain && VT.is128BitVector() &&
23113 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23114 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
23115 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
23117 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
23118 bool Lo = Mask[0] == 0;
23119 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23120 if (Depth == 1 && Root->getOpcode() == Shuffle)
23121 return false; // Nothing to do!
23123 switch (Mask.size()) {
23125 ShuffleVT = MVT::v8i16;
23128 ShuffleVT = MVT::v16i8;
23131 llvm_unreachable("Impossible mask size!");
23133 Op = DAG.getBitcast(ShuffleVT, Input);
23134 DCI.AddToWorklist(Op.getNode());
23135 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23136 DCI.AddToWorklist(Op.getNode());
23137 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23142 // Don't try to re-form single instruction chains under any circumstances now
23143 // that we've done encoding canonicalization for them.
23147 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
23148 // can replace them with a single PSHUFB instruction profitably. Intel's
23149 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
23150 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
23151 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
23152 SmallVector<SDValue, 16> PSHUFBMask;
23153 int NumBytes = VT.getSizeInBits() / 8;
23154 int Ratio = NumBytes / Mask.size();
23155 for (int i = 0; i < NumBytes; ++i) {
23156 if (Mask[i / Ratio] == SM_SentinelUndef) {
23157 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
23160 int M = Mask[i / Ratio] != SM_SentinelZero
23161 ? Ratio * Mask[i / Ratio] + i % Ratio
23163 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
23165 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
23166 Op = DAG.getBitcast(ByteVT, Input);
23167 DCI.AddToWorklist(Op.getNode());
23168 SDValue PSHUFBMaskOp =
23169 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
23170 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
23171 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
23172 DCI.AddToWorklist(Op.getNode());
23173 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23178 // Failed to find any combines.
23182 /// \brief Fully generic combining of x86 shuffle instructions.
23184 /// This should be the last combine run over the x86 shuffle instructions. Once
23185 /// they have been fully optimized, this will recursively consider all chains
23186 /// of single-use shuffle instructions, build a generic model of the cumulative
23187 /// shuffle operation, and check for simpler instructions which implement this
23188 /// operation. We use this primarily for two purposes:
23190 /// 1) Collapse generic shuffles to specialized single instructions when
23191 /// equivalent. In most cases, this is just an encoding size win, but
23192 /// sometimes we will collapse multiple generic shuffles into a single
23193 /// special-purpose shuffle.
23194 /// 2) Look for sequences of shuffle instructions with 3 or more total
23195 /// instructions, and replace them with the slightly more expensive SSSE3
23196 /// PSHUFB instruction if available. We do this as the last combining step
23197 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
23198 /// a suitable short sequence of other instructions. The PHUFB will either
23199 /// use a register or have to read from memory and so is slightly (but only
23200 /// slightly) more expensive than the other shuffle instructions.
23202 /// Because this is inherently a quadratic operation (for each shuffle in
23203 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
23204 /// This should never be an issue in practice as the shuffle lowering doesn't
23205 /// produce sequences of more than 8 instructions.
23207 /// FIXME: We will currently miss some cases where the redundant shuffling
23208 /// would simplify under the threshold for PSHUFB formation because of
23209 /// combine-ordering. To fix this, we should do the redundant instruction
23210 /// combining in this recursive walk.
23211 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
23212 ArrayRef<int> RootMask,
23213 int Depth, bool HasPSHUFB,
23215 TargetLowering::DAGCombinerInfo &DCI,
23216 const X86Subtarget *Subtarget) {
23217 // Bound the depth of our recursive combine because this is ultimately
23218 // quadratic in nature.
23222 // Directly rip through bitcasts to find the underlying operand.
23223 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
23224 Op = Op.getOperand(0);
23226 MVT VT = Op.getSimpleValueType();
23227 if (!VT.isVector())
23228 return false; // Bail if we hit a non-vector.
23230 assert(Root.getSimpleValueType().isVector() &&
23231 "Shuffles operate on vector types!");
23232 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
23233 "Can only combine shuffles of the same vector register size.");
23235 if (!isTargetShuffle(Op.getOpcode()))
23237 SmallVector<int, 16> OpMask;
23239 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, true, OpMask, IsUnary);
23240 // We only can combine unary shuffles which we can decode the mask for.
23241 if (!HaveMask || !IsUnary)
23244 assert(VT.getVectorNumElements() == OpMask.size() &&
23245 "Different mask size from vector size!");
23246 assert(((RootMask.size() > OpMask.size() &&
23247 RootMask.size() % OpMask.size() == 0) ||
23248 (OpMask.size() > RootMask.size() &&
23249 OpMask.size() % RootMask.size() == 0) ||
23250 OpMask.size() == RootMask.size()) &&
23251 "The smaller number of elements must divide the larger.");
23252 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
23253 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
23254 assert(((RootRatio == 1 && OpRatio == 1) ||
23255 (RootRatio == 1) != (OpRatio == 1)) &&
23256 "Must not have a ratio for both incoming and op masks!");
23258 SmallVector<int, 16> Mask;
23259 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
23261 // Merge this shuffle operation's mask into our accumulated mask. Note that
23262 // this shuffle's mask will be the first applied to the input, followed by the
23263 // root mask to get us all the way to the root value arrangement. The reason
23264 // for this order is that we are recursing up the operation chain.
23265 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
23266 int RootIdx = i / RootRatio;
23267 if (RootMask[RootIdx] < 0) {
23268 // This is a zero or undef lane, we're done.
23269 Mask.push_back(RootMask[RootIdx]);
23273 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
23274 int OpIdx = RootMaskedIdx / OpRatio;
23275 if (OpMask[OpIdx] < 0) {
23276 // The incoming lanes are zero or undef, it doesn't matter which ones we
23278 Mask.push_back(OpMask[OpIdx]);
23282 // Ok, we have non-zero lanes, map them through.
23283 Mask.push_back(OpMask[OpIdx] * OpRatio +
23284 RootMaskedIdx % OpRatio);
23287 // See if we can recurse into the operand to combine more things.
23288 switch (Op.getOpcode()) {
23289 case X86ISD::PSHUFB:
23291 case X86ISD::PSHUFD:
23292 case X86ISD::PSHUFHW:
23293 case X86ISD::PSHUFLW:
23294 if (Op.getOperand(0).hasOneUse() &&
23295 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23296 HasPSHUFB, DAG, DCI, Subtarget))
23300 case X86ISD::UNPCKL:
23301 case X86ISD::UNPCKH:
23302 assert(Op.getOperand(0) == Op.getOperand(1) &&
23303 "We only combine unary shuffles!");
23304 // We can't check for single use, we have to check that this shuffle is the
23306 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
23307 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23308 HasPSHUFB, DAG, DCI, Subtarget))
23313 // Minor canonicalization of the accumulated shuffle mask to make it easier
23314 // to match below. All this does is detect masks with squential pairs of
23315 // elements, and shrink them to the half-width mask. It does this in a loop
23316 // so it will reduce the size of the mask to the minimal width mask which
23317 // performs an equivalent shuffle.
23318 SmallVector<int, 16> WidenedMask;
23319 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
23320 Mask = std::move(WidenedMask);
23321 WidenedMask.clear();
23324 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
23328 /// \brief Get the PSHUF-style mask from PSHUF node.
23330 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
23331 /// PSHUF-style masks that can be reused with such instructions.
23332 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
23333 MVT VT = N.getSimpleValueType();
23334 SmallVector<int, 4> Mask;
23336 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, false, Mask, IsUnary);
23340 // If we have more than 128-bits, only the low 128-bits of shuffle mask
23341 // matter. Check that the upper masks are repeats and remove them.
23342 if (VT.getSizeInBits() > 128) {
23343 int LaneElts = 128 / VT.getScalarSizeInBits();
23345 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
23346 for (int j = 0; j < LaneElts; ++j)
23347 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
23348 "Mask doesn't repeat in high 128-bit lanes!");
23350 Mask.resize(LaneElts);
23353 switch (N.getOpcode()) {
23354 case X86ISD::PSHUFD:
23356 case X86ISD::PSHUFLW:
23359 case X86ISD::PSHUFHW:
23360 Mask.erase(Mask.begin(), Mask.begin() + 4);
23361 for (int &M : Mask)
23365 llvm_unreachable("No valid shuffle instruction found!");
23369 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
23371 /// We walk up the chain and look for a combinable shuffle, skipping over
23372 /// shuffles that we could hoist this shuffle's transformation past without
23373 /// altering anything.
23375 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
23377 TargetLowering::DAGCombinerInfo &DCI) {
23378 assert(N.getOpcode() == X86ISD::PSHUFD &&
23379 "Called with something other than an x86 128-bit half shuffle!");
23382 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
23383 // of the shuffles in the chain so that we can form a fresh chain to replace
23385 SmallVector<SDValue, 8> Chain;
23386 SDValue V = N.getOperand(0);
23387 for (; V.hasOneUse(); V = V.getOperand(0)) {
23388 switch (V.getOpcode()) {
23390 return SDValue(); // Nothing combined!
23393 // Skip bitcasts as we always know the type for the target specific
23397 case X86ISD::PSHUFD:
23398 // Found another dword shuffle.
23401 case X86ISD::PSHUFLW:
23402 // Check that the low words (being shuffled) are the identity in the
23403 // dword shuffle, and the high words are self-contained.
23404 if (Mask[0] != 0 || Mask[1] != 1 ||
23405 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
23408 Chain.push_back(V);
23411 case X86ISD::PSHUFHW:
23412 // Check that the high words (being shuffled) are the identity in the
23413 // dword shuffle, and the low words are self-contained.
23414 if (Mask[2] != 2 || Mask[3] != 3 ||
23415 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
23418 Chain.push_back(V);
23421 case X86ISD::UNPCKL:
23422 case X86ISD::UNPCKH:
23423 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
23424 // shuffle into a preceding word shuffle.
23425 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
23426 V.getSimpleValueType().getVectorElementType() != MVT::i16)
23429 // Search for a half-shuffle which we can combine with.
23430 unsigned CombineOp =
23431 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
23432 if (V.getOperand(0) != V.getOperand(1) ||
23433 !V->isOnlyUserOf(V.getOperand(0).getNode()))
23435 Chain.push_back(V);
23436 V = V.getOperand(0);
23438 switch (V.getOpcode()) {
23440 return SDValue(); // Nothing to combine.
23442 case X86ISD::PSHUFLW:
23443 case X86ISD::PSHUFHW:
23444 if (V.getOpcode() == CombineOp)
23447 Chain.push_back(V);
23451 V = V.getOperand(0);
23455 } while (V.hasOneUse());
23458 // Break out of the loop if we break out of the switch.
23462 if (!V.hasOneUse())
23463 // We fell out of the loop without finding a viable combining instruction.
23466 // Merge this node's mask and our incoming mask.
23467 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23468 for (int &M : Mask)
23470 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
23471 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23473 // Rebuild the chain around this new shuffle.
23474 while (!Chain.empty()) {
23475 SDValue W = Chain.pop_back_val();
23477 if (V.getValueType() != W.getOperand(0).getValueType())
23478 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
23480 switch (W.getOpcode()) {
23482 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
23484 case X86ISD::UNPCKL:
23485 case X86ISD::UNPCKH:
23486 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
23489 case X86ISD::PSHUFD:
23490 case X86ISD::PSHUFLW:
23491 case X86ISD::PSHUFHW:
23492 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
23496 if (V.getValueType() != N.getValueType())
23497 V = DAG.getBitcast(N.getValueType(), V);
23499 // Return the new chain to replace N.
23503 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
23506 /// We walk up the chain, skipping shuffles of the other half and looking
23507 /// through shuffles which switch halves trying to find a shuffle of the same
23508 /// pair of dwords.
23509 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
23511 TargetLowering::DAGCombinerInfo &DCI) {
23513 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
23514 "Called with something other than an x86 128-bit half shuffle!");
23516 unsigned CombineOpcode = N.getOpcode();
23518 // Walk up a single-use chain looking for a combinable shuffle.
23519 SDValue V = N.getOperand(0);
23520 for (; V.hasOneUse(); V = V.getOperand(0)) {
23521 switch (V.getOpcode()) {
23523 return false; // Nothing combined!
23526 // Skip bitcasts as we always know the type for the target specific
23530 case X86ISD::PSHUFLW:
23531 case X86ISD::PSHUFHW:
23532 if (V.getOpcode() == CombineOpcode)
23535 // Other-half shuffles are no-ops.
23538 // Break out of the loop if we break out of the switch.
23542 if (!V.hasOneUse())
23543 // We fell out of the loop without finding a viable combining instruction.
23546 // Combine away the bottom node as its shuffle will be accumulated into
23547 // a preceding shuffle.
23548 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23550 // Record the old value.
23553 // Merge this node's mask and our incoming mask (adjusted to account for all
23554 // the pshufd instructions encountered).
23555 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23556 for (int &M : Mask)
23558 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
23559 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23561 // Check that the shuffles didn't cancel each other out. If not, we need to
23562 // combine to the new one.
23564 // Replace the combinable shuffle with the combined one, updating all users
23565 // so that we re-evaluate the chain here.
23566 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
23571 /// \brief Try to combine x86 target specific shuffles.
23572 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
23573 TargetLowering::DAGCombinerInfo &DCI,
23574 const X86Subtarget *Subtarget) {
23576 MVT VT = N.getSimpleValueType();
23577 SmallVector<int, 4> Mask;
23579 switch (N.getOpcode()) {
23580 case X86ISD::PSHUFD:
23581 case X86ISD::PSHUFLW:
23582 case X86ISD::PSHUFHW:
23583 Mask = getPSHUFShuffleMask(N);
23584 assert(Mask.size() == 4);
23586 case X86ISD::UNPCKL: {
23587 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
23588 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
23589 // moves upper half elements into the lower half part. For example:
23591 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
23593 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
23595 // will be combined to:
23597 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
23599 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
23600 // happen due to advanced instructions.
23601 if (!VT.is128BitVector())
23604 auto Op0 = N.getOperand(0);
23605 auto Op1 = N.getOperand(1);
23606 if (Op0.getOpcode() == ISD::UNDEF &&
23607 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
23608 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
23610 unsigned NumElts = VT.getVectorNumElements();
23611 SmallVector<int, 8> ExpectedMask(NumElts, -1);
23612 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
23615 auto ShufOp = Op1.getOperand(0);
23616 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
23617 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
23621 case X86ISD::BLENDI: {
23622 SDValue V0 = N->getOperand(0);
23623 SDValue V1 = N->getOperand(1);
23624 assert(VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() &&
23625 "Unexpected input vector types");
23627 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23628 // operands and changing the mask to 1. This saves us a bunch of
23629 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23630 // x86InstrInfo knows how to commute this back after instruction selection
23631 // if it would help register allocation.
23633 // TODO: If optimizing for size or a processor that doesn't suffer from
23634 // partial register update stalls, this should be transformed into a MOVSD
23635 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23637 if (VT == MVT::v2f64)
23638 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23639 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23640 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
23641 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23650 // Nuke no-op shuffles that show up after combining.
23651 if (isNoopShuffleMask(Mask))
23652 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23654 // Look for simplifications involving one or two shuffle instructions.
23655 SDValue V = N.getOperand(0);
23656 switch (N.getOpcode()) {
23659 case X86ISD::PSHUFLW:
23660 case X86ISD::PSHUFHW:
23661 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
23663 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
23664 return SDValue(); // We combined away this shuffle, so we're done.
23666 // See if this reduces to a PSHUFD which is no more expensive and can
23667 // combine with more operations. Note that it has to at least flip the
23668 // dwords as otherwise it would have been removed as a no-op.
23669 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
23670 int DMask[] = {0, 1, 2, 3};
23671 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
23672 DMask[DOffset + 0] = DOffset + 1;
23673 DMask[DOffset + 1] = DOffset + 0;
23674 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
23675 V = DAG.getBitcast(DVT, V);
23676 DCI.AddToWorklist(V.getNode());
23677 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
23678 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
23679 DCI.AddToWorklist(V.getNode());
23680 return DAG.getBitcast(VT, V);
23683 // Look for shuffle patterns which can be implemented as a single unpack.
23684 // FIXME: This doesn't handle the location of the PSHUFD generically, and
23685 // only works when we have a PSHUFD followed by two half-shuffles.
23686 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
23687 (V.getOpcode() == X86ISD::PSHUFLW ||
23688 V.getOpcode() == X86ISD::PSHUFHW) &&
23689 V.getOpcode() != N.getOpcode() &&
23691 SDValue D = V.getOperand(0);
23692 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
23693 D = D.getOperand(0);
23694 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
23695 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23696 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
23697 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23698 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23700 for (int i = 0; i < 4; ++i) {
23701 WordMask[i + NOffset] = Mask[i] + NOffset;
23702 WordMask[i + VOffset] = VMask[i] + VOffset;
23704 // Map the word mask through the DWord mask.
23706 for (int i = 0; i < 8; ++i)
23707 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23708 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23709 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23710 // We can replace all three shuffles with an unpack.
23711 V = DAG.getBitcast(VT, D.getOperand(0));
23712 DCI.AddToWorklist(V.getNode());
23713 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23722 case X86ISD::PSHUFD:
23723 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23732 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23734 /// We combine this directly on the abstract vector shuffle nodes so it is
23735 /// easier to generically match. We also insert dummy vector shuffle nodes for
23736 /// the operands which explicitly discard the lanes which are unused by this
23737 /// operation to try to flow through the rest of the combiner the fact that
23738 /// they're unused.
23739 static SDValue combineShuffleToAddSub(SDNode *N, const X86Subtarget *Subtarget,
23740 SelectionDAG &DAG) {
23742 EVT VT = N->getValueType(0);
23743 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
23744 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
23747 // We only handle target-independent shuffles.
23748 // FIXME: It would be easy and harmless to use the target shuffle mask
23749 // extraction tool to support more.
23750 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23753 auto *SVN = cast<ShuffleVectorSDNode>(N);
23754 SmallVector<int, 8> Mask;
23755 for (int M : SVN->getMask())
23758 SDValue V1 = N->getOperand(0);
23759 SDValue V2 = N->getOperand(1);
23761 // We require the first shuffle operand to be the FSUB node, and the second to
23762 // be the FADD node.
23763 if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
23764 ShuffleVectorSDNode::commuteMask(Mask);
23766 } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
23769 // If there are other uses of these operations we can't fold them.
23770 if (!V1->hasOneUse() || !V2->hasOneUse())
23773 // Ensure that both operations have the same operands. Note that we can
23774 // commute the FADD operands.
23775 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23776 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23777 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23780 // We're looking for blends between FADD and FSUB nodes. We insist on these
23781 // nodes being lined up in a specific expected pattern.
23782 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23783 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23784 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23787 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23790 /// PerformShuffleCombine - Performs several different shuffle combines.
23791 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23792 TargetLowering::DAGCombinerInfo &DCI,
23793 const X86Subtarget *Subtarget) {
23795 SDValue N0 = N->getOperand(0);
23796 SDValue N1 = N->getOperand(1);
23797 EVT VT = N->getValueType(0);
23799 // Don't create instructions with illegal types after legalize types has run.
23800 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23801 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23804 // If we have legalized the vector types, look for blends of FADD and FSUB
23805 // nodes that we can fuse into an ADDSUB node.
23806 if (TLI.isTypeLegal(VT))
23807 if (SDValue AddSub = combineShuffleToAddSub(N, Subtarget, DAG))
23810 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23811 if (TLI.isTypeLegal(VT) && Subtarget->hasFp256() && VT.is256BitVector() &&
23812 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23813 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23815 // During Type Legalization, when promoting illegal vector types,
23816 // the backend might introduce new shuffle dag nodes and bitcasts.
23818 // This code performs the following transformation:
23819 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23820 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23822 // We do this only if both the bitcast and the BINOP dag nodes have
23823 // one use. Also, perform this transformation only if the new binary
23824 // operation is legal. This is to avoid introducing dag nodes that
23825 // potentially need to be further expanded (or custom lowered) into a
23826 // less optimal sequence of dag nodes.
23827 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23828 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23829 N0.getOpcode() == ISD::BITCAST) {
23830 SDValue BC0 = N0.getOperand(0);
23831 EVT SVT = BC0.getValueType();
23832 unsigned Opcode = BC0.getOpcode();
23833 unsigned NumElts = VT.getVectorNumElements();
23835 if (BC0.hasOneUse() && SVT.isVector() &&
23836 SVT.getVectorNumElements() * 2 == NumElts &&
23837 TLI.isOperationLegal(Opcode, VT)) {
23838 bool CanFold = false;
23850 unsigned SVTNumElts = SVT.getVectorNumElements();
23851 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23852 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23853 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23854 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23855 CanFold = SVOp->getMaskElt(i) < 0;
23858 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23859 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23860 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23861 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23866 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23867 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23868 // consecutive, non-overlapping, and in the right order.
23869 SmallVector<SDValue, 16> Elts;
23870 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23871 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23873 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23876 if (isTargetShuffle(N->getOpcode())) {
23878 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23879 if (Shuffle.getNode())
23882 // Try recursively combining arbitrary sequences of x86 shuffle
23883 // instructions into higher-order shuffles. We do this after combining
23884 // specific PSHUF instruction sequences into their minimal form so that we
23885 // can evaluate how many specialized shuffle instructions are involved in
23886 // a particular chain.
23887 SmallVector<int, 1> NonceMask; // Just a placeholder.
23888 NonceMask.push_back(0);
23889 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23890 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23892 return SDValue(); // This routine will use CombineTo to replace N.
23898 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23899 /// specific shuffle of a load can be folded into a single element load.
23900 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23901 /// shuffles have been custom lowered so we need to handle those here.
23902 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23903 TargetLowering::DAGCombinerInfo &DCI) {
23904 if (DCI.isBeforeLegalizeOps())
23907 SDValue InVec = N->getOperand(0);
23908 SDValue EltNo = N->getOperand(1);
23909 EVT EltVT = N->getValueType(0);
23911 if (!isa<ConstantSDNode>(EltNo))
23914 EVT OriginalVT = InVec.getValueType();
23916 if (InVec.getOpcode() == ISD::BITCAST) {
23917 // Don't duplicate a load with other uses.
23918 if (!InVec.hasOneUse())
23920 EVT BCVT = InVec.getOperand(0).getValueType();
23921 if (!BCVT.isVector() ||
23922 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23924 InVec = InVec.getOperand(0);
23927 EVT CurrentVT = InVec.getValueType();
23929 if (!isTargetShuffle(InVec.getOpcode()))
23932 // Don't duplicate a load with other uses.
23933 if (!InVec.hasOneUse())
23936 SmallVector<int, 16> ShuffleMask;
23938 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(), true,
23939 ShuffleMask, UnaryShuffle))
23942 // Select the input vector, guarding against out of range extract vector.
23943 unsigned NumElems = CurrentVT.getVectorNumElements();
23944 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23945 int Idx = (Elt > (int)NumElems) ? SM_SentinelUndef : ShuffleMask[Elt];
23947 if (Idx == SM_SentinelZero)
23948 return EltVT.isInteger() ? DAG.getConstant(0, SDLoc(N), EltVT)
23949 : DAG.getConstantFP(+0.0, SDLoc(N), EltVT);
23950 if (Idx == SM_SentinelUndef)
23951 return DAG.getUNDEF(EltVT);
23953 assert(0 <= Idx && Idx < (int)(2 * NumElems) && "Shuffle index out of range");
23954 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23955 : InVec.getOperand(1);
23957 // If inputs to shuffle are the same for both ops, then allow 2 uses
23958 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23959 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23961 if (LdNode.getOpcode() == ISD::BITCAST) {
23962 // Don't duplicate a load with other uses.
23963 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23966 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23967 LdNode = LdNode.getOperand(0);
23970 if (!ISD::isNormalLoad(LdNode.getNode()))
23973 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23975 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23978 // If there's a bitcast before the shuffle, check if the load type and
23979 // alignment is valid.
23980 unsigned Align = LN0->getAlignment();
23981 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23982 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23983 EltVT.getTypeForEVT(*DAG.getContext()));
23985 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23988 // All checks match so transform back to vector_shuffle so that DAG combiner
23989 // can finish the job
23992 // Create shuffle node taking into account the case that its a unary shuffle
23993 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23994 : InVec.getOperand(1);
23995 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23996 InVec.getOperand(0), Shuffle,
23998 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23999 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
24003 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
24004 const X86Subtarget *Subtarget) {
24005 SDValue N0 = N->getOperand(0);
24006 EVT VT = N->getValueType(0);
24008 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
24009 // special and don't usually play with other vector types, it's better to
24010 // handle them early to be sure we emit efficient code by avoiding
24011 // store-load conversions.
24012 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
24013 N0.getValueType() == MVT::v2i32 &&
24014 isNullConstant(N0.getOperand(1))) {
24015 SDValue N00 = N0->getOperand(0);
24016 if (N00.getValueType() == MVT::i32)
24017 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
24020 // Convert a bitcasted integer logic operation that has one bitcasted
24021 // floating-point operand and one constant operand into a floating-point
24022 // logic operation. This may create a load of the constant, but that is
24023 // cheaper than materializing the constant in an integer register and
24024 // transferring it to an SSE register or transferring the SSE operand to
24025 // integer register and back.
24027 switch (N0.getOpcode()) {
24028 case ISD::AND: FPOpcode = X86ISD::FAND; break;
24029 case ISD::OR: FPOpcode = X86ISD::FOR; break;
24030 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
24031 default: return SDValue();
24033 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
24034 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
24035 isa<ConstantSDNode>(N0.getOperand(1)) &&
24036 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
24037 N0.getOperand(0).getOperand(0).getValueType() == VT) {
24038 SDValue N000 = N0.getOperand(0).getOperand(0);
24039 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
24040 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
24046 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
24047 /// generation and convert it from being a bunch of shuffles and extracts
24048 /// into a somewhat faster sequence. For i686, the best sequence is apparently
24049 /// storing the value and loading scalars back, while for x64 we should
24050 /// use 64-bit extracts and shifts.
24051 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
24052 TargetLowering::DAGCombinerInfo &DCI) {
24053 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
24056 SDValue InputVector = N->getOperand(0);
24057 SDLoc dl(InputVector);
24058 // Detect mmx to i32 conversion through a v2i32 elt extract.
24059 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
24060 N->getValueType(0) == MVT::i32 &&
24061 InputVector.getValueType() == MVT::v2i32) {
24063 // The bitcast source is a direct mmx result.
24064 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
24065 if (MMXSrc.getValueType() == MVT::x86mmx)
24066 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24067 N->getValueType(0),
24068 InputVector.getNode()->getOperand(0));
24070 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
24071 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
24072 MMXSrc.getValueType() == MVT::i64) {
24073 SDValue MMXSrcOp = MMXSrc.getOperand(0);
24074 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
24075 MMXSrcOp.getValueType() == MVT::v1i64 &&
24076 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
24077 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24078 N->getValueType(0), MMXSrcOp.getOperand(0));
24082 EVT VT = N->getValueType(0);
24084 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
24085 InputVector.getOpcode() == ISD::BITCAST &&
24086 isa<ConstantSDNode>(InputVector.getOperand(0))) {
24087 uint64_t ExtractedElt =
24088 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
24089 uint64_t InputValue =
24090 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
24091 uint64_t Res = (InputValue >> ExtractedElt) & 1;
24092 return DAG.getConstant(Res, dl, MVT::i1);
24094 // Only operate on vectors of 4 elements, where the alternative shuffling
24095 // gets to be more expensive.
24096 if (InputVector.getValueType() != MVT::v4i32)
24099 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
24100 // single use which is a sign-extend or zero-extend, and all elements are
24102 SmallVector<SDNode *, 4> Uses;
24103 unsigned ExtractedElements = 0;
24104 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
24105 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
24106 if (UI.getUse().getResNo() != InputVector.getResNo())
24109 SDNode *Extract = *UI;
24110 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
24113 if (Extract->getValueType(0) != MVT::i32)
24115 if (!Extract->hasOneUse())
24117 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
24118 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
24120 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
24123 // Record which element was extracted.
24124 ExtractedElements |=
24125 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
24127 Uses.push_back(Extract);
24130 // If not all the elements were used, this may not be worthwhile.
24131 if (ExtractedElements != 15)
24134 // Ok, we've now decided to do the transformation.
24135 // If 64-bit shifts are legal, use the extract-shift sequence,
24136 // otherwise bounce the vector off the cache.
24137 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24140 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
24141 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
24142 auto &DL = DAG.getDataLayout();
24143 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
24144 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24145 DAG.getConstant(0, dl, VecIdxTy));
24146 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24147 DAG.getConstant(1, dl, VecIdxTy));
24149 SDValue ShAmt = DAG.getConstant(
24150 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
24151 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
24152 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24153 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
24154 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
24155 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24156 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
24158 // Store the value to a temporary stack slot.
24159 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
24160 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
24161 MachinePointerInfo(), false, false, 0);
24163 EVT ElementType = InputVector.getValueType().getVectorElementType();
24164 unsigned EltSize = ElementType.getSizeInBits() / 8;
24166 // Replace each use (extract) with a load of the appropriate element.
24167 for (unsigned i = 0; i < 4; ++i) {
24168 uint64_t Offset = EltSize * i;
24169 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
24170 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
24172 SDValue ScalarAddr =
24173 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
24175 // Load the scalar.
24176 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
24177 ScalarAddr, MachinePointerInfo(),
24178 false, false, false, 0);
24183 // Replace the extracts
24184 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
24185 UE = Uses.end(); UI != UE; ++UI) {
24186 SDNode *Extract = *UI;
24188 SDValue Idx = Extract->getOperand(1);
24189 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
24190 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
24193 // The replacement was made in place; don't return anything.
24198 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
24199 const X86Subtarget *Subtarget) {
24201 SDValue Cond = N->getOperand(0);
24202 SDValue LHS = N->getOperand(1);
24203 SDValue RHS = N->getOperand(2);
24205 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
24206 SDValue CondSrc = Cond->getOperand(0);
24207 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
24208 Cond = CondSrc->getOperand(0);
24211 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
24214 // A vselect where all conditions and data are constants can be optimized into
24215 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
24216 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
24217 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
24220 unsigned MaskValue = 0;
24221 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
24224 MVT VT = N->getSimpleValueType(0);
24225 unsigned NumElems = VT.getVectorNumElements();
24226 SmallVector<int, 8> ShuffleMask(NumElems, -1);
24227 for (unsigned i = 0; i < NumElems; ++i) {
24228 // Be sure we emit undef where we can.
24229 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
24230 ShuffleMask[i] = -1;
24232 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
24235 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24236 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
24238 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
24241 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
24243 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
24244 TargetLowering::DAGCombinerInfo &DCI,
24245 const X86Subtarget *Subtarget) {
24247 SDValue Cond = N->getOperand(0);
24248 // Get the LHS/RHS of the select.
24249 SDValue LHS = N->getOperand(1);
24250 SDValue RHS = N->getOperand(2);
24251 EVT VT = LHS.getValueType();
24252 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24254 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
24255 // instructions match the semantics of the common C idiom x<y?x:y but not
24256 // x<=y?x:y, because of how they handle negative zero (which can be
24257 // ignored in unsafe-math mode).
24258 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
24259 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
24260 VT != MVT::f80 && VT != MVT::f128 &&
24261 (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
24262 (Subtarget->hasSSE2() ||
24263 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
24264 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24266 unsigned Opcode = 0;
24267 // Check for x CC y ? x : y.
24268 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24269 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24273 // Converting this to a min would handle NaNs incorrectly, and swapping
24274 // the operands would cause it to handle comparisons between positive
24275 // and negative zero incorrectly.
24276 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24277 if (!DAG.getTarget().Options.UnsafeFPMath &&
24278 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24280 std::swap(LHS, RHS);
24282 Opcode = X86ISD::FMIN;
24285 // Converting this to a min would handle comparisons between positive
24286 // and negative zero incorrectly.
24287 if (!DAG.getTarget().Options.UnsafeFPMath &&
24288 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24290 Opcode = X86ISD::FMIN;
24293 // Converting this to a min would handle both negative zeros and NaNs
24294 // incorrectly, but we can swap the operands to fix both.
24295 std::swap(LHS, RHS);
24299 Opcode = X86ISD::FMIN;
24303 // Converting this to a max would handle comparisons between positive
24304 // and negative zero incorrectly.
24305 if (!DAG.getTarget().Options.UnsafeFPMath &&
24306 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24308 Opcode = X86ISD::FMAX;
24311 // Converting this to a max would handle NaNs incorrectly, and swapping
24312 // the operands would cause it to handle comparisons between positive
24313 // and negative zero incorrectly.
24314 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24315 if (!DAG.getTarget().Options.UnsafeFPMath &&
24316 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24318 std::swap(LHS, RHS);
24320 Opcode = X86ISD::FMAX;
24323 // Converting this to a max would handle both negative zeros and NaNs
24324 // incorrectly, but we can swap the operands to fix both.
24325 std::swap(LHS, RHS);
24329 Opcode = X86ISD::FMAX;
24332 // Check for x CC y ? y : x -- a min/max with reversed arms.
24333 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
24334 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
24338 // Converting this to a min would handle comparisons between positive
24339 // and negative zero incorrectly, and swapping the operands would
24340 // cause it to handle NaNs incorrectly.
24341 if (!DAG.getTarget().Options.UnsafeFPMath &&
24342 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
24343 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24345 std::swap(LHS, RHS);
24347 Opcode = X86ISD::FMIN;
24350 // Converting this to a min would handle NaNs incorrectly.
24351 if (!DAG.getTarget().Options.UnsafeFPMath &&
24352 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
24354 Opcode = X86ISD::FMIN;
24357 // Converting this to a min would handle both negative zeros and NaNs
24358 // incorrectly, but we can swap the operands to fix both.
24359 std::swap(LHS, RHS);
24363 Opcode = X86ISD::FMIN;
24367 // Converting this to a max would handle NaNs incorrectly.
24368 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24370 Opcode = X86ISD::FMAX;
24373 // Converting this to a max would handle comparisons between positive
24374 // and negative zero incorrectly, and swapping the operands would
24375 // cause it to handle NaNs incorrectly.
24376 if (!DAG.getTarget().Options.UnsafeFPMath &&
24377 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
24378 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24380 std::swap(LHS, RHS);
24382 Opcode = X86ISD::FMAX;
24385 // Converting this to a max would handle both negative zeros and NaNs
24386 // incorrectly, but we can swap the operands to fix both.
24387 std::swap(LHS, RHS);
24391 Opcode = X86ISD::FMAX;
24397 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
24400 EVT CondVT = Cond.getValueType();
24401 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
24402 CondVT.getVectorElementType() == MVT::i1) {
24403 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
24404 // lowering on KNL. In this case we convert it to
24405 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
24406 // The same situation for all 128 and 256-bit vectors of i8 and i16.
24407 // Since SKX these selects have a proper lowering.
24408 EVT OpVT = LHS.getValueType();
24409 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
24410 (OpVT.getVectorElementType() == MVT::i8 ||
24411 OpVT.getVectorElementType() == MVT::i16) &&
24412 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
24413 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
24414 DCI.AddToWorklist(Cond.getNode());
24415 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
24418 // If this is a select between two integer constants, try to do some
24420 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
24421 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
24422 // Don't do this for crazy integer types.
24423 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
24424 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
24425 // so that TrueC (the true value) is larger than FalseC.
24426 bool NeedsCondInvert = false;
24428 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
24429 // Efficiently invertible.
24430 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
24431 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
24432 isa<ConstantSDNode>(Cond.getOperand(1))))) {
24433 NeedsCondInvert = true;
24434 std::swap(TrueC, FalseC);
24437 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
24438 if (FalseC->getAPIntValue() == 0 &&
24439 TrueC->getAPIntValue().isPowerOf2()) {
24440 if (NeedsCondInvert) // Invert the condition if needed.
24441 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24442 DAG.getConstant(1, DL, Cond.getValueType()));
24444 // Zero extend the condition if needed.
24445 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
24447 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24448 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
24449 DAG.getConstant(ShAmt, DL, MVT::i8));
24452 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
24453 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24454 if (NeedsCondInvert) // Invert the condition if needed.
24455 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24456 DAG.getConstant(1, DL, Cond.getValueType()));
24458 // Zero extend the condition if needed.
24459 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24460 FalseC->getValueType(0), Cond);
24461 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24462 SDValue(FalseC, 0));
24465 // Optimize cases that will turn into an LEA instruction. This requires
24466 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24467 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24468 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24469 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24471 bool isFastMultiplier = false;
24473 switch ((unsigned char)Diff) {
24475 case 1: // result = add base, cond
24476 case 2: // result = lea base( , cond*2)
24477 case 3: // result = lea base(cond, cond*2)
24478 case 4: // result = lea base( , cond*4)
24479 case 5: // result = lea base(cond, cond*4)
24480 case 8: // result = lea base( , cond*8)
24481 case 9: // result = lea base(cond, cond*8)
24482 isFastMultiplier = true;
24487 if (isFastMultiplier) {
24488 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24489 if (NeedsCondInvert) // Invert the condition if needed.
24490 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24491 DAG.getConstant(1, DL, Cond.getValueType()));
24493 // Zero extend the condition if needed.
24494 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24496 // Scale the condition by the difference.
24498 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24499 DAG.getConstant(Diff, DL,
24500 Cond.getValueType()));
24502 // Add the base if non-zero.
24503 if (FalseC->getAPIntValue() != 0)
24504 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24505 SDValue(FalseC, 0));
24512 // Canonicalize max and min:
24513 // (x > y) ? x : y -> (x >= y) ? x : y
24514 // (x < y) ? x : y -> (x <= y) ? x : y
24515 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
24516 // the need for an extra compare
24517 // against zero. e.g.
24518 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
24520 // testl %edi, %edi
24522 // cmovgl %edi, %eax
24526 // cmovsl %eax, %edi
24527 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
24528 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24529 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24530 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24535 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
24536 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
24537 Cond.getOperand(0), Cond.getOperand(1), NewCC);
24538 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
24543 // Early exit check
24544 if (!TLI.isTypeLegal(VT))
24547 // Match VSELECTs into subs with unsigned saturation.
24548 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
24549 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
24550 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
24551 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
24552 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24554 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
24555 // left side invert the predicate to simplify logic below.
24557 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
24559 CC = ISD::getSetCCInverse(CC, true);
24560 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
24564 if (Other.getNode() && Other->getNumOperands() == 2 &&
24565 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
24566 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
24567 SDValue CondRHS = Cond->getOperand(1);
24569 // Look for a general sub with unsigned saturation first.
24570 // x >= y ? x-y : 0 --> subus x, y
24571 // x > y ? x-y : 0 --> subus x, y
24572 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
24573 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
24574 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
24576 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
24577 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
24578 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
24579 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
24580 // If the RHS is a constant we have to reverse the const
24581 // canonicalization.
24582 // x > C-1 ? x+-C : 0 --> subus x, C
24583 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
24584 CondRHSConst->getAPIntValue() ==
24585 (-OpRHSConst->getAPIntValue() - 1))
24586 return DAG.getNode(
24587 X86ISD::SUBUS, DL, VT, OpLHS,
24588 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
24590 // Another special case: If C was a sign bit, the sub has been
24591 // canonicalized into a xor.
24592 // FIXME: Would it be better to use computeKnownBits to determine
24593 // whether it's safe to decanonicalize the xor?
24594 // x s< 0 ? x^C : 0 --> subus x, C
24595 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
24596 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
24597 OpRHSConst->getAPIntValue().isSignBit())
24598 // Note that we have to rebuild the RHS constant here to ensure we
24599 // don't rely on particular values of undef lanes.
24600 return DAG.getNode(
24601 X86ISD::SUBUS, DL, VT, OpLHS,
24602 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
24607 // Simplify vector selection if condition value type matches vselect
24609 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
24610 assert(Cond.getValueType().isVector() &&
24611 "vector select expects a vector selector!");
24613 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
24614 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
24616 // Try invert the condition if true value is not all 1s and false value
24618 if (!TValIsAllOnes && !FValIsAllZeros &&
24619 // Check if the selector will be produced by CMPP*/PCMP*
24620 Cond.getOpcode() == ISD::SETCC &&
24621 // Check if SETCC has already been promoted
24622 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
24624 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
24625 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
24627 if (TValIsAllZeros || FValIsAllOnes) {
24628 SDValue CC = Cond.getOperand(2);
24629 ISD::CondCode NewCC =
24630 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
24631 Cond.getOperand(0).getValueType().isInteger());
24632 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
24633 std::swap(LHS, RHS);
24634 TValIsAllOnes = FValIsAllOnes;
24635 FValIsAllZeros = TValIsAllZeros;
24639 if (TValIsAllOnes || FValIsAllZeros) {
24642 if (TValIsAllOnes && FValIsAllZeros)
24644 else if (TValIsAllOnes)
24646 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
24647 else if (FValIsAllZeros)
24648 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24649 DAG.getBitcast(CondVT, LHS));
24651 return DAG.getBitcast(VT, Ret);
24655 // We should generate an X86ISD::BLENDI from a vselect if its argument
24656 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24657 // constants. This specific pattern gets generated when we split a
24658 // selector for a 512 bit vector in a machine without AVX512 (but with
24659 // 256-bit vectors), during legalization:
24661 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24663 // Iff we find this pattern and the build_vectors are built from
24664 // constants, we translate the vselect into a shuffle_vector that we
24665 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24666 if ((N->getOpcode() == ISD::VSELECT ||
24667 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24668 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
24669 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24670 if (Shuffle.getNode())
24674 // If this is a *dynamic* select (non-constant condition) and we can match
24675 // this node with one of the variable blend instructions, restructure the
24676 // condition so that the blends can use the high bit of each element and use
24677 // SimplifyDemandedBits to simplify the condition operand.
24678 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24679 !DCI.isBeforeLegalize() &&
24680 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24681 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
24683 // Don't optimize vector selects that map to mask-registers.
24687 // We can only handle the cases where VSELECT is directly legal on the
24688 // subtarget. We custom lower VSELECT nodes with constant conditions and
24689 // this makes it hard to see whether a dynamic VSELECT will correctly
24690 // lower, so we both check the operation's status and explicitly handle the
24691 // cases where a *dynamic* blend will fail even though a constant-condition
24692 // blend could be custom lowered.
24693 // FIXME: We should find a better way to handle this class of problems.
24694 // Potentially, we should combine constant-condition vselect nodes
24695 // pre-legalization into shuffles and not mark as many types as custom
24697 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
24699 // FIXME: We don't support i16-element blends currently. We could and
24700 // should support them by making *all* the bits in the condition be set
24701 // rather than just the high bit and using an i8-element blend.
24702 if (VT.getVectorElementType() == MVT::i16)
24704 // Dynamic blending was only available from SSE4.1 onward.
24705 if (VT.is128BitVector() && !Subtarget->hasSSE41())
24707 // Byte blends are only available in AVX2
24708 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24711 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24712 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24714 APInt KnownZero, KnownOne;
24715 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24716 DCI.isBeforeLegalizeOps());
24717 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24718 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24720 // If we changed the computation somewhere in the DAG, this change
24721 // will affect all users of Cond.
24722 // Make sure it is fine and update all the nodes so that we do not
24723 // use the generic VSELECT anymore. Otherwise, we may perform
24724 // wrong optimizations as we messed up with the actual expectation
24725 // for the vector boolean values.
24726 if (Cond != TLO.Old) {
24727 // Check all uses of that condition operand to check whether it will be
24728 // consumed by non-BLEND instructions, which may depend on all bits are
24730 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24732 if (I->getOpcode() != ISD::VSELECT)
24733 // TODO: Add other opcodes eventually lowered into BLEND.
24736 // Update all the users of the condition, before committing the change,
24737 // so that the VSELECT optimizations that expect the correct vector
24738 // boolean value will not be triggered.
24739 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24741 DAG.ReplaceAllUsesOfValueWith(
24743 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24744 Cond, I->getOperand(1), I->getOperand(2)));
24745 DCI.CommitTargetLoweringOpt(TLO);
24748 // At this point, only Cond is changed. Change the condition
24749 // just for N to keep the opportunity to optimize all other
24750 // users their own way.
24751 DAG.ReplaceAllUsesOfValueWith(
24753 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24754 TLO.New, N->getOperand(1), N->getOperand(2)));
24762 // Check whether a boolean test is testing a boolean value generated by
24763 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24766 // Simplify the following patterns:
24767 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24768 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24769 // to (Op EFLAGS Cond)
24771 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24772 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24773 // to (Op EFLAGS !Cond)
24775 // where Op could be BRCOND or CMOV.
24777 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24778 // Quit if not CMP and SUB with its value result used.
24779 if (Cmp.getOpcode() != X86ISD::CMP &&
24780 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24783 // Quit if not used as a boolean value.
24784 if (CC != X86::COND_E && CC != X86::COND_NE)
24787 // Check CMP operands. One of them should be 0 or 1 and the other should be
24788 // an SetCC or extended from it.
24789 SDValue Op1 = Cmp.getOperand(0);
24790 SDValue Op2 = Cmp.getOperand(1);
24793 const ConstantSDNode* C = nullptr;
24794 bool needOppositeCond = (CC == X86::COND_E);
24795 bool checkAgainstTrue = false; // Is it a comparison against 1?
24797 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24799 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24801 else // Quit if all operands are not constants.
24804 if (C->getZExtValue() == 1) {
24805 needOppositeCond = !needOppositeCond;
24806 checkAgainstTrue = true;
24807 } else if (C->getZExtValue() != 0)
24808 // Quit if the constant is neither 0 or 1.
24811 bool truncatedToBoolWithAnd = false;
24812 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24813 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24814 SetCC.getOpcode() == ISD::TRUNCATE ||
24815 SetCC.getOpcode() == ISD::AND) {
24816 if (SetCC.getOpcode() == ISD::AND) {
24818 if (isOneConstant(SetCC.getOperand(0)))
24820 if (isOneConstant(SetCC.getOperand(1)))
24824 SetCC = SetCC.getOperand(OpIdx);
24825 truncatedToBoolWithAnd = true;
24827 SetCC = SetCC.getOperand(0);
24830 switch (SetCC.getOpcode()) {
24831 case X86ISD::SETCC_CARRY:
24832 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24833 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24834 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24835 // truncated to i1 using 'and'.
24836 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24838 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24839 "Invalid use of SETCC_CARRY!");
24841 case X86ISD::SETCC:
24842 // Set the condition code or opposite one if necessary.
24843 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24844 if (needOppositeCond)
24845 CC = X86::GetOppositeBranchCondition(CC);
24846 return SetCC.getOperand(1);
24847 case X86ISD::CMOV: {
24848 // Check whether false/true value has canonical one, i.e. 0 or 1.
24849 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24850 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24851 // Quit if true value is not a constant.
24854 // Quit if false value is not a constant.
24856 SDValue Op = SetCC.getOperand(0);
24857 // Skip 'zext' or 'trunc' node.
24858 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24859 Op.getOpcode() == ISD::TRUNCATE)
24860 Op = Op.getOperand(0);
24861 // A special case for rdrand/rdseed, where 0 is set if false cond is
24863 if ((Op.getOpcode() != X86ISD::RDRAND &&
24864 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24867 // Quit if false value is not the constant 0 or 1.
24868 bool FValIsFalse = true;
24869 if (FVal && FVal->getZExtValue() != 0) {
24870 if (FVal->getZExtValue() != 1)
24872 // If FVal is 1, opposite cond is needed.
24873 needOppositeCond = !needOppositeCond;
24874 FValIsFalse = false;
24876 // Quit if TVal is not the constant opposite of FVal.
24877 if (FValIsFalse && TVal->getZExtValue() != 1)
24879 if (!FValIsFalse && TVal->getZExtValue() != 0)
24881 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24882 if (needOppositeCond)
24883 CC = X86::GetOppositeBranchCondition(CC);
24884 return SetCC.getOperand(3);
24891 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24893 /// (X86or (X86setcc) (X86setcc))
24894 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24895 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24896 X86::CondCode &CC1, SDValue &Flags,
24898 if (Cond->getOpcode() == X86ISD::CMP) {
24899 if (!isNullConstant(Cond->getOperand(1)))
24902 Cond = Cond->getOperand(0);
24907 SDValue SetCC0, SetCC1;
24908 switch (Cond->getOpcode()) {
24909 default: return false;
24916 SetCC0 = Cond->getOperand(0);
24917 SetCC1 = Cond->getOperand(1);
24921 // Make sure we have SETCC nodes, using the same flags value.
24922 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24923 SetCC1.getOpcode() != X86ISD::SETCC ||
24924 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24927 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24928 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24929 Flags = SetCC0->getOperand(1);
24933 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24934 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24935 TargetLowering::DAGCombinerInfo &DCI,
24936 const X86Subtarget *Subtarget) {
24939 // If the flag operand isn't dead, don't touch this CMOV.
24940 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24943 SDValue FalseOp = N->getOperand(0);
24944 SDValue TrueOp = N->getOperand(1);
24945 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24946 SDValue Cond = N->getOperand(3);
24948 if (CC == X86::COND_E || CC == X86::COND_NE) {
24949 switch (Cond.getOpcode()) {
24953 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24954 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24955 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24961 Flags = checkBoolTestSetCCCombine(Cond, CC);
24962 if (Flags.getNode() &&
24963 // Extra check as FCMOV only supports a subset of X86 cond.
24964 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24965 SDValue Ops[] = { FalseOp, TrueOp,
24966 DAG.getConstant(CC, DL, MVT::i8), Flags };
24967 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24970 // If this is a select between two integer constants, try to do some
24971 // optimizations. Note that the operands are ordered the opposite of SELECT
24973 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24974 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24975 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24976 // larger than FalseC (the false value).
24977 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24978 CC = X86::GetOppositeBranchCondition(CC);
24979 std::swap(TrueC, FalseC);
24980 std::swap(TrueOp, FalseOp);
24983 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24984 // This is efficient for any integer data type (including i8/i16) and
24986 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24987 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24988 DAG.getConstant(CC, DL, MVT::i8), Cond);
24990 // Zero extend the condition if needed.
24991 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24993 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24994 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24995 DAG.getConstant(ShAmt, DL, MVT::i8));
24996 if (N->getNumValues() == 2) // Dead flag value?
24997 return DCI.CombineTo(N, Cond, SDValue());
25001 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
25002 // for any integer data type, including i8/i16.
25003 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
25004 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
25005 DAG.getConstant(CC, DL, MVT::i8), Cond);
25007 // Zero extend the condition if needed.
25008 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
25009 FalseC->getValueType(0), Cond);
25010 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
25011 SDValue(FalseC, 0));
25013 if (N->getNumValues() == 2) // Dead flag value?
25014 return DCI.CombineTo(N, Cond, SDValue());
25018 // Optimize cases that will turn into an LEA instruction. This requires
25019 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
25020 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
25021 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
25022 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
25024 bool isFastMultiplier = false;
25026 switch ((unsigned char)Diff) {
25028 case 1: // result = add base, cond
25029 case 2: // result = lea base( , cond*2)
25030 case 3: // result = lea base(cond, cond*2)
25031 case 4: // result = lea base( , cond*4)
25032 case 5: // result = lea base(cond, cond*4)
25033 case 8: // result = lea base( , cond*8)
25034 case 9: // result = lea base(cond, cond*8)
25035 isFastMultiplier = true;
25040 if (isFastMultiplier) {
25041 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
25042 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
25043 DAG.getConstant(CC, DL, MVT::i8), Cond);
25044 // Zero extend the condition if needed.
25045 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
25047 // Scale the condition by the difference.
25049 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
25050 DAG.getConstant(Diff, DL, Cond.getValueType()));
25052 // Add the base if non-zero.
25053 if (FalseC->getAPIntValue() != 0)
25054 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
25055 SDValue(FalseC, 0));
25056 if (N->getNumValues() == 2) // Dead flag value?
25057 return DCI.CombineTo(N, Cond, SDValue());
25064 // Handle these cases:
25065 // (select (x != c), e, c) -> select (x != c), e, x),
25066 // (select (x == c), c, e) -> select (x == c), x, e)
25067 // where the c is an integer constant, and the "select" is the combination
25068 // of CMOV and CMP.
25070 // The rationale for this change is that the conditional-move from a constant
25071 // needs two instructions, however, conditional-move from a register needs
25072 // only one instruction.
25074 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
25075 // some instruction-combining opportunities. This opt needs to be
25076 // postponed as late as possible.
25078 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
25079 // the DCI.xxxx conditions are provided to postpone the optimization as
25080 // late as possible.
25082 ConstantSDNode *CmpAgainst = nullptr;
25083 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
25084 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
25085 !isa<ConstantSDNode>(Cond.getOperand(0))) {
25087 if (CC == X86::COND_NE &&
25088 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
25089 CC = X86::GetOppositeBranchCondition(CC);
25090 std::swap(TrueOp, FalseOp);
25093 if (CC == X86::COND_E &&
25094 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
25095 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
25096 DAG.getConstant(CC, DL, MVT::i8), Cond };
25097 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
25102 // Fold and/or of setcc's to double CMOV:
25103 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
25104 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
25106 // This combine lets us generate:
25107 // cmovcc1 (jcc1 if we don't have CMOV)
25113 // cmovne (jne if we don't have CMOV)
25114 // When we can't use the CMOV instruction, it might increase branch
25116 // When we can use CMOV, or when there is no mispredict, this improves
25117 // throughput and reduces register pressure.
25119 if (CC == X86::COND_NE) {
25121 X86::CondCode CC0, CC1;
25123 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
25125 std::swap(FalseOp, TrueOp);
25126 CC0 = X86::GetOppositeBranchCondition(CC0);
25127 CC1 = X86::GetOppositeBranchCondition(CC1);
25130 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
25132 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
25133 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
25134 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
25135 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
25143 /// PerformMulCombine - Optimize a single multiply with constant into two
25144 /// in order to implement it with two cheaper instructions, e.g.
25145 /// LEA + SHL, LEA + LEA.
25146 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
25147 TargetLowering::DAGCombinerInfo &DCI) {
25148 // An imul is usually smaller than the alternative sequence.
25149 if (DAG.getMachineFunction().getFunction()->optForMinSize())
25152 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
25155 EVT VT = N->getValueType(0);
25156 if (VT != MVT::i64 && VT != MVT::i32)
25159 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
25162 uint64_t MulAmt = C->getZExtValue();
25163 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
25166 uint64_t MulAmt1 = 0;
25167 uint64_t MulAmt2 = 0;
25168 if ((MulAmt % 9) == 0) {
25170 MulAmt2 = MulAmt / 9;
25171 } else if ((MulAmt % 5) == 0) {
25173 MulAmt2 = MulAmt / 5;
25174 } else if ((MulAmt % 3) == 0) {
25176 MulAmt2 = MulAmt / 3;
25182 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
25184 if (isPowerOf2_64(MulAmt2) &&
25185 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
25186 // If second multiplifer is pow2, issue it first. We want the multiply by
25187 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
25189 std::swap(MulAmt1, MulAmt2);
25191 if (isPowerOf2_64(MulAmt1))
25192 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25193 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
25195 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
25196 DAG.getConstant(MulAmt1, DL, VT));
25198 if (isPowerOf2_64(MulAmt2))
25199 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
25200 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
25202 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
25203 DAG.getConstant(MulAmt2, DL, VT));
25207 assert(MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX)
25208 && "Both cases that could cause potential overflows should have "
25209 "already been handled.");
25210 if (isPowerOf2_64(MulAmt - 1))
25211 // (mul x, 2^N + 1) => (add (shl x, N), x)
25212 NewMul = DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
25213 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25214 DAG.getConstant(Log2_64(MulAmt - 1), DL,
25217 else if (isPowerOf2_64(MulAmt + 1))
25218 // (mul x, 2^N - 1) => (sub (shl x, N), x)
25219 NewMul = DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::SHL, DL, VT,
25221 DAG.getConstant(Log2_64(MulAmt + 1),
25222 DL, MVT::i8)), N->getOperand(0));
25226 // Do not add new nodes to DAG combiner worklist.
25227 DCI.CombineTo(N, NewMul, false);
25232 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
25233 SDValue N0 = N->getOperand(0);
25234 SDValue N1 = N->getOperand(1);
25235 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
25236 EVT VT = N0.getValueType();
25238 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
25239 // since the result of setcc_c is all zero's or all ones.
25240 if (VT.isInteger() && !VT.isVector() &&
25241 N1C && N0.getOpcode() == ISD::AND &&
25242 N0.getOperand(1).getOpcode() == ISD::Constant) {
25243 SDValue N00 = N0.getOperand(0);
25244 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
25245 APInt ShAmt = N1C->getAPIntValue();
25246 Mask = Mask.shl(ShAmt);
25247 bool MaskOK = false;
25248 // We can handle cases concerning bit-widening nodes containing setcc_c if
25249 // we carefully interrogate the mask to make sure we are semantics
25251 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
25252 // of the underlying setcc_c operation if the setcc_c was zero extended.
25253 // Consider the following example:
25254 // zext(setcc_c) -> i32 0x0000FFFF
25255 // c1 -> i32 0x0000FFFF
25256 // c2 -> i32 0x00000001
25257 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
25258 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
25259 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25261 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
25262 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25264 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
25265 N00.getOpcode() == ISD::ANY_EXTEND) &&
25266 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25267 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
25269 if (MaskOK && Mask != 0) {
25271 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
25275 // Hardware support for vector shifts is sparse which makes us scalarize the
25276 // vector operations in many cases. Also, on sandybridge ADD is faster than
25278 // (shl V, 1) -> add V,V
25279 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
25280 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
25281 assert(N0.getValueType().isVector() && "Invalid vector shift type");
25282 // We shift all of the values by one. In many cases we do not have
25283 // hardware support for this operation. This is better expressed as an ADD
25285 if (N1SplatC->getAPIntValue() == 1)
25286 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
25292 static SDValue PerformSRACombine(SDNode *N, SelectionDAG &DAG) {
25293 SDValue N0 = N->getOperand(0);
25294 SDValue N1 = N->getOperand(1);
25295 EVT VT = N0.getValueType();
25296 unsigned Size = VT.getSizeInBits();
25298 // fold (ashr (shl, a, [56,48,32,24,16]), SarConst)
25299 // into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or
25300 // into (lshr, (sext (a), SarConst - [56,48,32,24,16]))
25301 // depending on sign of (SarConst - [56,48,32,24,16])
25303 // sexts in X86 are MOVs. The MOVs have the same code size
25304 // as above SHIFTs (only SHIFT on 1 has lower code size).
25305 // However the MOVs have 2 advantages to a SHIFT:
25306 // 1. MOVs can write to a register that differs from source
25307 // 2. MOVs accept memory operands
25309 if (!VT.isInteger() || VT.isVector() || N1.getOpcode() != ISD::Constant ||
25310 N0.getOpcode() != ISD::SHL || !N0.hasOneUse() ||
25311 N0.getOperand(1).getOpcode() != ISD::Constant)
25314 SDValue N00 = N0.getOperand(0);
25315 SDValue N01 = N0.getOperand(1);
25316 APInt ShlConst = (cast<ConstantSDNode>(N01))->getAPIntValue();
25317 APInt SarConst = (cast<ConstantSDNode>(N1))->getAPIntValue();
25318 EVT CVT = N1.getValueType();
25320 if (SarConst.isNegative())
25323 for (MVT SVT : MVT::integer_valuetypes()) {
25324 unsigned ShiftSize = SVT.getSizeInBits();
25325 // skipping types without corresponding sext/zext and
25326 // ShlConst that is not one of [56,48,32,24,16]
25327 if (ShiftSize < 8 || ShiftSize > 64 || ShlConst != Size - ShiftSize)
25331 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, N00, DAG.getValueType(SVT));
25332 SarConst = SarConst - (Size - ShiftSize);
25335 else if (SarConst.isNegative())
25336 return DAG.getNode(ISD::SHL, DL, VT, NN,
25337 DAG.getConstant(-SarConst, DL, CVT));
25339 return DAG.getNode(ISD::SRA, DL, VT, NN,
25340 DAG.getConstant(SarConst, DL, CVT));
25345 /// \brief Returns a vector of 0s if the node in input is a vector logical
25346 /// shift by a constant amount which is known to be bigger than or equal
25347 /// to the vector element size in bits.
25348 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
25349 const X86Subtarget *Subtarget) {
25350 EVT VT = N->getValueType(0);
25352 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
25353 (!Subtarget->hasInt256() ||
25354 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
25357 SDValue Amt = N->getOperand(1);
25359 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
25360 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
25361 APInt ShiftAmt = AmtSplat->getAPIntValue();
25362 unsigned MaxAmount =
25363 VT.getSimpleVT().getVectorElementType().getSizeInBits();
25365 // SSE2/AVX2 logical shifts always return a vector of 0s
25366 // if the shift amount is bigger than or equal to
25367 // the element size. The constant shift amount will be
25368 // encoded as a 8-bit immediate.
25369 if (ShiftAmt.trunc(8).uge(MaxAmount))
25370 return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL);
25376 /// PerformShiftCombine - Combine shifts.
25377 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
25378 TargetLowering::DAGCombinerInfo &DCI,
25379 const X86Subtarget *Subtarget) {
25380 if (N->getOpcode() == ISD::SHL)
25381 if (SDValue V = PerformSHLCombine(N, DAG))
25384 if (N->getOpcode() == ISD::SRA)
25385 if (SDValue V = PerformSRACombine(N, DAG))
25388 // Try to fold this logical shift into a zero vector.
25389 if (N->getOpcode() != ISD::SRA)
25390 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
25396 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
25397 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
25398 // and friends. Likewise for OR -> CMPNEQSS.
25399 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
25400 TargetLowering::DAGCombinerInfo &DCI,
25401 const X86Subtarget *Subtarget) {
25404 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
25405 // we're requiring SSE2 for both.
25406 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
25407 SDValue N0 = N->getOperand(0);
25408 SDValue N1 = N->getOperand(1);
25409 SDValue CMP0 = N0->getOperand(1);
25410 SDValue CMP1 = N1->getOperand(1);
25413 // The SETCCs should both refer to the same CMP.
25414 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
25417 SDValue CMP00 = CMP0->getOperand(0);
25418 SDValue CMP01 = CMP0->getOperand(1);
25419 EVT VT = CMP00.getValueType();
25421 if (VT == MVT::f32 || VT == MVT::f64) {
25422 bool ExpectingFlags = false;
25423 // Check for any users that want flags:
25424 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
25425 !ExpectingFlags && UI != UE; ++UI)
25426 switch (UI->getOpcode()) {
25431 ExpectingFlags = true;
25433 case ISD::CopyToReg:
25434 case ISD::SIGN_EXTEND:
25435 case ISD::ZERO_EXTEND:
25436 case ISD::ANY_EXTEND:
25440 if (!ExpectingFlags) {
25441 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
25442 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
25444 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
25445 X86::CondCode tmp = cc0;
25450 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
25451 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
25452 // FIXME: need symbolic constants for these magic numbers.
25453 // See X86ATTInstPrinter.cpp:printSSECC().
25454 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
25455 if (Subtarget->hasAVX512()) {
25456 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
25458 DAG.getConstant(x86cc, DL, MVT::i8));
25459 if (N->getValueType(0) != MVT::i1)
25460 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
25464 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
25465 CMP00.getValueType(), CMP00, CMP01,
25466 DAG.getConstant(x86cc, DL,
25469 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
25470 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
25472 if (is64BitFP && !Subtarget->is64Bit()) {
25473 // On a 32-bit target, we cannot bitcast the 64-bit float to a
25474 // 64-bit integer, since that's not a legal type. Since
25475 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
25476 // bits, but can do this little dance to extract the lowest 32 bits
25477 // and work with those going forward.
25478 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
25480 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
25481 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
25482 Vector32, DAG.getIntPtrConstant(0, DL));
25486 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
25487 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
25488 DAG.getConstant(1, DL, IntVT));
25489 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
25491 return OneBitOfTruth;
25499 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
25500 /// so it can be folded inside ANDNP.
25501 static bool CanFoldXORWithAllOnes(const SDNode *N) {
25502 EVT VT = N->getValueType(0);
25504 // Match direct AllOnes for 128 and 256-bit vectors
25505 if (ISD::isBuildVectorAllOnes(N))
25508 // Look through a bit convert.
25509 if (N->getOpcode() == ISD::BITCAST)
25510 N = N->getOperand(0).getNode();
25512 // Sometimes the operand may come from a insert_subvector building a 256-bit
25514 if (VT.is256BitVector() &&
25515 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
25516 SDValue V1 = N->getOperand(0);
25517 SDValue V2 = N->getOperand(1);
25519 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
25520 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
25521 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
25522 ISD::isBuildVectorAllOnes(V2.getNode()))
25529 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
25530 // register. In most cases we actually compare or select YMM-sized registers
25531 // and mixing the two types creates horrible code. This method optimizes
25532 // some of the transition sequences.
25533 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
25534 TargetLowering::DAGCombinerInfo &DCI,
25535 const X86Subtarget *Subtarget) {
25536 EVT VT = N->getValueType(0);
25537 if (!VT.is256BitVector())
25540 assert((N->getOpcode() == ISD::ANY_EXTEND ||
25541 N->getOpcode() == ISD::ZERO_EXTEND ||
25542 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
25544 SDValue Narrow = N->getOperand(0);
25545 EVT NarrowVT = Narrow->getValueType(0);
25546 if (!NarrowVT.is128BitVector())
25549 if (Narrow->getOpcode() != ISD::XOR &&
25550 Narrow->getOpcode() != ISD::AND &&
25551 Narrow->getOpcode() != ISD::OR)
25554 SDValue N0 = Narrow->getOperand(0);
25555 SDValue N1 = Narrow->getOperand(1);
25558 // The Left side has to be a trunc.
25559 if (N0.getOpcode() != ISD::TRUNCATE)
25562 // The type of the truncated inputs.
25563 EVT WideVT = N0->getOperand(0)->getValueType(0);
25567 // The right side has to be a 'trunc' or a constant vector.
25568 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
25569 ConstantSDNode *RHSConstSplat = nullptr;
25570 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
25571 RHSConstSplat = RHSBV->getConstantSplatNode();
25572 if (!RHSTrunc && !RHSConstSplat)
25575 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25577 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
25580 // Set N0 and N1 to hold the inputs to the new wide operation.
25581 N0 = N0->getOperand(0);
25582 if (RHSConstSplat) {
25583 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
25584 SDValue(RHSConstSplat, 0));
25585 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
25586 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
25587 } else if (RHSTrunc) {
25588 N1 = N1->getOperand(0);
25591 // Generate the wide operation.
25592 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
25593 unsigned Opcode = N->getOpcode();
25595 case ISD::ANY_EXTEND:
25597 case ISD::ZERO_EXTEND: {
25598 unsigned InBits = NarrowVT.getScalarSizeInBits();
25599 APInt Mask = APInt::getAllOnesValue(InBits);
25600 Mask = Mask.zext(VT.getScalarSizeInBits());
25601 return DAG.getNode(ISD::AND, DL, VT,
25602 Op, DAG.getConstant(Mask, DL, VT));
25604 case ISD::SIGN_EXTEND:
25605 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
25606 Op, DAG.getValueType(NarrowVT));
25608 llvm_unreachable("Unexpected opcode");
25612 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
25613 TargetLowering::DAGCombinerInfo &DCI,
25614 const X86Subtarget *Subtarget) {
25615 SDValue N0 = N->getOperand(0);
25616 SDValue N1 = N->getOperand(1);
25619 // A vector zext_in_reg may be represented as a shuffle,
25620 // feeding into a bitcast (this represents anyext) feeding into
25621 // an and with a mask.
25622 // We'd like to try to combine that into a shuffle with zero
25623 // plus a bitcast, removing the and.
25624 if (N0.getOpcode() != ISD::BITCAST ||
25625 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
25628 // The other side of the AND should be a splat of 2^C, where C
25629 // is the number of bits in the source type.
25630 if (N1.getOpcode() == ISD::BITCAST)
25631 N1 = N1.getOperand(0);
25632 if (N1.getOpcode() != ISD::BUILD_VECTOR)
25634 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
25636 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
25637 EVT SrcType = Shuffle->getValueType(0);
25639 // We expect a single-source shuffle
25640 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
25643 unsigned SrcSize = SrcType.getScalarSizeInBits();
25645 APInt SplatValue, SplatUndef;
25646 unsigned SplatBitSize;
25648 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
25649 SplatBitSize, HasAnyUndefs))
25652 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
25653 // Make sure the splat matches the mask we expect
25654 if (SplatBitSize > ResSize ||
25655 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
25658 // Make sure the input and output size make sense
25659 if (SrcSize >= ResSize || ResSize % SrcSize)
25662 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
25663 // The number of u's between each two values depends on the ratio between
25664 // the source and dest type.
25665 unsigned ZextRatio = ResSize / SrcSize;
25666 bool IsZext = true;
25667 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
25668 if (i % ZextRatio) {
25669 if (Shuffle->getMaskElt(i) > 0) {
25675 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
25676 // Expected element number
25686 // Ok, perform the transformation - replace the shuffle with
25687 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
25688 // (instead of undef) where the k elements come from the zero vector.
25689 SmallVector<int, 8> Mask;
25690 unsigned NumElems = SrcType.getVectorNumElements();
25691 for (unsigned i = 0; i < NumElems; ++i)
25693 Mask.push_back(NumElems);
25695 Mask.push_back(i / ZextRatio);
25697 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
25698 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
25699 return DAG.getBitcast(N0.getValueType(), NewShuffle);
25702 /// If both input operands of a logic op are being cast from floating point
25703 /// types, try to convert this into a floating point logic node to avoid
25704 /// unnecessary moves from SSE to integer registers.
25705 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
25706 const X86Subtarget *Subtarget) {
25707 unsigned FPOpcode = ISD::DELETED_NODE;
25708 if (N->getOpcode() == ISD::AND)
25709 FPOpcode = X86ISD::FAND;
25710 else if (N->getOpcode() == ISD::OR)
25711 FPOpcode = X86ISD::FOR;
25712 else if (N->getOpcode() == ISD::XOR)
25713 FPOpcode = X86ISD::FXOR;
25715 assert(FPOpcode != ISD::DELETED_NODE &&
25716 "Unexpected input node for FP logic conversion");
25718 EVT VT = N->getValueType(0);
25719 SDValue N0 = N->getOperand(0);
25720 SDValue N1 = N->getOperand(1);
25722 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
25723 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
25724 (Subtarget->hasSSE2() && VT == MVT::i64))) {
25725 SDValue N00 = N0.getOperand(0);
25726 SDValue N10 = N1.getOperand(0);
25727 EVT N00Type = N00.getValueType();
25728 EVT N10Type = N10.getValueType();
25729 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
25730 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
25731 return DAG.getBitcast(VT, FPLogic);
25737 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
25738 TargetLowering::DAGCombinerInfo &DCI,
25739 const X86Subtarget *Subtarget) {
25740 if (DCI.isBeforeLegalizeOps())
25743 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
25746 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25749 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25752 EVT VT = N->getValueType(0);
25753 SDValue N0 = N->getOperand(0);
25754 SDValue N1 = N->getOperand(1);
25757 // Create BEXTR instructions
25758 // BEXTR is ((X >> imm) & (2**size-1))
25759 if (VT == MVT::i32 || VT == MVT::i64) {
25760 // Check for BEXTR.
25761 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
25762 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
25763 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
25764 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25765 if (MaskNode && ShiftNode) {
25766 uint64_t Mask = MaskNode->getZExtValue();
25767 uint64_t Shift = ShiftNode->getZExtValue();
25768 if (isMask_64(Mask)) {
25769 uint64_t MaskSize = countPopulation(Mask);
25770 if (Shift + MaskSize <= VT.getSizeInBits())
25771 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25772 DAG.getConstant(Shift | (MaskSize << 8), DL,
25781 // Want to form ANDNP nodes:
25782 // 1) In the hopes of then easily combining them with OR and AND nodes
25783 // to form PBLEND/PSIGN.
25784 // 2) To match ANDN packed intrinsics
25785 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25788 // Check LHS for vnot
25789 if (N0.getOpcode() == ISD::XOR &&
25790 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25791 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25792 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25794 // Check RHS for vnot
25795 if (N1.getOpcode() == ISD::XOR &&
25796 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25797 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25798 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25803 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25804 TargetLowering::DAGCombinerInfo &DCI,
25805 const X86Subtarget *Subtarget) {
25806 if (DCI.isBeforeLegalizeOps())
25809 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25812 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25815 SDValue N0 = N->getOperand(0);
25816 SDValue N1 = N->getOperand(1);
25817 EVT VT = N->getValueType(0);
25819 // look for psign/blend
25820 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25821 if (!Subtarget->hasSSSE3() ||
25822 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25825 // Canonicalize pandn to RHS
25826 if (N0.getOpcode() == X86ISD::ANDNP)
25828 // or (and (m, y), (pandn m, x))
25829 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25830 SDValue Mask = N1.getOperand(0);
25831 SDValue X = N1.getOperand(1);
25833 if (N0.getOperand(0) == Mask)
25834 Y = N0.getOperand(1);
25835 if (N0.getOperand(1) == Mask)
25836 Y = N0.getOperand(0);
25838 // Check to see if the mask appeared in both the AND and ANDNP and
25842 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25843 // Look through mask bitcast.
25844 if (Mask.getOpcode() == ISD::BITCAST)
25845 Mask = Mask.getOperand(0);
25846 if (X.getOpcode() == ISD::BITCAST)
25847 X = X.getOperand(0);
25848 if (Y.getOpcode() == ISD::BITCAST)
25849 Y = Y.getOperand(0);
25851 EVT MaskVT = Mask.getValueType();
25853 // Validate that the Mask operand is a vector sra node.
25854 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25855 // there is no psrai.b
25856 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25857 unsigned SraAmt = ~0;
25858 if (Mask.getOpcode() == ISD::SRA) {
25859 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25860 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25861 SraAmt = AmtConst->getZExtValue();
25862 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25863 SDValue SraC = Mask.getOperand(1);
25864 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25866 if ((SraAmt + 1) != EltBits)
25871 // Now we know we at least have a plendvb with the mask val. See if
25872 // we can form a psignb/w/d.
25873 // psign = x.type == y.type == mask.type && y = sub(0, x);
25874 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25875 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25876 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25877 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25878 "Unsupported VT for PSIGN");
25879 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25880 return DAG.getBitcast(VT, Mask);
25882 // PBLENDVB only available on SSE 4.1
25883 if (!Subtarget->hasSSE41())
25886 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25888 X = DAG.getBitcast(BlendVT, X);
25889 Y = DAG.getBitcast(BlendVT, Y);
25890 Mask = DAG.getBitcast(BlendVT, Mask);
25891 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25892 return DAG.getBitcast(VT, Mask);
25896 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25899 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25900 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25902 // SHLD/SHRD instructions have lower register pressure, but on some
25903 // platforms they have higher latency than the equivalent
25904 // series of shifts/or that would otherwise be generated.
25905 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25906 // have higher latencies and we are not optimizing for size.
25907 if (!OptForSize && Subtarget->isSHLDSlow())
25910 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25912 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25914 if (!N0.hasOneUse() || !N1.hasOneUse())
25917 SDValue ShAmt0 = N0.getOperand(1);
25918 if (ShAmt0.getValueType() != MVT::i8)
25920 SDValue ShAmt1 = N1.getOperand(1);
25921 if (ShAmt1.getValueType() != MVT::i8)
25923 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25924 ShAmt0 = ShAmt0.getOperand(0);
25925 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25926 ShAmt1 = ShAmt1.getOperand(0);
25929 unsigned Opc = X86ISD::SHLD;
25930 SDValue Op0 = N0.getOperand(0);
25931 SDValue Op1 = N1.getOperand(0);
25932 if (ShAmt0.getOpcode() == ISD::SUB) {
25933 Opc = X86ISD::SHRD;
25934 std::swap(Op0, Op1);
25935 std::swap(ShAmt0, ShAmt1);
25938 unsigned Bits = VT.getSizeInBits();
25939 if (ShAmt1.getOpcode() == ISD::SUB) {
25940 SDValue Sum = ShAmt1.getOperand(0);
25941 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25942 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25943 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25944 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25945 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25946 return DAG.getNode(Opc, DL, VT,
25948 DAG.getNode(ISD::TRUNCATE, DL,
25951 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25952 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25954 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25955 return DAG.getNode(Opc, DL, VT,
25956 N0.getOperand(0), N1.getOperand(0),
25957 DAG.getNode(ISD::TRUNCATE, DL,
25964 // Generate NEG and CMOV for integer abs.
25965 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25966 EVT VT = N->getValueType(0);
25968 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25969 // 8-bit integer abs to NEG and CMOV.
25970 if (VT.isInteger() && VT.getSizeInBits() == 8)
25973 SDValue N0 = N->getOperand(0);
25974 SDValue N1 = N->getOperand(1);
25977 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25978 // and change it to SUB and CMOV.
25979 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25980 N0.getOpcode() == ISD::ADD &&
25981 N0.getOperand(1) == N1 &&
25982 N1.getOpcode() == ISD::SRA &&
25983 N1.getOperand(0) == N0.getOperand(0))
25984 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25985 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25986 // Generate SUB & CMOV.
25987 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25988 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25990 SDValue Ops[] = { N0.getOperand(0), Neg,
25991 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25992 SDValue(Neg.getNode(), 1) };
25993 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25998 // Try to turn tests against the signbit in the form of:
25999 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
26002 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
26003 // This is only worth doing if the output type is i8.
26004 if (N->getValueType(0) != MVT::i8)
26007 SDValue N0 = N->getOperand(0);
26008 SDValue N1 = N->getOperand(1);
26010 // We should be performing an xor against a truncated shift.
26011 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
26014 // Make sure we are performing an xor against one.
26015 if (!isOneConstant(N1))
26018 // SetCC on x86 zero extends so only act on this if it's a logical shift.
26019 SDValue Shift = N0.getOperand(0);
26020 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
26023 // Make sure we are truncating from one of i16, i32 or i64.
26024 EVT ShiftTy = Shift.getValueType();
26025 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
26028 // Make sure the shift amount extracts the sign bit.
26029 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
26030 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
26033 // Create a greater-than comparison against -1.
26034 // N.B. Using SETGE against 0 works but we want a canonical looking
26035 // comparison, using SETGT matches up with what TranslateX86CC.
26037 SDValue ShiftOp = Shift.getOperand(0);
26038 EVT ShiftOpTy = ShiftOp.getValueType();
26039 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
26040 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
26044 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
26045 TargetLowering::DAGCombinerInfo &DCI,
26046 const X86Subtarget *Subtarget) {
26047 if (DCI.isBeforeLegalizeOps())
26050 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
26053 if (Subtarget->hasCMov())
26054 if (SDValue RV = performIntegerAbsCombine(N, DAG))
26057 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
26063 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
26064 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
26065 /// X86ISD::AVG instruction.
26066 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
26067 const X86Subtarget *Subtarget, SDLoc DL) {
26068 if (!VT.isVector() || !VT.isSimple())
26070 EVT InVT = In.getValueType();
26071 unsigned NumElems = VT.getVectorNumElements();
26073 EVT ScalarVT = VT.getVectorElementType();
26074 if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
26075 isPowerOf2_32(NumElems)))
26078 // InScalarVT is the intermediate type in AVG pattern and it should be greater
26079 // than the original input type (i8/i16).
26080 EVT InScalarVT = InVT.getVectorElementType();
26081 if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
26084 if (Subtarget->hasAVX512()) {
26085 if (VT.getSizeInBits() > 512)
26087 } else if (Subtarget->hasAVX2()) {
26088 if (VT.getSizeInBits() > 256)
26091 if (VT.getSizeInBits() > 128)
26095 // Detect the following pattern:
26097 // %1 = zext <N x i8> %a to <N x i32>
26098 // %2 = zext <N x i8> %b to <N x i32>
26099 // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
26100 // %4 = add nuw nsw <N x i32> %3, %2
26101 // %5 = lshr <N x i32> %N, <i32 1 x N>
26102 // %6 = trunc <N x i32> %5 to <N x i8>
26104 // In AVX512, the last instruction can also be a trunc store.
26106 if (In.getOpcode() != ISD::SRL)
26109 // A lambda checking the given SDValue is a constant vector and each element
26110 // is in the range [Min, Max].
26111 auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
26112 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
26113 if (!BV || !BV->isConstant())
26115 for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
26116 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
26119 uint64_t Val = C->getZExtValue();
26120 if (Val < Min || Val > Max)
26126 // Check if each element of the vector is left-shifted by one.
26127 auto LHS = In.getOperand(0);
26128 auto RHS = In.getOperand(1);
26129 if (!IsConstVectorInRange(RHS, 1, 1))
26131 if (LHS.getOpcode() != ISD::ADD)
26134 // Detect a pattern of a + b + 1 where the order doesn't matter.
26135 SDValue Operands[3];
26136 Operands[0] = LHS.getOperand(0);
26137 Operands[1] = LHS.getOperand(1);
26139 // Take care of the case when one of the operands is a constant vector whose
26140 // element is in the range [1, 256].
26141 if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
26142 Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
26143 Operands[0].getOperand(0).getValueType() == VT) {
26144 // The pattern is detected. Subtract one from the constant vector, then
26145 // demote it and emit X86ISD::AVG instruction.
26146 SDValue One = DAG.getConstant(1, DL, InScalarVT);
26147 SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
26148 SmallVector<SDValue, 8>(NumElems, One));
26149 Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
26150 Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
26151 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26155 if (Operands[0].getOpcode() == ISD::ADD)
26156 std::swap(Operands[0], Operands[1]);
26157 else if (Operands[1].getOpcode() != ISD::ADD)
26159 Operands[2] = Operands[1].getOperand(0);
26160 Operands[1] = Operands[1].getOperand(1);
26162 // Now we have three operands of two additions. Check that one of them is a
26163 // constant vector with ones, and the other two are promoted from i8/i16.
26164 for (int i = 0; i < 3; ++i) {
26165 if (!IsConstVectorInRange(Operands[i], 1, 1))
26167 std::swap(Operands[i], Operands[2]);
26169 // Check if Operands[0] and Operands[1] are results of type promotion.
26170 for (int j = 0; j < 2; ++j)
26171 if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
26172 Operands[j].getOperand(0).getValueType() != VT)
26175 // The pattern is detected, emit X86ISD::AVG instruction.
26176 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26177 Operands[1].getOperand(0));
26183 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
26184 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
26185 TargetLowering::DAGCombinerInfo &DCI,
26186 const X86Subtarget *Subtarget) {
26187 LoadSDNode *Ld = cast<LoadSDNode>(N);
26188 EVT RegVT = Ld->getValueType(0);
26189 EVT MemVT = Ld->getMemoryVT();
26191 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26193 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
26194 // into two 16-byte operations.
26195 ISD::LoadExtType Ext = Ld->getExtensionType();
26197 unsigned AddressSpace = Ld->getAddressSpace();
26198 unsigned Alignment = Ld->getAlignment();
26199 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
26200 Ext == ISD::NON_EXTLOAD &&
26201 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
26202 AddressSpace, Alignment, &Fast) && !Fast) {
26203 unsigned NumElems = RegVT.getVectorNumElements();
26207 SDValue Ptr = Ld->getBasePtr();
26208 SDValue Increment =
26209 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26211 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
26213 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26214 Ld->getPointerInfo(), Ld->isVolatile(),
26215 Ld->isNonTemporal(), Ld->isInvariant(),
26217 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26218 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26219 Ld->getPointerInfo(), Ld->isVolatile(),
26220 Ld->isNonTemporal(), Ld->isInvariant(),
26221 std::min(16U, Alignment));
26222 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
26224 Load2.getValue(1));
26226 SDValue NewVec = DAG.getUNDEF(RegVT);
26227 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
26228 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
26229 return DCI.CombineTo(N, NewVec, TF, true);
26235 /// PerformMLOADCombine - Resolve extending loads
26236 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
26237 TargetLowering::DAGCombinerInfo &DCI,
26238 const X86Subtarget *Subtarget) {
26239 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
26240 if (Mld->getExtensionType() != ISD::SEXTLOAD)
26243 EVT VT = Mld->getValueType(0);
26244 unsigned NumElems = VT.getVectorNumElements();
26245 EVT LdVT = Mld->getMemoryVT();
26248 assert(LdVT != VT && "Cannot extend to the same type");
26249 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
26250 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
26251 // From, To sizes and ElemCount must be pow of two
26252 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26253 "Unexpected size for extending masked load");
26255 unsigned SizeRatio = ToSz / FromSz;
26256 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
26258 // Create a type on which we perform the shuffle
26259 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26260 LdVT.getScalarType(), NumElems*SizeRatio);
26261 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26263 // Convert Src0 value
26264 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
26265 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
26266 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26267 for (unsigned i = 0; i != NumElems; ++i)
26268 ShuffleVec[i] = i * SizeRatio;
26270 // Can't shuffle using an illegal type.
26271 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26272 "WideVecVT should be legal");
26273 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
26274 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
26276 // Prepare the new mask
26278 SDValue Mask = Mld->getMask();
26279 if (Mask.getValueType() == VT) {
26280 // Mask and original value have the same type
26281 NewMask = DAG.getBitcast(WideVecVT, Mask);
26282 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26283 for (unsigned i = 0; i != NumElems; ++i)
26284 ShuffleVec[i] = i * SizeRatio;
26285 for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
26286 ShuffleVec[i] = NumElems * SizeRatio;
26287 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26288 DAG.getConstant(0, dl, WideVecVT),
26292 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26293 unsigned WidenNumElts = NumElems*SizeRatio;
26294 unsigned MaskNumElts = VT.getVectorNumElements();
26295 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26298 unsigned NumConcat = WidenNumElts / MaskNumElts;
26299 SmallVector<SDValue, 16> Ops(NumConcat);
26300 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26302 for (unsigned i = 1; i != NumConcat; ++i)
26305 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26308 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
26309 Mld->getBasePtr(), NewMask, WideSrc0,
26310 Mld->getMemoryVT(), Mld->getMemOperand(),
26312 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
26313 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
26315 /// PerformMSTORECombine - Resolve truncating stores
26316 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
26317 const X86Subtarget *Subtarget) {
26318 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
26319 if (!Mst->isTruncatingStore())
26322 EVT VT = Mst->getValue().getValueType();
26323 unsigned NumElems = VT.getVectorNumElements();
26324 EVT StVT = Mst->getMemoryVT();
26327 assert(StVT != VT && "Cannot truncate to the same type");
26328 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26329 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26331 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26333 // The truncating store is legal in some cases. For example
26334 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26335 // are designated for truncate store.
26336 // In this case we don't need any further transformations.
26337 if (TLI.isTruncStoreLegal(VT, StVT))
26340 // From, To sizes and ElemCount must be pow of two
26341 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26342 "Unexpected size for truncating masked store");
26343 // We are going to use the original vector elt for storing.
26344 // Accumulated smaller vector elements must be a multiple of the store size.
26345 assert (((NumElems * FromSz) % ToSz) == 0 &&
26346 "Unexpected ratio for truncating masked store");
26348 unsigned SizeRatio = FromSz / ToSz;
26349 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26351 // Create a type on which we perform the shuffle
26352 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26353 StVT.getScalarType(), NumElems*SizeRatio);
26355 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26357 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
26358 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26359 for (unsigned i = 0; i != NumElems; ++i)
26360 ShuffleVec[i] = i * SizeRatio;
26362 // Can't shuffle using an illegal type.
26363 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26364 "WideVecVT should be legal");
26366 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26367 DAG.getUNDEF(WideVecVT),
26371 SDValue Mask = Mst->getMask();
26372 if (Mask.getValueType() == VT) {
26373 // Mask and original value have the same type
26374 NewMask = DAG.getBitcast(WideVecVT, Mask);
26375 for (unsigned i = 0; i != NumElems; ++i)
26376 ShuffleVec[i] = i * SizeRatio;
26377 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
26378 ShuffleVec[i] = NumElems*SizeRatio;
26379 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26380 DAG.getConstant(0, dl, WideVecVT),
26384 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26385 unsigned WidenNumElts = NumElems*SizeRatio;
26386 unsigned MaskNumElts = VT.getVectorNumElements();
26387 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26390 unsigned NumConcat = WidenNumElts / MaskNumElts;
26391 SmallVector<SDValue, 16> Ops(NumConcat);
26392 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26394 for (unsigned i = 1; i != NumConcat; ++i)
26397 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26400 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
26401 Mst->getBasePtr(), NewMask, StVT,
26402 Mst->getMemOperand(), false);
26404 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
26405 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
26406 const X86Subtarget *Subtarget) {
26407 StoreSDNode *St = cast<StoreSDNode>(N);
26408 EVT VT = St->getValue().getValueType();
26409 EVT StVT = St->getMemoryVT();
26411 SDValue StoredVal = St->getOperand(1);
26412 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26414 // If we are saving a concatenation of two XMM registers and 32-byte stores
26415 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
26417 unsigned AddressSpace = St->getAddressSpace();
26418 unsigned Alignment = St->getAlignment();
26419 if (VT.is256BitVector() && StVT == VT &&
26420 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
26421 AddressSpace, Alignment, &Fast) && !Fast) {
26422 unsigned NumElems = VT.getVectorNumElements();
26426 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
26427 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
26430 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26431 SDValue Ptr0 = St->getBasePtr();
26432 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
26434 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
26435 St->getPointerInfo(), St->isVolatile(),
26436 St->isNonTemporal(), Alignment);
26437 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
26438 St->getPointerInfo(), St->isVolatile(),
26439 St->isNonTemporal(),
26440 std::min(16U, Alignment));
26441 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
26444 // Optimize trunc store (of multiple scalars) to shuffle and store.
26445 // First, pack all of the elements in one place. Next, store to memory
26446 // in fewer chunks.
26447 if (St->isTruncatingStore() && VT.isVector()) {
26448 // Check if we can detect an AVG pattern from the truncation. If yes,
26449 // replace the trunc store by a normal store with the result of X86ISD::AVG
26452 detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
26454 return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
26455 St->getPointerInfo(), St->isVolatile(),
26456 St->isNonTemporal(), St->getAlignment());
26458 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26459 unsigned NumElems = VT.getVectorNumElements();
26460 assert(StVT != VT && "Cannot truncate to the same type");
26461 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26462 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26464 // The truncating store is legal in some cases. For example
26465 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26466 // are designated for truncate store.
26467 // In this case we don't need any further transformations.
26468 if (TLI.isTruncStoreLegal(VT, StVT))
26471 // From, To sizes and ElemCount must be pow of two
26472 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
26473 // We are going to use the original vector elt for storing.
26474 // Accumulated smaller vector elements must be a multiple of the store size.
26475 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
26477 unsigned SizeRatio = FromSz / ToSz;
26479 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26481 // Create a type on which we perform the shuffle
26482 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26483 StVT.getScalarType(), NumElems*SizeRatio);
26485 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26487 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
26488 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
26489 for (unsigned i = 0; i != NumElems; ++i)
26490 ShuffleVec[i] = i * SizeRatio;
26492 // Can't shuffle using an illegal type.
26493 if (!TLI.isTypeLegal(WideVecVT))
26496 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26497 DAG.getUNDEF(WideVecVT),
26499 // At this point all of the data is stored at the bottom of the
26500 // register. We now need to save it to mem.
26502 // Find the largest store unit
26503 MVT StoreType = MVT::i8;
26504 for (MVT Tp : MVT::integer_valuetypes()) {
26505 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
26509 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
26510 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
26511 (64 <= NumElems * ToSz))
26512 StoreType = MVT::f64;
26514 // Bitcast the original vector into a vector of store-size units
26515 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
26516 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
26517 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
26518 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
26519 SmallVector<SDValue, 8> Chains;
26520 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
26521 TLI.getPointerTy(DAG.getDataLayout()));
26522 SDValue Ptr = St->getBasePtr();
26524 // Perform one or more big stores into memory.
26525 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
26526 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
26527 StoreType, ShuffWide,
26528 DAG.getIntPtrConstant(i, dl));
26529 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
26530 St->getPointerInfo(), St->isVolatile(),
26531 St->isNonTemporal(), St->getAlignment());
26532 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26533 Chains.push_back(Ch);
26536 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
26539 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
26540 // the FP state in cases where an emms may be missing.
26541 // A preferable solution to the general problem is to figure out the right
26542 // places to insert EMMS. This qualifies as a quick hack.
26544 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
26545 if (VT.getSizeInBits() != 64)
26548 const Function *F = DAG.getMachineFunction().getFunction();
26549 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
26551 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
26552 if ((VT.isVector() ||
26553 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
26554 isa<LoadSDNode>(St->getValue()) &&
26555 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
26556 St->getChain().hasOneUse() && !St->isVolatile()) {
26557 SDNode* LdVal = St->getValue().getNode();
26558 LoadSDNode *Ld = nullptr;
26559 int TokenFactorIndex = -1;
26560 SmallVector<SDValue, 8> Ops;
26561 SDNode* ChainVal = St->getChain().getNode();
26562 // Must be a store of a load. We currently handle two cases: the load
26563 // is a direct child, and it's under an intervening TokenFactor. It is
26564 // possible to dig deeper under nested TokenFactors.
26565 if (ChainVal == LdVal)
26566 Ld = cast<LoadSDNode>(St->getChain());
26567 else if (St->getValue().hasOneUse() &&
26568 ChainVal->getOpcode() == ISD::TokenFactor) {
26569 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
26570 if (ChainVal->getOperand(i).getNode() == LdVal) {
26571 TokenFactorIndex = i;
26572 Ld = cast<LoadSDNode>(St->getValue());
26574 Ops.push_back(ChainVal->getOperand(i));
26578 if (!Ld || !ISD::isNormalLoad(Ld))
26581 // If this is not the MMX case, i.e. we are just turning i64 load/store
26582 // into f64 load/store, avoid the transformation if there are multiple
26583 // uses of the loaded value.
26584 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
26589 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
26590 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
26592 if (Subtarget->is64Bit() || F64IsLegal) {
26593 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
26594 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
26595 Ld->getPointerInfo(), Ld->isVolatile(),
26596 Ld->isNonTemporal(), Ld->isInvariant(),
26597 Ld->getAlignment());
26598 SDValue NewChain = NewLd.getValue(1);
26599 if (TokenFactorIndex != -1) {
26600 Ops.push_back(NewChain);
26601 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26603 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
26604 St->getPointerInfo(),
26605 St->isVolatile(), St->isNonTemporal(),
26606 St->getAlignment());
26609 // Otherwise, lower to two pairs of 32-bit loads / stores.
26610 SDValue LoAddr = Ld->getBasePtr();
26611 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
26612 DAG.getConstant(4, LdDL, MVT::i32));
26614 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
26615 Ld->getPointerInfo(),
26616 Ld->isVolatile(), Ld->isNonTemporal(),
26617 Ld->isInvariant(), Ld->getAlignment());
26618 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
26619 Ld->getPointerInfo().getWithOffset(4),
26620 Ld->isVolatile(), Ld->isNonTemporal(),
26622 MinAlign(Ld->getAlignment(), 4));
26624 SDValue NewChain = LoLd.getValue(1);
26625 if (TokenFactorIndex != -1) {
26626 Ops.push_back(LoLd);
26627 Ops.push_back(HiLd);
26628 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26631 LoAddr = St->getBasePtr();
26632 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
26633 DAG.getConstant(4, StDL, MVT::i32));
26635 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
26636 St->getPointerInfo(),
26637 St->isVolatile(), St->isNonTemporal(),
26638 St->getAlignment());
26639 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
26640 St->getPointerInfo().getWithOffset(4),
26642 St->isNonTemporal(),
26643 MinAlign(St->getAlignment(), 4));
26644 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
26647 // This is similar to the above case, but here we handle a scalar 64-bit
26648 // integer store that is extracted from a vector on a 32-bit target.
26649 // If we have SSE2, then we can treat it like a floating-point double
26650 // to get past legalization. The execution dependencies fixup pass will
26651 // choose the optimal machine instruction for the store if this really is
26652 // an integer or v2f32 rather than an f64.
26653 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
26654 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
26655 SDValue OldExtract = St->getOperand(1);
26656 SDValue ExtOp0 = OldExtract.getOperand(0);
26657 unsigned VecSize = ExtOp0.getValueSizeInBits();
26658 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
26659 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
26660 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
26661 BitCast, OldExtract.getOperand(1));
26662 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
26663 St->getPointerInfo(), St->isVolatile(),
26664 St->isNonTemporal(), St->getAlignment());
26670 /// Return 'true' if this vector operation is "horizontal"
26671 /// and return the operands for the horizontal operation in LHS and RHS. A
26672 /// horizontal operation performs the binary operation on successive elements
26673 /// of its first operand, then on successive elements of its second operand,
26674 /// returning the resulting values in a vector. For example, if
26675 /// A = < float a0, float a1, float a2, float a3 >
26677 /// B = < float b0, float b1, float b2, float b3 >
26678 /// then the result of doing a horizontal operation on A and B is
26679 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
26680 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
26681 /// A horizontal-op B, for some already available A and B, and if so then LHS is
26682 /// set to A, RHS to B, and the routine returns 'true'.
26683 /// Note that the binary operation should have the property that if one of the
26684 /// operands is UNDEF then the result is UNDEF.
26685 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
26686 // Look for the following pattern: if
26687 // A = < float a0, float a1, float a2, float a3 >
26688 // B = < float b0, float b1, float b2, float b3 >
26690 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
26691 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
26692 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
26693 // which is A horizontal-op B.
26695 // At least one of the operands should be a vector shuffle.
26696 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
26697 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
26700 MVT VT = LHS.getSimpleValueType();
26702 assert((VT.is128BitVector() || VT.is256BitVector()) &&
26703 "Unsupported vector type for horizontal add/sub");
26705 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
26706 // operate independently on 128-bit lanes.
26707 unsigned NumElts = VT.getVectorNumElements();
26708 unsigned NumLanes = VT.getSizeInBits()/128;
26709 unsigned NumLaneElts = NumElts / NumLanes;
26710 assert((NumLaneElts % 2 == 0) &&
26711 "Vector type should have an even number of elements in each lane");
26712 unsigned HalfLaneElts = NumLaneElts/2;
26714 // View LHS in the form
26715 // LHS = VECTOR_SHUFFLE A, B, LMask
26716 // If LHS is not a shuffle then pretend it is the shuffle
26717 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
26718 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
26721 SmallVector<int, 16> LMask(NumElts);
26722 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26723 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
26724 A = LHS.getOperand(0);
26725 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
26726 B = LHS.getOperand(1);
26727 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
26728 std::copy(Mask.begin(), Mask.end(), LMask.begin());
26730 if (LHS.getOpcode() != ISD::UNDEF)
26732 for (unsigned i = 0; i != NumElts; ++i)
26736 // Likewise, view RHS in the form
26737 // RHS = VECTOR_SHUFFLE C, D, RMask
26739 SmallVector<int, 16> RMask(NumElts);
26740 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26741 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
26742 C = RHS.getOperand(0);
26743 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
26744 D = RHS.getOperand(1);
26745 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
26746 std::copy(Mask.begin(), Mask.end(), RMask.begin());
26748 if (RHS.getOpcode() != ISD::UNDEF)
26750 for (unsigned i = 0; i != NumElts; ++i)
26754 // Check that the shuffles are both shuffling the same vectors.
26755 if (!(A == C && B == D) && !(A == D && B == C))
26758 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
26759 if (!A.getNode() && !B.getNode())
26762 // If A and B occur in reverse order in RHS, then "swap" them (which means
26763 // rewriting the mask).
26765 ShuffleVectorSDNode::commuteMask(RMask);
26767 // At this point LHS and RHS are equivalent to
26768 // LHS = VECTOR_SHUFFLE A, B, LMask
26769 // RHS = VECTOR_SHUFFLE A, B, RMask
26770 // Check that the masks correspond to performing a horizontal operation.
26771 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
26772 for (unsigned i = 0; i != NumLaneElts; ++i) {
26773 int LIdx = LMask[i+l], RIdx = RMask[i+l];
26775 // Ignore any UNDEF components.
26776 if (LIdx < 0 || RIdx < 0 ||
26777 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
26778 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
26781 // Check that successive elements are being operated on. If not, this is
26782 // not a horizontal operation.
26783 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
26784 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
26785 if (!(LIdx == Index && RIdx == Index + 1) &&
26786 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
26791 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
26792 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
26796 /// Do target-specific dag combines on floating point adds.
26797 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
26798 const X86Subtarget *Subtarget) {
26799 EVT VT = N->getValueType(0);
26800 SDValue LHS = N->getOperand(0);
26801 SDValue RHS = N->getOperand(1);
26803 // Try to synthesize horizontal adds from adds of shuffles.
26804 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26805 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26806 isHorizontalBinOp(LHS, RHS, true))
26807 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
26811 /// Do target-specific dag combines on floating point subs.
26812 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
26813 const X86Subtarget *Subtarget) {
26814 EVT VT = N->getValueType(0);
26815 SDValue LHS = N->getOperand(0);
26816 SDValue RHS = N->getOperand(1);
26818 // Try to synthesize horizontal subs from subs of shuffles.
26819 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26820 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26821 isHorizontalBinOp(LHS, RHS, false))
26822 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
26826 /// Truncate a group of v4i32 into v16i8/v8i16 using X86ISD::PACKUS.
26828 combineVectorTruncationWithPACKUS(SDNode *N, SelectionDAG &DAG,
26829 SmallVector<SDValue, 8> &Regs) {
26830 assert(Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 ||
26831 Regs[0].getValueType() == MVT::v2i64));
26832 EVT OutVT = N->getValueType(0);
26833 EVT OutSVT = OutVT.getVectorElementType();
26834 EVT InVT = Regs[0].getValueType();
26835 EVT InSVT = InVT.getVectorElementType();
26838 // First, use mask to unset all bits that won't appear in the result.
26839 assert((OutSVT == MVT::i8 || OutSVT == MVT::i16) &&
26840 "OutSVT can only be either i8 or i16.");
26842 DAG.getConstant(OutSVT == MVT::i8 ? 0xFF : 0xFFFF, DL, InSVT);
26843 SDValue MaskVec = DAG.getNode(
26844 ISD::BUILD_VECTOR, DL, InVT,
26845 SmallVector<SDValue, 8>(InVT.getVectorNumElements(), MaskVal));
26846 for (auto &Reg : Regs)
26847 Reg = DAG.getNode(ISD::AND, DL, InVT, MaskVec, Reg);
26849 MVT UnpackedVT, PackedVT;
26850 if (OutSVT == MVT::i8) {
26851 UnpackedVT = MVT::v8i16;
26852 PackedVT = MVT::v16i8;
26854 UnpackedVT = MVT::v4i32;
26855 PackedVT = MVT::v8i16;
26858 // In each iteration, truncate the type by a half size.
26859 auto RegNum = Regs.size();
26860 for (unsigned j = 1, e = InSVT.getSizeInBits() / OutSVT.getSizeInBits();
26861 j < e; j *= 2, RegNum /= 2) {
26862 for (unsigned i = 0; i < RegNum; i++)
26863 Regs[i] = DAG.getNode(ISD::BITCAST, DL, UnpackedVT, Regs[i]);
26864 for (unsigned i = 0; i < RegNum / 2; i++)
26865 Regs[i] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[i * 2],
26869 // If the type of the result is v8i8, we need do one more X86ISD::PACKUS, and
26870 // then extract a subvector as the result since v8i8 is not a legal type.
26871 if (OutVT == MVT::v8i8) {
26872 Regs[0] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[0], Regs[0]);
26873 Regs[0] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OutVT, Regs[0],
26874 DAG.getIntPtrConstant(0, DL));
26876 } else if (RegNum > 1) {
26877 Regs.resize(RegNum);
26878 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26883 /// Truncate a group of v4i32 into v8i16 using X86ISD::PACKSS.
26885 combineVectorTruncationWithPACKSS(SDNode *N, SelectionDAG &DAG,
26886 SmallVector<SDValue, 8> &Regs) {
26887 assert(Regs.size() > 0 && Regs[0].getValueType() == MVT::v4i32);
26888 EVT OutVT = N->getValueType(0);
26891 // Shift left by 16 bits, then arithmetic-shift right by 16 bits.
26892 SDValue ShAmt = DAG.getConstant(16, DL, MVT::i32);
26893 for (auto &Reg : Regs) {
26894 Reg = getTargetVShiftNode(X86ISD::VSHLI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26895 Reg = getTargetVShiftNode(X86ISD::VSRAI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26898 for (unsigned i = 0, e = Regs.size() / 2; i < e; i++)
26899 Regs[i] = DAG.getNode(X86ISD::PACKSS, DL, MVT::v8i16, Regs[i * 2],
26902 if (Regs.size() > 2) {
26903 Regs.resize(Regs.size() / 2);
26904 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26909 /// This function transforms truncation from vXi32/vXi64 to vXi8/vXi16 into
26910 /// X86ISD::PACKUS/X86ISD::PACKSS operations. We do it here because after type
26911 /// legalization the truncation will be translated into a BUILD_VECTOR with each
26912 /// element that is extracted from a vector and then truncated, and it is
26913 /// diffcult to do this optimization based on them.
26914 static SDValue combineVectorTruncation(SDNode *N, SelectionDAG &DAG,
26915 const X86Subtarget *Subtarget) {
26916 EVT OutVT = N->getValueType(0);
26917 if (!OutVT.isVector())
26920 SDValue In = N->getOperand(0);
26921 if (!In.getValueType().isSimple())
26924 EVT InVT = In.getValueType();
26925 unsigned NumElems = OutVT.getVectorNumElements();
26927 // TODO: On AVX2, the behavior of X86ISD::PACKUS is different from that on
26928 // SSE2, and we need to take care of it specially.
26929 // AVX512 provides vpmovdb.
26930 if (!Subtarget->hasSSE2() || Subtarget->hasAVX2())
26933 EVT OutSVT = OutVT.getVectorElementType();
26934 EVT InSVT = InVT.getVectorElementType();
26935 if (!((InSVT == MVT::i32 || InSVT == MVT::i64) &&
26936 (OutSVT == MVT::i8 || OutSVT == MVT::i16) && isPowerOf2_32(NumElems) &&
26940 // SSSE3's pshufb results in less instructions in the cases below.
26941 if (Subtarget->hasSSSE3() && NumElems == 8 &&
26942 ((OutSVT == MVT::i8 && InSVT != MVT::i64) ||
26943 (InSVT == MVT::i32 && OutSVT == MVT::i16)))
26948 // Split a long vector into vectors of legal type.
26949 unsigned RegNum = InVT.getSizeInBits() / 128;
26950 SmallVector<SDValue, 8> SubVec(RegNum);
26951 if (InSVT == MVT::i32) {
26952 for (unsigned i = 0; i < RegNum; i++)
26953 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
26954 DAG.getIntPtrConstant(i * 4, DL));
26956 for (unsigned i = 0; i < RegNum; i++)
26957 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
26958 DAG.getIntPtrConstant(i * 2, DL));
26961 // SSE2 provides PACKUS for only 2 x v8i16 -> v16i8 and SSE4.1 provides PAKCUS
26962 // for 2 x v4i32 -> v8i16. For SSSE3 and below, we need to use PACKSS to
26963 // truncate 2 x v4i32 to v8i16.
26964 if (Subtarget->hasSSE41() || OutSVT == MVT::i8)
26965 return combineVectorTruncationWithPACKUS(N, DAG, SubVec);
26966 else if (InSVT == MVT::i32)
26967 return combineVectorTruncationWithPACKSS(N, DAG, SubVec);
26972 static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
26973 const X86Subtarget *Subtarget) {
26974 // Try to detect AVG pattern first.
26975 SDValue Avg = detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG,
26976 Subtarget, SDLoc(N));
26980 return combineVectorTruncation(N, DAG, Subtarget);
26983 /// Do target-specific dag combines on floating point negations.
26984 static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
26985 const X86Subtarget *Subtarget) {
26986 EVT VT = N->getValueType(0);
26987 EVT SVT = VT.getScalarType();
26988 SDValue Arg = N->getOperand(0);
26991 // Let legalize expand this if it isn't a legal type yet.
26992 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26995 // If we're negating a FMUL node on a target with FMA, then we can avoid the
26996 // use of a constant by performing (-0 - A*B) instead.
26997 // FIXME: Check rounding control flags as well once it becomes available.
26998 if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) &&
26999 Arg->getFlags()->hasNoSignedZeros() && Subtarget->hasAnyFMA()) {
27000 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
27001 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
27002 Arg.getOperand(1), Zero);
27005 // If we're negating a FMA node, then we can adjust the
27006 // instruction to include the extra negation.
27007 if (Arg.hasOneUse()) {
27008 switch (Arg.getOpcode()) {
27009 case X86ISD::FMADD:
27010 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
27011 Arg.getOperand(1), Arg.getOperand(2));
27012 case X86ISD::FMSUB:
27013 return DAG.getNode(X86ISD::FNMADD, DL, VT, Arg.getOperand(0),
27014 Arg.getOperand(1), Arg.getOperand(2));
27015 case X86ISD::FNMADD:
27016 return DAG.getNode(X86ISD::FMSUB, DL, VT, Arg.getOperand(0),
27017 Arg.getOperand(1), Arg.getOperand(2));
27018 case X86ISD::FNMSUB:
27019 return DAG.getNode(X86ISD::FMADD, DL, VT, Arg.getOperand(0),
27020 Arg.getOperand(1), Arg.getOperand(2));
27026 static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
27027 const X86Subtarget *Subtarget) {
27028 EVT VT = N->getValueType(0);
27029 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
27030 // VXORPS, VORPS, VANDPS, VANDNPS are supported only under DQ extention.
27031 // These logic operations may be executed in the integer domain.
27033 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
27034 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
27036 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
27037 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
27038 unsigned IntOpcode = 0;
27039 switch (N->getOpcode()) {
27040 default: llvm_unreachable("Unexpected FP logic op");
27041 case X86ISD::FOR: IntOpcode = ISD::OR; break;
27042 case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
27043 case X86ISD::FAND: IntOpcode = ISD::AND; break;
27044 case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
27046 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
27047 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
27051 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
27052 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
27053 const X86Subtarget *Subtarget) {
27054 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
27056 // F[X]OR(0.0, x) -> x
27057 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27058 if (C->getValueAPF().isPosZero())
27059 return N->getOperand(1);
27061 // F[X]OR(x, 0.0) -> x
27062 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27063 if (C->getValueAPF().isPosZero())
27064 return N->getOperand(0);
27066 return lowerX86FPLogicOp(N, DAG, Subtarget);
27069 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
27070 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
27071 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
27073 // Only perform optimizations if UnsafeMath is used.
27074 if (!DAG.getTarget().Options.UnsafeFPMath)
27077 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
27078 // into FMINC and FMAXC, which are Commutative operations.
27079 unsigned NewOp = 0;
27080 switch (N->getOpcode()) {
27081 default: llvm_unreachable("unknown opcode");
27082 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
27083 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
27086 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
27087 N->getOperand(0), N->getOperand(1));
27090 static SDValue performFMinNumFMaxNumCombine(SDNode *N, SelectionDAG &DAG,
27091 const X86Subtarget *Subtarget) {
27092 if (Subtarget->useSoftFloat())
27095 // TODO: Check for global or instruction-level "nnan". In that case, we
27096 // should be able to lower to FMAX/FMIN alone.
27097 // TODO: If an operand is already known to be a NaN or not a NaN, this
27098 // should be an optional swap and FMAX/FMIN.
27100 EVT VT = N->getValueType(0);
27101 if (!((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
27102 (Subtarget->hasSSE2() && (VT == MVT::f64 || VT == MVT::v2f64)) ||
27103 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))))
27106 // This takes at least 3 instructions, so favor a library call when operating
27107 // on a scalar and minimizing code size.
27108 if (!VT.isVector() && DAG.getMachineFunction().getFunction()->optForMinSize())
27111 SDValue Op0 = N->getOperand(0);
27112 SDValue Op1 = N->getOperand(1);
27114 EVT SetCCType = DAG.getTargetLoweringInfo().getSetCCResultType(
27115 DAG.getDataLayout(), *DAG.getContext(), VT);
27117 // There are 4 possibilities involving NaN inputs, and these are the required
27121 // ----------------
27122 // Num | Max | Op0 |
27123 // Op0 ----------------
27124 // NaN | Op1 | NaN |
27125 // ----------------
27127 // The SSE FP max/min instructions were not designed for this case, but rather
27129 // Min = Op1 < Op0 ? Op1 : Op0
27130 // Max = Op1 > Op0 ? Op1 : Op0
27132 // So they always return Op0 if either input is a NaN. However, we can still
27133 // use those instructions for fmaxnum by selecting away a NaN input.
27135 // If either operand is NaN, the 2nd source operand (Op0) is passed through.
27136 auto MinMaxOp = N->getOpcode() == ISD::FMAXNUM ? X86ISD::FMAX : X86ISD::FMIN;
27137 SDValue MinOrMax = DAG.getNode(MinMaxOp, DL, VT, Op1, Op0);
27138 SDValue IsOp0Nan = DAG.getSetCC(DL, SetCCType , Op0, Op0, ISD::SETUO);
27140 // If Op0 is a NaN, select Op1. Otherwise, select the max. If both operands
27141 // are NaN, the NaN value of Op1 is the result.
27142 auto SelectOpcode = VT.isVector() ? ISD::VSELECT : ISD::SELECT;
27143 return DAG.getNode(SelectOpcode, DL, VT, IsOp0Nan, Op1, MinOrMax);
27146 /// Do target-specific dag combines on X86ISD::FAND nodes.
27147 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG,
27148 const X86Subtarget *Subtarget) {
27149 // FAND(0.0, x) -> 0.0
27150 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27151 if (C->getValueAPF().isPosZero())
27152 return N->getOperand(0);
27154 // FAND(x, 0.0) -> 0.0
27155 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27156 if (C->getValueAPF().isPosZero())
27157 return N->getOperand(1);
27159 return lowerX86FPLogicOp(N, DAG, Subtarget);
27162 /// Do target-specific dag combines on X86ISD::FANDN nodes
27163 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG,
27164 const X86Subtarget *Subtarget) {
27165 // FANDN(0.0, x) -> x
27166 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27167 if (C->getValueAPF().isPosZero())
27168 return N->getOperand(1);
27170 // FANDN(x, 0.0) -> 0.0
27171 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27172 if (C->getValueAPF().isPosZero())
27173 return N->getOperand(1);
27175 return lowerX86FPLogicOp(N, DAG, Subtarget);
27178 static SDValue PerformBTCombine(SDNode *N,
27180 TargetLowering::DAGCombinerInfo &DCI) {
27181 // BT ignores high bits in the bit index operand.
27182 SDValue Op1 = N->getOperand(1);
27183 if (Op1.hasOneUse()) {
27184 unsigned BitWidth = Op1.getValueSizeInBits();
27185 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
27186 APInt KnownZero, KnownOne;
27187 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
27188 !DCI.isBeforeLegalizeOps());
27189 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27190 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
27191 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
27192 DCI.CommitTargetLoweringOpt(TLO);
27197 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
27198 SDValue Op = N->getOperand(0);
27199 if (Op.getOpcode() == ISD::BITCAST)
27200 Op = Op.getOperand(0);
27201 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
27202 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
27203 VT.getVectorElementType().getSizeInBits() ==
27204 OpVT.getVectorElementType().getSizeInBits()) {
27205 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
27210 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
27211 const X86Subtarget *Subtarget) {
27212 EVT VT = N->getValueType(0);
27213 if (!VT.isVector())
27216 SDValue N0 = N->getOperand(0);
27217 SDValue N1 = N->getOperand(1);
27218 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
27221 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
27222 // both SSE and AVX2 since there is no sign-extended shift right
27223 // operation on a vector with 64-bit elements.
27224 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
27225 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
27226 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
27227 N0.getOpcode() == ISD::SIGN_EXTEND)) {
27228 SDValue N00 = N0.getOperand(0);
27230 // EXTLOAD has a better solution on AVX2,
27231 // it may be replaced with X86ISD::VSEXT node.
27232 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
27233 if (!ISD::isNormalLoad(N00.getNode()))
27236 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
27237 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
27239 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
27245 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
27246 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
27247 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
27248 /// eliminate extend, add, and shift instructions.
27249 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
27250 const X86Subtarget *Subtarget) {
27251 // TODO: This should be valid for other integer types.
27252 EVT VT = Sext->getValueType(0);
27253 if (VT != MVT::i64)
27256 // We need an 'add nsw' feeding into the 'sext'.
27257 SDValue Add = Sext->getOperand(0);
27258 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
27261 // Having a constant operand to the 'add' ensures that we are not increasing
27262 // the instruction count because the constant is extended for free below.
27263 // A constant operand can also become the displacement field of an LEA.
27264 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
27268 // Don't make the 'add' bigger if there's no hope of combining it with some
27269 // other 'add' or 'shl' instruction.
27270 // TODO: It may be profitable to generate simpler LEA instructions in place
27271 // of single 'add' instructions, but the cost model for selecting an LEA
27272 // currently has a high threshold.
27273 bool HasLEAPotential = false;
27274 for (auto *User : Sext->uses()) {
27275 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
27276 HasLEAPotential = true;
27280 if (!HasLEAPotential)
27283 // Everything looks good, so pull the 'sext' ahead of the 'add'.
27284 int64_t AddConstant = AddOp1->getSExtValue();
27285 SDValue AddOp0 = Add.getOperand(0);
27286 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
27287 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
27289 // The wider add is guaranteed to not wrap because both operands are
27292 Flags.setNoSignedWrap(true);
27293 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
27296 /// (i8,i32 {s/z}ext ({s/u}divrem (i8 x, i8 y)) ->
27297 /// (i8,i32 ({s/u}divrem_sext_hreg (i8 x, i8 y)
27298 /// This exposes the {s/z}ext to the sdivrem lowering, so that it directly
27299 /// extends from AH (which we otherwise need to do contortions to access).
27300 static SDValue getDivRem8(SDNode *N, SelectionDAG &DAG) {
27301 SDValue N0 = N->getOperand(0);
27302 auto OpcodeN = N->getOpcode();
27303 auto OpcodeN0 = N0.getOpcode();
27304 if (!((OpcodeN == ISD::SIGN_EXTEND && OpcodeN0 == ISD::SDIVREM) ||
27305 (OpcodeN == ISD::ZERO_EXTEND && OpcodeN0 == ISD::UDIVREM)))
27308 EVT VT = N->getValueType(0);
27309 EVT InVT = N0.getValueType();
27310 if (N0.getResNo() != 1 || InVT != MVT::i8 || VT != MVT::i32)
27313 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
27314 auto DivRemOpcode = OpcodeN0 == ISD::SDIVREM ? X86ISD::SDIVREM8_SEXT_HREG
27315 : X86ISD::UDIVREM8_ZEXT_HREG;
27316 SDValue R = DAG.getNode(DivRemOpcode, SDLoc(N), NodeTys, N0.getOperand(0),
27318 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
27319 return R.getValue(1);
27322 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
27323 TargetLowering::DAGCombinerInfo &DCI,
27324 const X86Subtarget *Subtarget) {
27325 SDValue N0 = N->getOperand(0);
27326 EVT VT = N->getValueType(0);
27327 EVT SVT = VT.getScalarType();
27328 EVT InVT = N0.getValueType();
27329 EVT InSVT = InVT.getScalarType();
27332 if (SDValue DivRem8 = getDivRem8(N, DAG))
27335 if (!DCI.isBeforeLegalizeOps()) {
27336 if (InVT == MVT::i1) {
27337 SDValue Zero = DAG.getConstant(0, DL, VT);
27339 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
27340 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
27345 if (VT.isVector() && Subtarget->hasSSE2()) {
27346 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
27347 EVT InVT = N.getValueType();
27348 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
27349 Size / InVT.getScalarSizeInBits());
27350 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
27351 DAG.getUNDEF(InVT));
27353 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
27356 // If target-size is less than 128-bits, extend to a type that would extend
27357 // to 128 bits, extend that and extract the original target vector.
27358 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
27359 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27360 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27361 unsigned Scale = 128 / VT.getSizeInBits();
27363 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
27364 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
27365 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
27366 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
27367 DAG.getIntPtrConstant(0, DL));
27370 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
27371 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
27372 if (VT.getSizeInBits() == 128 &&
27373 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27374 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27375 SDValue ExOp = ExtendVecSize(DL, N0, 128);
27376 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
27379 // On pre-AVX2 targets, split into 128-bit nodes of
27380 // ISD::SIGN_EXTEND_VECTOR_INREG.
27381 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
27382 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27383 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27384 unsigned NumVecs = VT.getSizeInBits() / 128;
27385 unsigned NumSubElts = 128 / SVT.getSizeInBits();
27386 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
27387 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
27389 SmallVector<SDValue, 8> Opnds;
27390 for (unsigned i = 0, Offset = 0; i != NumVecs;
27391 ++i, Offset += NumSubElts) {
27392 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
27393 DAG.getIntPtrConstant(Offset, DL));
27394 SrcVec = ExtendVecSize(DL, SrcVec, 128);
27395 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
27396 Opnds.push_back(SrcVec);
27398 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
27402 if (Subtarget->hasAVX() && VT.is256BitVector())
27403 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27406 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
27412 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
27413 const X86Subtarget* Subtarget) {
27415 EVT VT = N->getValueType(0);
27417 // Let legalize expand this if it isn't a legal type yet.
27418 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
27421 EVT ScalarVT = VT.getScalarType();
27422 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
27425 SDValue A = N->getOperand(0);
27426 SDValue B = N->getOperand(1);
27427 SDValue C = N->getOperand(2);
27429 bool NegA = (A.getOpcode() == ISD::FNEG);
27430 bool NegB = (B.getOpcode() == ISD::FNEG);
27431 bool NegC = (C.getOpcode() == ISD::FNEG);
27433 // Negative multiplication when NegA xor NegB
27434 bool NegMul = (NegA != NegB);
27436 A = A.getOperand(0);
27438 B = B.getOperand(0);
27440 C = C.getOperand(0);
27444 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
27446 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
27448 return DAG.getNode(Opcode, dl, VT, A, B, C);
27451 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
27452 TargetLowering::DAGCombinerInfo &DCI,
27453 const X86Subtarget *Subtarget) {
27454 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
27455 // (and (i32 x86isd::setcc_carry), 1)
27456 // This eliminates the zext. This transformation is necessary because
27457 // ISD::SETCC is always legalized to i8.
27459 SDValue N0 = N->getOperand(0);
27460 EVT VT = N->getValueType(0);
27462 if (N0.getOpcode() == ISD::AND &&
27464 N0.getOperand(0).hasOneUse()) {
27465 SDValue N00 = N0.getOperand(0);
27466 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27467 if (!isOneConstant(N0.getOperand(1)))
27469 return DAG.getNode(ISD::AND, dl, VT,
27470 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27471 N00.getOperand(0), N00.getOperand(1)),
27472 DAG.getConstant(1, dl, VT));
27476 if (N0.getOpcode() == ISD::TRUNCATE &&
27478 N0.getOperand(0).hasOneUse()) {
27479 SDValue N00 = N0.getOperand(0);
27480 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27481 return DAG.getNode(ISD::AND, dl, VT,
27482 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27483 N00.getOperand(0), N00.getOperand(1)),
27484 DAG.getConstant(1, dl, VT));
27488 if (VT.is256BitVector())
27489 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27492 if (SDValue DivRem8 = getDivRem8(N, DAG))
27498 // Optimize x == -y --> x+y == 0
27499 // x != -y --> x+y != 0
27500 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
27501 const X86Subtarget* Subtarget) {
27502 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
27503 SDValue LHS = N->getOperand(0);
27504 SDValue RHS = N->getOperand(1);
27505 EVT VT = N->getValueType(0);
27508 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
27509 if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
27510 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
27511 LHS.getOperand(1));
27512 return DAG.getSetCC(DL, N->getValueType(0), addV,
27513 DAG.getConstant(0, DL, addV.getValueType()), CC);
27515 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
27516 if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
27517 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
27518 RHS.getOperand(1));
27519 return DAG.getSetCC(DL, N->getValueType(0), addV,
27520 DAG.getConstant(0, DL, addV.getValueType()), CC);
27523 if (VT.getScalarType() == MVT::i1 &&
27524 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
27526 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27527 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27528 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27530 if (!IsSEXT0 || !IsVZero1) {
27531 // Swap the operands and update the condition code.
27532 std::swap(LHS, RHS);
27533 CC = ISD::getSetCCSwappedOperands(CC);
27535 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27536 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27537 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27540 if (IsSEXT0 && IsVZero1) {
27541 assert(VT == LHS.getOperand(0).getValueType() &&
27542 "Uexpected operand type");
27543 if (CC == ISD::SETGT)
27544 return DAG.getConstant(0, DL, VT);
27545 if (CC == ISD::SETLE)
27546 return DAG.getConstant(1, DL, VT);
27547 if (CC == ISD::SETEQ || CC == ISD::SETGE)
27548 return DAG.getNOT(DL, LHS.getOperand(0), VT);
27550 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
27551 "Unexpected condition code!");
27552 return LHS.getOperand(0);
27559 static SDValue PerformGatherScatterCombine(SDNode *N, SelectionDAG &DAG) {
27561 // Gather and Scatter instructions use k-registers for masks. The type of
27562 // the masks is v*i1. So the mask will be truncated anyway.
27563 // The SIGN_EXTEND_INREG my be dropped.
27564 SDValue Mask = N->getOperand(2);
27565 if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) {
27566 SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end());
27567 NewOps[2] = Mask.getOperand(0);
27568 DAG.UpdateNodeOperands(N, NewOps);
27573 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
27574 // as "sbb reg,reg", since it can be extended without zext and produces
27575 // an all-ones bit which is more useful than 0/1 in some cases.
27576 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
27579 return DAG.getNode(ISD::AND, DL, VT,
27580 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27581 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27583 DAG.getConstant(1, DL, VT));
27584 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
27585 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
27586 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27587 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27591 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
27592 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
27593 TargetLowering::DAGCombinerInfo &DCI,
27594 const X86Subtarget *Subtarget) {
27596 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
27597 SDValue EFLAGS = N->getOperand(1);
27599 if (CC == X86::COND_A) {
27600 // Try to convert COND_A into COND_B in an attempt to facilitate
27601 // materializing "setb reg".
27603 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
27604 // cannot take an immediate as its first operand.
27606 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
27607 EFLAGS.getValueType().isInteger() &&
27608 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
27609 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
27610 EFLAGS.getNode()->getVTList(),
27611 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
27612 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
27613 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
27617 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
27618 // a zext and produces an all-ones bit which is more useful than 0/1 in some
27620 if (CC == X86::COND_B)
27621 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
27623 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27624 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27625 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
27631 // Optimize branch condition evaluation.
27633 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
27634 TargetLowering::DAGCombinerInfo &DCI,
27635 const X86Subtarget *Subtarget) {
27637 SDValue Chain = N->getOperand(0);
27638 SDValue Dest = N->getOperand(1);
27639 SDValue EFLAGS = N->getOperand(3);
27640 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
27642 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27643 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27644 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
27651 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
27652 SelectionDAG &DAG) {
27653 // Take advantage of vector comparisons producing 0 or -1 in each lane to
27654 // optimize away operation when it's from a constant.
27656 // The general transformation is:
27657 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
27658 // AND(VECTOR_CMP(x,y), constant2)
27659 // constant2 = UNARYOP(constant)
27661 // Early exit if this isn't a vector operation, the operand of the
27662 // unary operation isn't a bitwise AND, or if the sizes of the operations
27663 // aren't the same.
27664 EVT VT = N->getValueType(0);
27665 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
27666 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
27667 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
27670 // Now check that the other operand of the AND is a constant. We could
27671 // make the transformation for non-constant splats as well, but it's unclear
27672 // that would be a benefit as it would not eliminate any operations, just
27673 // perform one more step in scalar code before moving to the vector unit.
27674 if (BuildVectorSDNode *BV =
27675 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
27676 // Bail out if the vector isn't a constant.
27677 if (!BV->isConstant())
27680 // Everything checks out. Build up the new and improved node.
27682 EVT IntVT = BV->getValueType(0);
27683 // Create a new constant of the appropriate type for the transformed
27685 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
27686 // The AND node needs bitcasts to/from an integer vector type around it.
27687 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
27688 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
27689 N->getOperand(0)->getOperand(0), MaskConst);
27690 SDValue Res = DAG.getBitcast(VT, NewAnd);
27697 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27698 const X86Subtarget *Subtarget) {
27699 SDValue Op0 = N->getOperand(0);
27700 EVT VT = N->getValueType(0);
27701 EVT InVT = Op0.getValueType();
27702 EVT InSVT = InVT.getScalarType();
27703 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27705 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
27706 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
27707 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27709 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27710 InVT.getVectorNumElements());
27711 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
27713 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
27714 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
27716 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27722 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27723 const X86Subtarget *Subtarget) {
27724 // First try to optimize away the conversion entirely when it's
27725 // conditionally from a constant. Vectors only.
27726 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
27729 // Now move on to more general possibilities.
27730 SDValue Op0 = N->getOperand(0);
27731 EVT VT = N->getValueType(0);
27732 EVT InVT = Op0.getValueType();
27733 EVT InSVT = InVT.getScalarType();
27735 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
27736 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
27737 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27739 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27740 InVT.getVectorNumElements());
27741 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
27742 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27745 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
27746 // a 32-bit target where SSE doesn't support i64->FP operations.
27747 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
27748 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
27749 EVT LdVT = Ld->getValueType(0);
27751 // This transformation is not supported if the result type is f16
27752 if (VT == MVT::f16)
27755 if (!Ld->isVolatile() && !VT.isVector() &&
27756 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
27757 !Subtarget->is64Bit() && LdVT == MVT::i64) {
27758 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
27759 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
27760 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
27767 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
27768 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
27769 X86TargetLowering::DAGCombinerInfo &DCI) {
27770 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
27771 // the result is either zero or one (depending on the input carry bit).
27772 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
27773 if (X86::isZeroNode(N->getOperand(0)) &&
27774 X86::isZeroNode(N->getOperand(1)) &&
27775 // We don't have a good way to replace an EFLAGS use, so only do this when
27777 SDValue(N, 1).use_empty()) {
27779 EVT VT = N->getValueType(0);
27780 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
27781 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
27782 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
27783 DAG.getConstant(X86::COND_B, DL,
27786 DAG.getConstant(1, DL, VT));
27787 return DCI.CombineTo(N, Res1, CarryOut);
27793 // fold (add Y, (sete X, 0)) -> adc 0, Y
27794 // (add Y, (setne X, 0)) -> sbb -1, Y
27795 // (sub (sete X, 0), Y) -> sbb 0, Y
27796 // (sub (setne X, 0), Y) -> adc -1, Y
27797 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
27800 // Look through ZExts.
27801 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
27802 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
27805 SDValue SetCC = Ext.getOperand(0);
27806 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
27809 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
27810 if (CC != X86::COND_E && CC != X86::COND_NE)
27813 SDValue Cmp = SetCC.getOperand(1);
27814 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
27815 !X86::isZeroNode(Cmp.getOperand(1)) ||
27816 !Cmp.getOperand(0).getValueType().isInteger())
27819 SDValue CmpOp0 = Cmp.getOperand(0);
27820 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
27821 DAG.getConstant(1, DL, CmpOp0.getValueType()));
27823 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
27824 if (CC == X86::COND_NE)
27825 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
27826 DL, OtherVal.getValueType(), OtherVal,
27827 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
27829 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
27830 DL, OtherVal.getValueType(), OtherVal,
27831 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
27834 /// PerformADDCombine - Do target-specific dag combines on integer adds.
27835 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
27836 const X86Subtarget *Subtarget) {
27837 EVT VT = N->getValueType(0);
27838 SDValue Op0 = N->getOperand(0);
27839 SDValue Op1 = N->getOperand(1);
27841 // Try to synthesize horizontal adds from adds of shuffles.
27842 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27843 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27844 isHorizontalBinOp(Op0, Op1, true))
27845 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
27847 return OptimizeConditionalInDecrement(N, DAG);
27850 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
27851 const X86Subtarget *Subtarget) {
27852 SDValue Op0 = N->getOperand(0);
27853 SDValue Op1 = N->getOperand(1);
27855 // X86 can't encode an immediate LHS of a sub. See if we can push the
27856 // negation into a preceding instruction.
27857 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
27858 // If the RHS of the sub is a XOR with one use and a constant, invert the
27859 // immediate. Then add one to the LHS of the sub so we can turn
27860 // X-Y -> X+~Y+1, saving one register.
27861 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
27862 isa<ConstantSDNode>(Op1.getOperand(1))) {
27863 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
27864 EVT VT = Op0.getValueType();
27865 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
27867 DAG.getConstant(~XorC, SDLoc(Op1), VT));
27868 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
27869 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
27873 // Try to synthesize horizontal adds from adds of shuffles.
27874 EVT VT = N->getValueType(0);
27875 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27876 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27877 isHorizontalBinOp(Op0, Op1, true))
27878 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
27880 return OptimizeConditionalInDecrement(N, DAG);
27883 /// performVZEXTCombine - Performs build vector combines
27884 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
27885 TargetLowering::DAGCombinerInfo &DCI,
27886 const X86Subtarget *Subtarget) {
27888 MVT VT = N->getSimpleValueType(0);
27889 SDValue Op = N->getOperand(0);
27890 MVT OpVT = Op.getSimpleValueType();
27891 MVT OpEltVT = OpVT.getVectorElementType();
27892 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
27894 // (vzext (bitcast (vzext (x)) -> (vzext x)
27896 while (V.getOpcode() == ISD::BITCAST)
27897 V = V.getOperand(0);
27899 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
27900 MVT InnerVT = V.getSimpleValueType();
27901 MVT InnerEltVT = InnerVT.getVectorElementType();
27903 // If the element sizes match exactly, we can just do one larger vzext. This
27904 // is always an exact type match as vzext operates on integer types.
27905 if (OpEltVT == InnerEltVT) {
27906 assert(OpVT == InnerVT && "Types must match for vzext!");
27907 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
27910 // The only other way we can combine them is if only a single element of the
27911 // inner vzext is used in the input to the outer vzext.
27912 if (InnerEltVT.getSizeInBits() < InputBits)
27915 // In this case, the inner vzext is completely dead because we're going to
27916 // only look at bits inside of the low element. Just do the outer vzext on
27917 // a bitcast of the input to the inner.
27918 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
27921 // Check if we can bypass extracting and re-inserting an element of an input
27922 // vector. Essentially:
27923 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
27924 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
27925 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
27926 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
27927 SDValue ExtractedV = V.getOperand(0);
27928 SDValue OrigV = ExtractedV.getOperand(0);
27929 if (isNullConstant(ExtractedV.getOperand(1))) {
27930 MVT OrigVT = OrigV.getSimpleValueType();
27931 // Extract a subvector if necessary...
27932 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
27933 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
27934 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
27935 OrigVT.getVectorNumElements() / Ratio);
27936 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
27937 DAG.getIntPtrConstant(0, DL));
27939 Op = DAG.getBitcast(OpVT, OrigV);
27940 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
27947 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
27948 DAGCombinerInfo &DCI) const {
27949 SelectionDAG &DAG = DCI.DAG;
27950 switch (N->getOpcode()) {
27952 case ISD::EXTRACT_VECTOR_ELT:
27953 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
27956 case X86ISD::SHRUNKBLEND:
27957 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
27958 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
27959 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
27960 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
27961 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
27962 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
27963 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
27966 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
27967 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
27968 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
27969 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
27970 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
27971 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
27972 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
27973 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
27974 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
27975 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
27976 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
27977 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
27978 case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
27979 case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
27981 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
27983 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
27985 case ISD::FMAXNUM: return performFMinNumFMaxNumCombine(N, DAG,
27987 case X86ISD::FAND: return PerformFANDCombine(N, DAG, Subtarget);
27988 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG, Subtarget);
27989 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
27990 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
27991 case ISD::ANY_EXTEND:
27992 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
27993 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
27994 case ISD::SIGN_EXTEND_INREG:
27995 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
27996 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
27997 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
27998 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
27999 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
28000 case X86ISD::SHUFP: // Handle all target specific shuffles
28001 case X86ISD::PALIGNR:
28002 case X86ISD::BLENDI:
28003 case X86ISD::UNPCKH:
28004 case X86ISD::UNPCKL:
28005 case X86ISD::MOVHLPS:
28006 case X86ISD::MOVLHPS:
28007 case X86ISD::PSHUFB:
28008 case X86ISD::PSHUFD:
28009 case X86ISD::PSHUFHW:
28010 case X86ISD::PSHUFLW:
28011 case X86ISD::MOVSS:
28012 case X86ISD::MOVSD:
28013 case X86ISD::VPERMILPI:
28014 case X86ISD::VPERM2X128:
28015 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
28016 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
28018 case ISD::MSCATTER: return PerformGatherScatterCombine(N, DAG);
28024 /// isTypeDesirableForOp - Return true if the target has native support for
28025 /// the specified value type and it is 'desirable' to use the type for the
28026 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
28027 /// instruction encodings are longer and some i16 instructions are slow.
28028 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
28029 if (!isTypeLegal(VT))
28031 if (VT != MVT::i16)
28038 case ISD::SIGN_EXTEND:
28039 case ISD::ZERO_EXTEND:
28040 case ISD::ANY_EXTEND:
28053 /// This function checks if any of the users of EFLAGS copies the EFLAGS. We
28054 /// know that the code that lowers COPY of EFLAGS has to use the stack, and if
28055 /// we don't adjust the stack we clobber the first frame index.
28056 /// See X86InstrInfo::copyPhysReg.
28057 bool X86TargetLowering::hasCopyImplyingStackAdjustment(
28058 MachineFunction *MF) const {
28059 const MachineRegisterInfo &MRI = MF->getRegInfo();
28061 return any_of(MRI.reg_instructions(X86::EFLAGS),
28062 [](const MachineInstr &RI) { return RI.isCopy(); });
28065 /// IsDesirableToPromoteOp - This method query the target whether it is
28066 /// beneficial for dag combiner to promote the specified node. If true, it
28067 /// should return the desired promotion type by reference.
28068 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
28069 EVT VT = Op.getValueType();
28070 if (VT != MVT::i16)
28073 bool Promote = false;
28074 bool Commute = false;
28075 switch (Op.getOpcode()) {
28078 LoadSDNode *LD = cast<LoadSDNode>(Op);
28079 // If the non-extending load has a single use and it's not live out, then it
28080 // might be folded.
28081 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
28082 Op.hasOneUse()*/) {
28083 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
28084 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
28085 // The only case where we'd want to promote LOAD (rather then it being
28086 // promoted as an operand is when it's only use is liveout.
28087 if (UI->getOpcode() != ISD::CopyToReg)
28094 case ISD::SIGN_EXTEND:
28095 case ISD::ZERO_EXTEND:
28096 case ISD::ANY_EXTEND:
28101 SDValue N0 = Op.getOperand(0);
28102 // Look out for (store (shl (load), x)).
28103 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
28116 SDValue N0 = Op.getOperand(0);
28117 SDValue N1 = Op.getOperand(1);
28118 if (!Commute && MayFoldLoad(N1))
28120 // Avoid disabling potential load folding opportunities.
28121 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
28123 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
28133 //===----------------------------------------------------------------------===//
28134 // X86 Inline Assembly Support
28135 //===----------------------------------------------------------------------===//
28137 // Helper to match a string separated by whitespace.
28138 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
28139 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
28141 for (StringRef Piece : Pieces) {
28142 if (!S.startswith(Piece)) // Check if the piece matches.
28145 S = S.substr(Piece.size());
28146 StringRef::size_type Pos = S.find_first_not_of(" \t");
28147 if (Pos == 0) // We matched a prefix.
28156 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
28158 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
28159 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
28160 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
28161 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
28163 if (AsmPieces.size() == 3)
28165 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
28172 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
28173 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
28175 std::string AsmStr = IA->getAsmString();
28177 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
28178 if (!Ty || Ty->getBitWidth() % 16 != 0)
28181 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
28182 SmallVector<StringRef, 4> AsmPieces;
28183 SplitString(AsmStr, AsmPieces, ";\n");
28185 switch (AsmPieces.size()) {
28186 default: return false;
28188 // FIXME: this should verify that we are targeting a 486 or better. If not,
28189 // we will turn this bswap into something that will be lowered to logical
28190 // ops instead of emitting the bswap asm. For now, we don't support 486 or
28191 // lower so don't worry about this.
28193 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
28194 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
28195 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
28196 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
28197 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
28198 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
28199 // No need to check constraints, nothing other than the equivalent of
28200 // "=r,0" would be valid here.
28201 return IntrinsicLowering::LowerToByteSwap(CI);
28204 // rorw $$8, ${0:w} --> llvm.bswap.i16
28205 if (CI->getType()->isIntegerTy(16) &&
28206 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28207 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
28208 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
28210 StringRef ConstraintsStr = IA->getConstraintString();
28211 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28212 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28213 if (clobbersFlagRegisters(AsmPieces))
28214 return IntrinsicLowering::LowerToByteSwap(CI);
28218 if (CI->getType()->isIntegerTy(32) &&
28219 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28220 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
28221 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
28222 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
28224 StringRef ConstraintsStr = IA->getConstraintString();
28225 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28226 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28227 if (clobbersFlagRegisters(AsmPieces))
28228 return IntrinsicLowering::LowerToByteSwap(CI);
28231 if (CI->getType()->isIntegerTy(64)) {
28232 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
28233 if (Constraints.size() >= 2 &&
28234 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
28235 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
28236 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
28237 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
28238 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
28239 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
28240 return IntrinsicLowering::LowerToByteSwap(CI);
28248 /// getConstraintType - Given a constraint letter, return the type of
28249 /// constraint it is for this target.
28250 X86TargetLowering::ConstraintType
28251 X86TargetLowering::getConstraintType(StringRef Constraint) const {
28252 if (Constraint.size() == 1) {
28253 switch (Constraint[0]) {
28264 return C_RegisterClass;
28288 return TargetLowering::getConstraintType(Constraint);
28291 /// Examine constraint type and operand type and determine a weight value.
28292 /// This object must already have been set up with the operand type
28293 /// and the current alternative constraint selected.
28294 TargetLowering::ConstraintWeight
28295 X86TargetLowering::getSingleConstraintMatchWeight(
28296 AsmOperandInfo &info, const char *constraint) const {
28297 ConstraintWeight weight = CW_Invalid;
28298 Value *CallOperandVal = info.CallOperandVal;
28299 // If we don't have a value, we can't do a match,
28300 // but allow it at the lowest weight.
28301 if (!CallOperandVal)
28303 Type *type = CallOperandVal->getType();
28304 // Look at the constraint type.
28305 switch (*constraint) {
28307 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
28318 if (CallOperandVal->getType()->isIntegerTy())
28319 weight = CW_SpecificReg;
28324 if (type->isFloatingPointTy())
28325 weight = CW_SpecificReg;
28328 if (type->isX86_MMXTy() && Subtarget->hasMMX())
28329 weight = CW_SpecificReg;
28333 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
28334 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
28335 weight = CW_Register;
28338 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
28339 if (C->getZExtValue() <= 31)
28340 weight = CW_Constant;
28344 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28345 if (C->getZExtValue() <= 63)
28346 weight = CW_Constant;
28350 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28351 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
28352 weight = CW_Constant;
28356 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28357 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
28358 weight = CW_Constant;
28362 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28363 if (C->getZExtValue() <= 3)
28364 weight = CW_Constant;
28368 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28369 if (C->getZExtValue() <= 0xff)
28370 weight = CW_Constant;
28375 if (isa<ConstantFP>(CallOperandVal)) {
28376 weight = CW_Constant;
28380 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28381 if ((C->getSExtValue() >= -0x80000000LL) &&
28382 (C->getSExtValue() <= 0x7fffffffLL))
28383 weight = CW_Constant;
28387 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28388 if (C->getZExtValue() <= 0xffffffff)
28389 weight = CW_Constant;
28396 /// LowerXConstraint - try to replace an X constraint, which matches anything,
28397 /// with another that has more specific requirements based on the type of the
28398 /// corresponding operand.
28399 const char *X86TargetLowering::
28400 LowerXConstraint(EVT ConstraintVT) const {
28401 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
28402 // 'f' like normal targets.
28403 if (ConstraintVT.isFloatingPoint()) {
28404 if (Subtarget->hasSSE2())
28406 if (Subtarget->hasSSE1())
28410 return TargetLowering::LowerXConstraint(ConstraintVT);
28413 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
28414 /// vector. If it is invalid, don't add anything to Ops.
28415 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
28416 std::string &Constraint,
28417 std::vector<SDValue>&Ops,
28418 SelectionDAG &DAG) const {
28421 // Only support length 1 constraints for now.
28422 if (Constraint.length() > 1) return;
28424 char ConstraintLetter = Constraint[0];
28425 switch (ConstraintLetter) {
28428 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28429 if (C->getZExtValue() <= 31) {
28430 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28431 Op.getValueType());
28437 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28438 if (C->getZExtValue() <= 63) {
28439 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28440 Op.getValueType());
28446 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28447 if (isInt<8>(C->getSExtValue())) {
28448 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28449 Op.getValueType());
28455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28456 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
28457 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
28458 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
28459 Op.getValueType());
28465 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28466 if (C->getZExtValue() <= 3) {
28467 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28468 Op.getValueType());
28474 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28475 if (C->getZExtValue() <= 255) {
28476 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28477 Op.getValueType());
28483 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28484 if (C->getZExtValue() <= 127) {
28485 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28486 Op.getValueType());
28492 // 32-bit signed value
28493 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28494 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28495 C->getSExtValue())) {
28496 // Widen to 64 bits here to get it sign extended.
28497 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
28500 // FIXME gcc accepts some relocatable values here too, but only in certain
28501 // memory models; it's complicated.
28506 // 32-bit unsigned value
28507 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28508 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28509 C->getZExtValue())) {
28510 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28511 Op.getValueType());
28515 // FIXME gcc accepts some relocatable values here too, but only in certain
28516 // memory models; it's complicated.
28520 // Literal immediates are always ok.
28521 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
28522 // Widen to 64 bits here to get it sign extended.
28523 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
28527 // In any sort of PIC mode addresses need to be computed at runtime by
28528 // adding in a register or some sort of table lookup. These can't
28529 // be used as immediates.
28530 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
28533 // If we are in non-pic codegen mode, we allow the address of a global (with
28534 // an optional displacement) to be used with 'i'.
28535 GlobalAddressSDNode *GA = nullptr;
28536 int64_t Offset = 0;
28538 // Match either (GA), (GA+C), (GA+C1+C2), etc.
28540 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
28541 Offset += GA->getOffset();
28543 } else if (Op.getOpcode() == ISD::ADD) {
28544 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28545 Offset += C->getZExtValue();
28546 Op = Op.getOperand(0);
28549 } else if (Op.getOpcode() == ISD::SUB) {
28550 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28551 Offset += -C->getZExtValue();
28552 Op = Op.getOperand(0);
28557 // Otherwise, this isn't something we can handle, reject it.
28561 const GlobalValue *GV = GA->getGlobal();
28562 // If we require an extra load to get this address, as in PIC mode, we
28563 // can't accept it.
28564 if (isGlobalStubReference(
28565 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
28568 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
28569 GA->getValueType(0), Offset);
28574 if (Result.getNode()) {
28575 Ops.push_back(Result);
28578 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
28581 std::pair<unsigned, const TargetRegisterClass *>
28582 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
28583 StringRef Constraint,
28585 // First, see if this is a constraint that directly corresponds to an LLVM
28587 if (Constraint.size() == 1) {
28588 // GCC Constraint Letters
28589 switch (Constraint[0]) {
28591 // TODO: Slight differences here in allocation order and leaving
28592 // RIP in the class. Do they matter any more here than they do
28593 // in the normal allocation?
28594 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
28595 if (Subtarget->is64Bit()) {
28596 if (VT == MVT::i32 || VT == MVT::f32)
28597 return std::make_pair(0U, &X86::GR32RegClass);
28598 if (VT == MVT::i16)
28599 return std::make_pair(0U, &X86::GR16RegClass);
28600 if (VT == MVT::i8 || VT == MVT::i1)
28601 return std::make_pair(0U, &X86::GR8RegClass);
28602 if (VT == MVT::i64 || VT == MVT::f64)
28603 return std::make_pair(0U, &X86::GR64RegClass);
28606 // 32-bit fallthrough
28607 case 'Q': // Q_REGS
28608 if (VT == MVT::i32 || VT == MVT::f32)
28609 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
28610 if (VT == MVT::i16)
28611 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
28612 if (VT == MVT::i8 || VT == MVT::i1)
28613 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
28614 if (VT == MVT::i64)
28615 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
28617 case 'r': // GENERAL_REGS
28618 case 'l': // INDEX_REGS
28619 if (VT == MVT::i8 || VT == MVT::i1)
28620 return std::make_pair(0U, &X86::GR8RegClass);
28621 if (VT == MVT::i16)
28622 return std::make_pair(0U, &X86::GR16RegClass);
28623 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
28624 return std::make_pair(0U, &X86::GR32RegClass);
28625 return std::make_pair(0U, &X86::GR64RegClass);
28626 case 'R': // LEGACY_REGS
28627 if (VT == MVT::i8 || VT == MVT::i1)
28628 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
28629 if (VT == MVT::i16)
28630 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
28631 if (VT == MVT::i32 || !Subtarget->is64Bit())
28632 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
28633 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
28634 case 'f': // FP Stack registers.
28635 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
28636 // value to the correct fpstack register class.
28637 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
28638 return std::make_pair(0U, &X86::RFP32RegClass);
28639 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
28640 return std::make_pair(0U, &X86::RFP64RegClass);
28641 return std::make_pair(0U, &X86::RFP80RegClass);
28642 case 'y': // MMX_REGS if MMX allowed.
28643 if (!Subtarget->hasMMX()) break;
28644 return std::make_pair(0U, &X86::VR64RegClass);
28645 case 'Y': // SSE_REGS if SSE2 allowed
28646 if (!Subtarget->hasSSE2()) break;
28648 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
28649 if (!Subtarget->hasSSE1()) break;
28651 switch (VT.SimpleTy) {
28653 // Scalar SSE types.
28656 return std::make_pair(0U, &X86::FR32RegClass);
28659 return std::make_pair(0U, &X86::FR64RegClass);
28660 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28668 return std::make_pair(0U, &X86::VR128RegClass);
28676 return std::make_pair(0U, &X86::VR256RegClass);
28681 return std::make_pair(0U, &X86::VR512RegClass);
28687 // Use the default implementation in TargetLowering to convert the register
28688 // constraint into a member of a register class.
28689 std::pair<unsigned, const TargetRegisterClass*> Res;
28690 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
28692 // Not found as a standard register?
28694 // Map st(0) -> st(7) -> ST0
28695 if (Constraint.size() == 7 && Constraint[0] == '{' &&
28696 tolower(Constraint[1]) == 's' &&
28697 tolower(Constraint[2]) == 't' &&
28698 Constraint[3] == '(' &&
28699 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
28700 Constraint[5] == ')' &&
28701 Constraint[6] == '}') {
28703 Res.first = X86::FP0+Constraint[4]-'0';
28704 Res.second = &X86::RFP80RegClass;
28708 // GCC allows "st(0)" to be called just plain "st".
28709 if (StringRef("{st}").equals_lower(Constraint)) {
28710 Res.first = X86::FP0;
28711 Res.second = &X86::RFP80RegClass;
28716 if (StringRef("{flags}").equals_lower(Constraint)) {
28717 Res.first = X86::EFLAGS;
28718 Res.second = &X86::CCRRegClass;
28722 // 'A' means EAX + EDX.
28723 if (Constraint == "A") {
28724 Res.first = X86::EAX;
28725 Res.second = &X86::GR32_ADRegClass;
28731 // Otherwise, check to see if this is a register class of the wrong value
28732 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
28733 // turn into {ax},{dx}.
28734 // MVT::Other is used to specify clobber names.
28735 if (Res.second->hasType(VT) || VT == MVT::Other)
28736 return Res; // Correct type already, nothing to do.
28738 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
28739 // return "eax". This should even work for things like getting 64bit integer
28740 // registers when given an f64 type.
28741 const TargetRegisterClass *Class = Res.second;
28742 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
28743 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
28744 unsigned Size = VT.getSizeInBits();
28745 if (Size == 1) Size = 8;
28746 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, Size);
28748 Res.first = DestReg;
28749 Res.second = Size == 8 ? &X86::GR8RegClass
28750 : Size == 16 ? &X86::GR16RegClass
28751 : Size == 32 ? &X86::GR32RegClass
28752 : &X86::GR64RegClass;
28753 assert(Res.second->contains(Res.first) && "Register in register class");
28755 // No register found/type mismatch.
28757 Res.second = nullptr;
28759 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
28760 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
28761 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
28762 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
28763 Class == &X86::VR512RegClass) {
28764 // Handle references to XMM physical registers that got mapped into the
28765 // wrong class. This can happen with constraints like {xmm0} where the
28766 // target independent register mapper will just pick the first match it can
28767 // find, ignoring the required type.
28769 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28770 if (VT == MVT::f32 || VT == MVT::i32)
28771 Res.second = &X86::FR32RegClass;
28772 else if (VT == MVT::f64 || VT == MVT::i64)
28773 Res.second = &X86::FR64RegClass;
28774 else if (X86::VR128RegClass.hasType(VT))
28775 Res.second = &X86::VR128RegClass;
28776 else if (X86::VR256RegClass.hasType(VT))
28777 Res.second = &X86::VR256RegClass;
28778 else if (X86::VR512RegClass.hasType(VT))
28779 Res.second = &X86::VR512RegClass;
28781 // Type mismatch and not a clobber: Return an error;
28783 Res.second = nullptr;
28790 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
28791 const AddrMode &AM, Type *Ty,
28792 unsigned AS) const {
28793 // Scaling factors are not free at all.
28794 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
28795 // will take 2 allocations in the out of order engine instead of 1
28796 // for plain addressing mode, i.e. inst (reg1).
28798 // vaddps (%rsi,%drx), %ymm0, %ymm1
28799 // Requires two allocations (one for the load, one for the computation)
28801 // vaddps (%rsi), %ymm0, %ymm1
28802 // Requires just 1 allocation, i.e., freeing allocations for other operations
28803 // and having less micro operations to execute.
28805 // For some X86 architectures, this is even worse because for instance for
28806 // stores, the complex addressing mode forces the instruction to use the
28807 // "load" ports instead of the dedicated "store" port.
28808 // E.g., on Haswell:
28809 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
28810 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
28811 if (isLegalAddressingMode(DL, AM, Ty, AS))
28812 // Scale represents reg2 * scale, thus account for 1
28813 // as soon as we use a second register.
28814 return AM.Scale != 0;
28818 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
28819 // Integer division on x86 is expensive. However, when aggressively optimizing
28820 // for code size, we prefer to use a div instruction, as it is usually smaller
28821 // than the alternative sequence.
28822 // The exception to this is vector division. Since x86 doesn't have vector
28823 // integer division, leaving the division as-is is a loss even in terms of
28824 // size, because it will have to be scalarized, while the alternative code
28825 // sequence can be performed in vector form.
28826 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
28827 Attribute::MinSize);
28828 return OptSize && !VT.isVector();