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);
270 // Scalar integer divide and remainder are lowered to use operations that
271 // produce two results, to match the available instructions. This exposes
272 // the two-result form to trivial CSE, which is able to combine x/y and x%y
273 // into a single instruction.
275 // Scalar integer multiply-high is also lowered to use two-result
276 // operations, to match the available instructions. However, plain multiply
277 // (low) operations are left as Legal, as there are single-result
278 // instructions for this in x86. Using the two-result multiply instructions
279 // when both high and low results are needed must be arranged by dagcombine.
280 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
281 setOperationAction(ISD::MULHS, VT, Expand);
282 setOperationAction(ISD::MULHU, VT, Expand);
283 setOperationAction(ISD::SDIV, VT, Expand);
284 setOperationAction(ISD::UDIV, VT, Expand);
285 setOperationAction(ISD::SREM, VT, Expand);
286 setOperationAction(ISD::UREM, VT, Expand);
288 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
289 setOperationAction(ISD::ADDC, VT, Custom);
290 setOperationAction(ISD::ADDE, VT, Custom);
291 setOperationAction(ISD::SUBC, VT, Custom);
292 setOperationAction(ISD::SUBE, VT, Custom);
295 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
296 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
297 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
298 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
299 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
300 setOperationAction(ISD::BR_CC , MVT::f128, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
303 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
304 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::f128, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
312 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
313 if (Subtarget->is64Bit())
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
318 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
320 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
321 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
322 // is. We should promote the value to 64-bits to solve this.
323 // This is what the CRT headers do - `fmodf` is an inline header
324 // function casting to f64 and calling `fmod`.
325 setOperationAction(ISD::FREM , MVT::f32 , Promote);
327 setOperationAction(ISD::FREM , MVT::f32 , Expand);
330 setOperationAction(ISD::FREM , MVT::f64 , Expand);
331 setOperationAction(ISD::FREM , MVT::f80 , Expand);
332 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
334 // Promote the i8 variants and force them on up to i32 which has a shorter
336 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
337 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
338 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
339 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
340 if (Subtarget->hasBMI()) {
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
342 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
343 if (Subtarget->is64Bit())
344 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
347 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
348 if (Subtarget->is64Bit())
349 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
352 if (Subtarget->hasLZCNT()) {
353 // When promoting the i8 variants, force them to i32 for a shorter
355 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
356 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
358 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
361 if (Subtarget->is64Bit())
362 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
364 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
365 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
366 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
367 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
368 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
369 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
370 if (Subtarget->is64Bit()) {
371 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
372 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
376 // Special handling for half-precision floating point conversions.
377 // If we don't have F16C support, then lower half float conversions
378 // into library calls.
379 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
380 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
381 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
384 // There's never any support for operations beyond MVT::f32.
385 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
386 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
387 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
388 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
390 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
391 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
392 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
393 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
394 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
395 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
397 if (Subtarget->hasPOPCNT()) {
398 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
400 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
401 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
402 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
403 if (Subtarget->is64Bit())
404 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
407 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
409 if (!Subtarget->hasMOVBE())
410 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
412 // These should be promoted to a larger select which is supported.
413 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
414 // X86 wants to expand cmov itself.
415 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
416 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
417 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
418 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
419 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
420 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
421 setOperationAction(ISD::SELECT , MVT::f128 , Custom);
422 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
423 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
424 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
425 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
426 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
427 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
428 setOperationAction(ISD::SETCC , MVT::f128 , Custom);
429 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
431 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
434 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
435 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
437 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
438 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
439 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
440 // support continuation, user-level threading, and etc.. As a result, no
441 // other SjLj exception interfaces are implemented and please don't build
442 // your own exception handling based on them.
443 // LLVM/Clang supports zero-cost DWARF exception handling.
444 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
445 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
448 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
449 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
450 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
451 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
452 if (Subtarget->is64Bit())
453 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
454 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
455 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
456 if (Subtarget->is64Bit()) {
457 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
458 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
459 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
460 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
461 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
463 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
464 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
465 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
466 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
467 if (Subtarget->is64Bit()) {
468 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
469 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
470 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
473 if (Subtarget->hasSSE1())
474 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
476 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
478 // Expand certain atomics
479 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
480 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
481 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
482 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
485 if (Subtarget->hasCmpxchg16b()) {
486 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
489 // FIXME - use subtarget debug flags
490 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
491 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
492 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
495 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
496 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
498 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
499 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
501 setOperationAction(ISD::TRAP, MVT::Other, Legal);
502 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
504 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
505 setOperationAction(ISD::VASTART , MVT::Other, Custom);
506 setOperationAction(ISD::VAEND , MVT::Other, Expand);
507 if (Subtarget->is64Bit()) {
508 setOperationAction(ISD::VAARG , MVT::Other, Custom);
509 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
511 // TargetInfo::CharPtrBuiltinVaList
512 setOperationAction(ISD::VAARG , MVT::Other, Expand);
513 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
516 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
517 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
519 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
521 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
522 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
523 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
525 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
526 // f32 and f64 use SSE.
527 // Set up the FP register classes.
528 addRegisterClass(MVT::f32, &X86::FR32RegClass);
529 addRegisterClass(MVT::f64, &X86::FR64RegClass);
531 // Use ANDPD to simulate FABS.
532 setOperationAction(ISD::FABS , MVT::f64, Custom);
533 setOperationAction(ISD::FABS , MVT::f32, Custom);
535 // Use XORP to simulate FNEG.
536 setOperationAction(ISD::FNEG , MVT::f64, Custom);
537 setOperationAction(ISD::FNEG , MVT::f32, Custom);
539 // Use ANDPD and ORPD to simulate FCOPYSIGN.
540 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
541 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
543 // Lower this to FGETSIGNx86 plus an AND.
544 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
545 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
547 // We don't support sin/cos/fmod
548 setOperationAction(ISD::FSIN , MVT::f64, Expand);
549 setOperationAction(ISD::FCOS , MVT::f64, Expand);
550 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
551 setOperationAction(ISD::FSIN , MVT::f32, Expand);
552 setOperationAction(ISD::FCOS , MVT::f32, Expand);
553 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
555 // Expand FP immediates into loads from the stack, except for the special
557 addLegalFPImmediate(APFloat(+0.0)); // xorpd
558 addLegalFPImmediate(APFloat(+0.0f)); // xorps
559 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
560 // Use SSE for f32, x87 for f64.
561 // Set up the FP register classes.
562 addRegisterClass(MVT::f32, &X86::FR32RegClass);
563 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
565 // Use ANDPS to simulate FABS.
566 setOperationAction(ISD::FABS , MVT::f32, Custom);
568 // Use XORP to simulate FNEG.
569 setOperationAction(ISD::FNEG , MVT::f32, Custom);
571 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
573 // Use ANDPS and ORPS to simulate FCOPYSIGN.
574 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
575 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
577 // We don't support sin/cos/fmod
578 setOperationAction(ISD::FSIN , MVT::f32, Expand);
579 setOperationAction(ISD::FCOS , MVT::f32, Expand);
580 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
582 // Special cases we handle for FP constants.
583 addLegalFPImmediate(APFloat(+0.0f)); // xorps
584 addLegalFPImmediate(APFloat(+0.0)); // FLD0
585 addLegalFPImmediate(APFloat(+1.0)); // FLD1
586 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
587 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
589 if (!TM.Options.UnsafeFPMath) {
590 setOperationAction(ISD::FSIN , MVT::f64, Expand);
591 setOperationAction(ISD::FCOS , MVT::f64, Expand);
592 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
594 } else if (!Subtarget->useSoftFloat()) {
595 // f32 and f64 in x87.
596 // Set up the FP register classes.
597 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
598 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
600 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
601 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
605 if (!TM.Options.UnsafeFPMath) {
606 setOperationAction(ISD::FSIN , MVT::f64, Expand);
607 setOperationAction(ISD::FSIN , MVT::f32, Expand);
608 setOperationAction(ISD::FCOS , MVT::f64, Expand);
609 setOperationAction(ISD::FCOS , MVT::f32, Expand);
610 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
611 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
613 addLegalFPImmediate(APFloat(+0.0)); // FLD0
614 addLegalFPImmediate(APFloat(+1.0)); // FLD1
615 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
616 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
617 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
618 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
619 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
620 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
623 // We don't support FMA.
624 setOperationAction(ISD::FMA, MVT::f64, Expand);
625 setOperationAction(ISD::FMA, MVT::f32, Expand);
627 // Long double always uses X87, except f128 in MMX.
628 if (!Subtarget->useSoftFloat()) {
629 if (Subtarget->is64Bit() && Subtarget->hasMMX()) {
630 addRegisterClass(MVT::f128, &X86::FR128RegClass);
631 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
632 setOperationAction(ISD::FABS , MVT::f128, Custom);
633 setOperationAction(ISD::FNEG , MVT::f128, Custom);
634 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
637 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
638 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
639 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
641 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
642 addLegalFPImmediate(TmpFlt); // FLD0
644 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
647 APFloat TmpFlt2(+1.0);
648 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
650 addLegalFPImmediate(TmpFlt2); // FLD1
651 TmpFlt2.changeSign();
652 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
655 if (!TM.Options.UnsafeFPMath) {
656 setOperationAction(ISD::FSIN , MVT::f80, Expand);
657 setOperationAction(ISD::FCOS , MVT::f80, Expand);
658 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
661 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
662 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
663 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
664 setOperationAction(ISD::FRINT, MVT::f80, Expand);
665 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
666 setOperationAction(ISD::FMA, MVT::f80, Expand);
669 // Always use a library call for pow.
670 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
671 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
672 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
674 setOperationAction(ISD::FLOG, MVT::f80, Expand);
675 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
676 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
677 setOperationAction(ISD::FEXP, MVT::f80, Expand);
678 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
679 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
680 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
682 // First set operation action for all vector types to either promote
683 // (for widening) or expand (for scalarization). Then we will selectively
684 // turn on ones that can be effectively codegen'd.
685 for (MVT VT : MVT::vector_valuetypes()) {
686 setOperationAction(ISD::ADD , VT, Expand);
687 setOperationAction(ISD::SUB , VT, Expand);
688 setOperationAction(ISD::FADD, VT, Expand);
689 setOperationAction(ISD::FNEG, VT, Expand);
690 setOperationAction(ISD::FSUB, VT, Expand);
691 setOperationAction(ISD::MUL , VT, Expand);
692 setOperationAction(ISD::FMUL, VT, Expand);
693 setOperationAction(ISD::SDIV, VT, Expand);
694 setOperationAction(ISD::UDIV, VT, Expand);
695 setOperationAction(ISD::FDIV, VT, Expand);
696 setOperationAction(ISD::SREM, VT, Expand);
697 setOperationAction(ISD::UREM, VT, Expand);
698 setOperationAction(ISD::LOAD, VT, Expand);
699 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
701 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
702 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
703 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
704 setOperationAction(ISD::FABS, VT, Expand);
705 setOperationAction(ISD::FSIN, VT, Expand);
706 setOperationAction(ISD::FSINCOS, VT, Expand);
707 setOperationAction(ISD::FCOS, VT, Expand);
708 setOperationAction(ISD::FSINCOS, VT, Expand);
709 setOperationAction(ISD::FREM, VT, Expand);
710 setOperationAction(ISD::FMA, VT, Expand);
711 setOperationAction(ISD::FPOWI, VT, Expand);
712 setOperationAction(ISD::FSQRT, VT, Expand);
713 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
714 setOperationAction(ISD::FFLOOR, VT, Expand);
715 setOperationAction(ISD::FCEIL, VT, Expand);
716 setOperationAction(ISD::FTRUNC, VT, Expand);
717 setOperationAction(ISD::FRINT, VT, Expand);
718 setOperationAction(ISD::FNEARBYINT, VT, Expand);
719 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
720 setOperationAction(ISD::MULHS, VT, Expand);
721 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
722 setOperationAction(ISD::MULHU, VT, Expand);
723 setOperationAction(ISD::SDIVREM, VT, Expand);
724 setOperationAction(ISD::UDIVREM, VT, Expand);
725 setOperationAction(ISD::FPOW, VT, Expand);
726 setOperationAction(ISD::CTPOP, VT, Expand);
727 setOperationAction(ISD::CTTZ, VT, Expand);
728 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
729 setOperationAction(ISD::CTLZ, VT, Expand);
730 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
731 setOperationAction(ISD::SHL, VT, Expand);
732 setOperationAction(ISD::SRA, VT, Expand);
733 setOperationAction(ISD::SRL, VT, Expand);
734 setOperationAction(ISD::ROTL, VT, Expand);
735 setOperationAction(ISD::ROTR, VT, Expand);
736 setOperationAction(ISD::BSWAP, VT, Expand);
737 setOperationAction(ISD::SETCC, VT, Expand);
738 setOperationAction(ISD::FLOG, VT, Expand);
739 setOperationAction(ISD::FLOG2, VT, Expand);
740 setOperationAction(ISD::FLOG10, VT, Expand);
741 setOperationAction(ISD::FEXP, VT, Expand);
742 setOperationAction(ISD::FEXP2, VT, Expand);
743 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
744 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
745 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
746 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
747 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
748 setOperationAction(ISD::TRUNCATE, VT, Expand);
749 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
750 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
751 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
752 setOperationAction(ISD::VSELECT, VT, Expand);
753 setOperationAction(ISD::SELECT_CC, VT, Expand);
754 for (MVT InnerVT : MVT::vector_valuetypes()) {
755 setTruncStoreAction(InnerVT, VT, Expand);
757 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
758 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
760 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
761 // types, we have to deal with them whether we ask for Expansion or not.
762 // Setting Expand causes its own optimisation problems though, so leave
764 if (VT.getVectorElementType() == MVT::i1)
765 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
767 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
768 // split/scalarized right now.
769 if (VT.getVectorElementType() == MVT::f16)
770 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
774 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
775 // with -msoft-float, disable use of MMX as well.
776 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
777 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
778 // No operations on x86mmx supported, everything uses intrinsics.
781 // MMX-sized vectors (other than x86mmx) are expected to be expanded
782 // into smaller operations.
783 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
784 setOperationAction(ISD::MULHS, MMXTy, Expand);
785 setOperationAction(ISD::AND, MMXTy, Expand);
786 setOperationAction(ISD::OR, MMXTy, Expand);
787 setOperationAction(ISD::XOR, MMXTy, Expand);
788 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
789 setOperationAction(ISD::SELECT, MMXTy, Expand);
790 setOperationAction(ISD::BITCAST, MMXTy, Expand);
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
794 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
795 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
797 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
798 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
799 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
800 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
801 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
802 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
803 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
804 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
805 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
806 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
807 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
808 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
809 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
810 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
813 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
814 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
816 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
817 // registers cannot be used even for integer operations.
818 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
819 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
820 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
821 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
823 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
824 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
825 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
826 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
827 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
828 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
829 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
830 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
831 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
832 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
833 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
834 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
835 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
836 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
837 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
838 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
839 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
840 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
841 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
842 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
843 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
844 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
845 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
847 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
848 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
849 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
850 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
852 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
853 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
854 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
855 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
857 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
858 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
859 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
861 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
863 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
864 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
865 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
866 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
868 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
869 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
870 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
871 // ISD::CTTZ v2i64 - scalarization is faster.
872 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
873 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
874 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
875 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
877 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
878 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
879 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
880 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
881 setOperationAction(ISD::VSELECT, VT, Custom);
882 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
885 // We support custom legalizing of sext and anyext loads for specific
886 // memory vector types which we can load as a scalar (or sequence of
887 // scalars) and extend in-register to a legal 128-bit vector type. For sext
888 // loads these must work with a single scalar load.
889 for (MVT VT : MVT::integer_vector_valuetypes()) {
890 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
891 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
892 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
893 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
894 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
895 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
896 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
897 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
898 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
901 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
902 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
903 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
904 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
905 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
906 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
907 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
908 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
910 if (Subtarget->is64Bit()) {
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
915 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
916 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
917 setOperationAction(ISD::AND, VT, Promote);
918 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
919 setOperationAction(ISD::OR, VT, Promote);
920 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
921 setOperationAction(ISD::XOR, VT, Promote);
922 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
923 setOperationAction(ISD::LOAD, VT, Promote);
924 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
925 setOperationAction(ISD::SELECT, VT, Promote);
926 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
929 // Custom lower v2i64 and v2f64 selects.
930 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
931 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
932 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
933 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
935 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
936 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
938 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
940 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
941 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
942 // As there is no 64-bit GPR available, we need build a special custom
943 // sequence to convert from v2i32 to v2f32.
944 if (!Subtarget->is64Bit())
945 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
947 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
948 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
950 for (MVT VT : MVT::fp_vector_valuetypes())
951 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
953 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
954 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
955 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
958 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
959 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
960 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
961 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
962 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
963 setOperationAction(ISD::FRINT, RoundedTy, Legal);
964 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
967 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
968 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
969 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
970 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
971 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
972 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
973 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
974 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
976 // FIXME: Do we need to handle scalar-to-vector here?
977 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
979 // We directly match byte blends in the backend as they match the VSELECT
981 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
983 // SSE41 brings specific instructions for doing vector sign extend even in
984 // cases where we don't have SRA.
985 for (MVT VT : MVT::integer_vector_valuetypes()) {
986 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
988 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
991 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
997 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1004 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1006 // i8 and i16 vectors are custom because the source register and source
1007 // source memory operand types are not the same width. f32 vectors are
1008 // custom since the immediate controlling the insert encodes additional
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1013 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1015 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1018 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1020 // FIXME: these should be Legal, but that's only for the case where
1021 // the index is constant. For now custom expand to deal with that.
1022 if (Subtarget->is64Bit()) {
1023 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1024 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1028 if (Subtarget->hasSSE2()) {
1029 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1031 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1034 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1036 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1037 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1039 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1040 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1042 // In the customized shift lowering, the legal cases in AVX2 will be
1044 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1045 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1047 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1048 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1050 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1051 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1054 if (Subtarget->hasXOP()) {
1055 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1062 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1065 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1066 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1071 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1073 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1074 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1075 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1077 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1087 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1088 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1090 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1100 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1101 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1103 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1104 // even though v8i16 is a legal type.
1105 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1106 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1107 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1109 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1110 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1111 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1113 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1114 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1116 for (MVT VT : MVT::fp_vector_valuetypes())
1117 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1119 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1120 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1122 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1123 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1125 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1126 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1128 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1131 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1133 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1135 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1137 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1139 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1140 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1142 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1143 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1145 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1146 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1148 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1150 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1153 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1155 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1158 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1162 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1164 if (Subtarget->hasAnyFMA()) {
1165 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1166 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1167 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1168 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1169 setOperationAction(ISD::FMA, MVT::f32, Legal);
1170 setOperationAction(ISD::FMA, MVT::f64, Legal);
1173 if (Subtarget->hasInt256()) {
1174 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1175 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1176 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1177 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1179 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1180 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1181 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1182 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1184 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1185 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1186 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1187 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1189 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1190 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1191 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1192 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1194 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1196 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1197 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1199 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1200 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1202 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1203 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1205 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1207 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1208 // when we have a 256bit-wide blend with immediate.
1209 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1211 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1217 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1224 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1226 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1227 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1228 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1229 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1231 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1232 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1233 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1234 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1236 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1237 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1238 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1239 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1241 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1243 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1244 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1246 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1247 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1249 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1250 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1252 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1255 // In the customized shift lowering, the legal cases in AVX2 will be
1257 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1258 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1260 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1261 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1263 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1264 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1266 // Custom lower several nodes for 256-bit types.
1267 for (MVT VT : MVT::vector_valuetypes()) {
1268 if (VT.getScalarSizeInBits() >= 32) {
1269 setOperationAction(ISD::MLOAD, VT, Legal);
1270 setOperationAction(ISD::MSTORE, VT, Legal);
1272 // Extract subvector is special because the value type
1273 // (result) is 128-bit but the source is 256-bit wide.
1274 if (VT.is128BitVector()) {
1275 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1277 // Do not attempt to custom lower other non-256-bit vectors
1278 if (!VT.is256BitVector())
1281 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1282 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1283 setOperationAction(ISD::VSELECT, VT, Custom);
1284 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1285 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1286 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1287 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1288 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1291 if (Subtarget->hasInt256())
1292 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1294 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1295 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1296 setOperationAction(ISD::AND, VT, Promote);
1297 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1298 setOperationAction(ISD::OR, VT, Promote);
1299 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1300 setOperationAction(ISD::XOR, VT, Promote);
1301 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1302 setOperationAction(ISD::LOAD, VT, Promote);
1303 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1304 setOperationAction(ISD::SELECT, VT, Promote);
1305 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1309 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1310 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1311 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1312 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1313 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1315 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1316 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1317 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1319 for (MVT VT : MVT::fp_vector_valuetypes())
1320 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1322 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1323 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1324 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1325 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1326 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1327 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1328 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1329 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1330 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1331 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1332 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1333 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1335 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1336 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1337 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1338 setOperationAction(ISD::XOR, MVT::i1, Legal);
1339 setOperationAction(ISD::OR, MVT::i1, Legal);
1340 setOperationAction(ISD::AND, MVT::i1, Legal);
1341 setOperationAction(ISD::SUB, MVT::i1, Custom);
1342 setOperationAction(ISD::ADD, MVT::i1, Custom);
1343 setOperationAction(ISD::MUL, MVT::i1, Custom);
1344 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1345 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1346 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1348 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1350 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1351 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1355 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1356 setOperationAction(ISD::FABS, MVT::v16f32, Custom);
1358 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1359 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1363 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1364 setOperationAction(ISD::FABS, MVT::v8f64, Custom);
1365 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1366 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1368 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1369 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1371 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1376 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1377 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1381 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1382 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1383 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1385 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1387 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1388 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1389 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1390 if (Subtarget->hasVLX()){
1391 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1393 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1394 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1395 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1397 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1399 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1400 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1401 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1403 setOperationAction(ISD::MLOAD, MVT::v8i32, Custom);
1404 setOperationAction(ISD::MLOAD, MVT::v8f32, Custom);
1405 setOperationAction(ISD::MSTORE, MVT::v8i32, Custom);
1406 setOperationAction(ISD::MSTORE, MVT::v8f32, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1409 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1410 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1411 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1412 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1413 if (Subtarget->hasDQI()) {
1414 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1415 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1417 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1418 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1419 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1420 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1421 if (Subtarget->hasVLX()) {
1422 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1423 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1424 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1425 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1426 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1427 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1428 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1429 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1432 if (Subtarget->hasVLX()) {
1433 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1434 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1435 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1438 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1439 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1440 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1442 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1443 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1444 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1445 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1446 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1447 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1448 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1449 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1452 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1453 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1454 if (Subtarget->hasDQI()) {
1455 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1456 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1458 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1459 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1460 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1461 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1462 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1463 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1464 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1465 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1466 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1467 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
1475 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1476 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1478 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1480 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1482 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1484 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1485 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1486 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1490 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1491 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1493 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1494 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1495 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1496 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1497 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1498 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1499 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1500 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1502 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1503 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1505 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1506 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1508 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1510 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1511 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1513 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1514 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1516 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1517 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1519 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1520 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1521 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1522 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1523 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1524 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1526 if (Subtarget->hasCDI()) {
1527 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1528 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Expand);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Expand);
1532 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1533 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1535 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1536 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Expand);
1537 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Expand);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Expand);
1539 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Expand);
1541 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1542 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1544 if (Subtarget->hasVLX()) {
1545 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1546 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1547 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1548 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1549 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1550 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1551 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1552 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1554 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1555 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1556 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1557 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1559 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1560 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1561 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1562 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1563 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1564 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1565 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1566 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1568 } // Subtarget->hasCDI()
1570 if (Subtarget->hasDQI()) {
1571 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1572 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1573 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1575 // Custom lower several nodes.
1576 for (MVT VT : MVT::vector_valuetypes()) {
1577 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1579 setOperationAction(ISD::AND, VT, Legal);
1580 setOperationAction(ISD::OR, VT, Legal);
1581 setOperationAction(ISD::XOR, VT, Legal);
1583 if ((VT.is128BitVector() || VT.is256BitVector()) && EltSize >= 32) {
1584 setOperationAction(ISD::MGATHER, VT, Custom);
1585 setOperationAction(ISD::MSCATTER, VT, Custom);
1587 // Extract subvector is special because the value type
1588 // (result) is 256/128-bit but the source is 512-bit wide.
1589 if (VT.is128BitVector() || VT.is256BitVector()) {
1590 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1592 if (VT.getVectorElementType() == MVT::i1)
1593 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1595 // Do not attempt to custom lower other non-512-bit vectors
1596 if (!VT.is512BitVector())
1599 if (EltSize >= 32) {
1600 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1601 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1602 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1603 setOperationAction(ISD::VSELECT, VT, Legal);
1604 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1605 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1606 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1607 setOperationAction(ISD::MLOAD, VT, Legal);
1608 setOperationAction(ISD::MSTORE, VT, Legal);
1609 setOperationAction(ISD::MGATHER, VT, Legal);
1610 setOperationAction(ISD::MSCATTER, VT, Custom);
1613 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1614 setOperationAction(ISD::SELECT, VT, Promote);
1615 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1619 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1620 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1621 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1623 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1624 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1626 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1627 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1628 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1629 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1630 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1631 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1632 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1633 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1634 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1635 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1636 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1637 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1640 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1641 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1644 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1645 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1646 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1647 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1648 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1649 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1650 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1651 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1652 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1655 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1656 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1660 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1661 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1662 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1663 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1664 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1665 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1669 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1670 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1671 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1672 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1673 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1674 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1675 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1676 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1678 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1679 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1680 if (Subtarget->hasVLX())
1681 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1683 if (Subtarget->hasCDI()) {
1684 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1685 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1686 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Expand);
1687 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Expand);
1690 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1691 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1692 setOperationAction(ISD::VSELECT, VT, Legal);
1693 setOperationAction(ISD::SRL, VT, Custom);
1694 setOperationAction(ISD::SHL, VT, Custom);
1695 setOperationAction(ISD::SRA, VT, Custom);
1697 setOperationAction(ISD::AND, VT, Promote);
1698 AddPromotedToType (ISD::AND, VT, MVT::v8i64);
1699 setOperationAction(ISD::OR, VT, Promote);
1700 AddPromotedToType (ISD::OR, VT, MVT::v8i64);
1701 setOperationAction(ISD::XOR, VT, Promote);
1702 AddPromotedToType (ISD::XOR, VT, MVT::v8i64);
1706 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1707 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1708 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1710 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1711 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1712 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1713 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1714 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1715 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1716 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1717 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1718 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1719 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1720 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1721 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1723 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1724 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1725 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1726 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1727 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1728 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1729 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1730 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1732 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1733 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1734 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1735 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1736 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1737 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1738 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1739 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1742 // We want to custom lower some of our intrinsics.
1743 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1744 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1745 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1746 if (!Subtarget->is64Bit()) {
1747 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1748 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1751 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1752 // handle type legalization for these operations here.
1754 // FIXME: We really should do custom legalization for addition and
1755 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1756 // than generic legalization for 64-bit multiplication-with-overflow, though.
1757 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1758 if (VT == MVT::i64 && !Subtarget->is64Bit())
1760 // Add/Sub/Mul with overflow operations are custom lowered.
1761 setOperationAction(ISD::SADDO, VT, Custom);
1762 setOperationAction(ISD::UADDO, VT, Custom);
1763 setOperationAction(ISD::SSUBO, VT, Custom);
1764 setOperationAction(ISD::USUBO, VT, Custom);
1765 setOperationAction(ISD::SMULO, VT, Custom);
1766 setOperationAction(ISD::UMULO, VT, Custom);
1769 if (!Subtarget->is64Bit()) {
1770 // These libcalls are not available in 32-bit.
1771 setLibcallName(RTLIB::SHL_I128, nullptr);
1772 setLibcallName(RTLIB::SRL_I128, nullptr);
1773 setLibcallName(RTLIB::SRA_I128, nullptr);
1776 // Combine sin / cos into one node or libcall if possible.
1777 if (Subtarget->hasSinCos()) {
1778 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1779 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1780 if (Subtarget->isTargetDarwin()) {
1781 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1782 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1783 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1784 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1788 if (Subtarget->isTargetWin64()) {
1789 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1790 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1791 setOperationAction(ISD::SREM, MVT::i128, Custom);
1792 setOperationAction(ISD::UREM, MVT::i128, Custom);
1793 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1794 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1797 // We have target-specific dag combine patterns for the following nodes:
1798 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1799 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1800 setTargetDAGCombine(ISD::BITCAST);
1801 setTargetDAGCombine(ISD::VSELECT);
1802 setTargetDAGCombine(ISD::SELECT);
1803 setTargetDAGCombine(ISD::SHL);
1804 setTargetDAGCombine(ISD::SRA);
1805 setTargetDAGCombine(ISD::SRL);
1806 setTargetDAGCombine(ISD::OR);
1807 setTargetDAGCombine(ISD::AND);
1808 setTargetDAGCombine(ISD::ADD);
1809 setTargetDAGCombine(ISD::FADD);
1810 setTargetDAGCombine(ISD::FSUB);
1811 setTargetDAGCombine(ISD::FNEG);
1812 setTargetDAGCombine(ISD::FMA);
1813 setTargetDAGCombine(ISD::FMINNUM);
1814 setTargetDAGCombine(ISD::FMAXNUM);
1815 setTargetDAGCombine(ISD::SUB);
1816 setTargetDAGCombine(ISD::LOAD);
1817 setTargetDAGCombine(ISD::MLOAD);
1818 setTargetDAGCombine(ISD::STORE);
1819 setTargetDAGCombine(ISD::MSTORE);
1820 setTargetDAGCombine(ISD::TRUNCATE);
1821 setTargetDAGCombine(ISD::ZERO_EXTEND);
1822 setTargetDAGCombine(ISD::ANY_EXTEND);
1823 setTargetDAGCombine(ISD::SIGN_EXTEND);
1824 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1825 setTargetDAGCombine(ISD::SINT_TO_FP);
1826 setTargetDAGCombine(ISD::UINT_TO_FP);
1827 setTargetDAGCombine(ISD::SETCC);
1828 setTargetDAGCombine(ISD::BUILD_VECTOR);
1829 setTargetDAGCombine(ISD::MUL);
1830 setTargetDAGCombine(ISD::XOR);
1831 setTargetDAGCombine(ISD::MSCATTER);
1832 setTargetDAGCombine(ISD::MGATHER);
1834 computeRegisterProperties(Subtarget->getRegisterInfo());
1836 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1837 MaxStoresPerMemsetOptSize = 8;
1838 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1839 MaxStoresPerMemcpyOptSize = 4;
1840 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1841 MaxStoresPerMemmoveOptSize = 4;
1842 setPrefLoopAlignment(4); // 2^4 bytes.
1844 // A predictable cmov does not hurt on an in-order CPU.
1845 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1846 PredictableSelectIsExpensive = !Subtarget->isAtom();
1847 EnableExtLdPromotion = true;
1848 setPrefFunctionAlignment(4); // 2^4 bytes.
1850 verifyIntrinsicTables();
1853 // This has so far only been implemented for 64-bit MachO.
1854 bool X86TargetLowering::useLoadStackGuardNode() const {
1855 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1858 TargetLoweringBase::LegalizeTypeAction
1859 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1860 if (ExperimentalVectorWideningLegalization &&
1861 VT.getVectorNumElements() != 1 &&
1862 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1863 return TypeWidenVector;
1865 return TargetLoweringBase::getPreferredVectorAction(VT);
1868 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1871 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1873 if (VT.isSimple()) {
1874 MVT VVT = VT.getSimpleVT();
1875 const unsigned NumElts = VVT.getVectorNumElements();
1876 const MVT EltVT = VVT.getVectorElementType();
1877 if (VVT.is512BitVector()) {
1878 if (Subtarget->hasAVX512())
1879 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1880 EltVT == MVT::f32 || EltVT == MVT::f64)
1882 case 8: return MVT::v8i1;
1883 case 16: return MVT::v16i1;
1885 if (Subtarget->hasBWI())
1886 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1888 case 32: return MVT::v32i1;
1889 case 64: return MVT::v64i1;
1893 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1894 if (Subtarget->hasVLX())
1895 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1896 EltVT == MVT::f32 || EltVT == MVT::f64)
1898 case 2: return MVT::v2i1;
1899 case 4: return MVT::v4i1;
1900 case 8: return MVT::v8i1;
1902 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1903 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1905 case 8: return MVT::v8i1;
1906 case 16: return MVT::v16i1;
1907 case 32: return MVT::v32i1;
1912 return VT.changeVectorElementTypeToInteger();
1915 /// Helper for getByValTypeAlignment to determine
1916 /// the desired ByVal argument alignment.
1917 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1920 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1921 if (VTy->getBitWidth() == 128)
1923 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1924 unsigned EltAlign = 0;
1925 getMaxByValAlign(ATy->getElementType(), EltAlign);
1926 if (EltAlign > MaxAlign)
1927 MaxAlign = EltAlign;
1928 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1929 for (auto *EltTy : STy->elements()) {
1930 unsigned EltAlign = 0;
1931 getMaxByValAlign(EltTy, EltAlign);
1932 if (EltAlign > MaxAlign)
1933 MaxAlign = EltAlign;
1940 /// Return the desired alignment for ByVal aggregate
1941 /// function arguments in the caller parameter area. For X86, aggregates
1942 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1943 /// are at 4-byte boundaries.
1944 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1945 const DataLayout &DL) const {
1946 if (Subtarget->is64Bit()) {
1947 // Max of 8 and alignment of type.
1948 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1955 if (Subtarget->hasSSE1())
1956 getMaxByValAlign(Ty, Align);
1960 /// Returns the target specific optimal type for load
1961 /// and store operations as a result of memset, memcpy, and memmove
1962 /// lowering. If DstAlign is zero that means it's safe to destination
1963 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1964 /// means there isn't a need to check it against alignment requirement,
1965 /// probably because the source does not need to be loaded. If 'IsMemset' is
1966 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1967 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1968 /// source is constant so it does not need to be loaded.
1969 /// It returns EVT::Other if the type should be determined using generic
1970 /// target-independent logic.
1972 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1973 unsigned DstAlign, unsigned SrcAlign,
1974 bool IsMemset, bool ZeroMemset,
1976 MachineFunction &MF) const {
1977 const Function *F = MF.getFunction();
1978 if ((!IsMemset || ZeroMemset) &&
1979 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1981 (!Subtarget->isUnalignedMem16Slow() ||
1982 ((DstAlign == 0 || DstAlign >= 16) &&
1983 (SrcAlign == 0 || SrcAlign >= 16)))) {
1985 // FIXME: Check if unaligned 32-byte accesses are slow.
1986 if (Subtarget->hasInt256())
1988 if (Subtarget->hasFp256())
1991 if (Subtarget->hasSSE2())
1993 if (Subtarget->hasSSE1())
1995 } else if (!MemcpyStrSrc && Size >= 8 &&
1996 !Subtarget->is64Bit() &&
1997 Subtarget->hasSSE2()) {
1998 // Do not use f64 to lower memcpy if source is string constant. It's
1999 // better to use i32 to avoid the loads.
2003 // This is a compromise. If we reach here, unaligned accesses may be slow on
2004 // this target. However, creating smaller, aligned accesses could be even
2005 // slower and would certainly be a lot more code.
2006 if (Subtarget->is64Bit() && Size >= 8)
2011 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2013 return X86ScalarSSEf32;
2014 else if (VT == MVT::f64)
2015 return X86ScalarSSEf64;
2020 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2025 switch (VT.getSizeInBits()) {
2027 // 8-byte and under are always assumed to be fast.
2031 *Fast = !Subtarget->isUnalignedMem16Slow();
2034 *Fast = !Subtarget->isUnalignedMem32Slow();
2036 // TODO: What about AVX-512 (512-bit) accesses?
2039 // Misaligned accesses of any size are always allowed.
2043 /// Return the entry encoding for a jump table in the
2044 /// current function. The returned value is a member of the
2045 /// MachineJumpTableInfo::JTEntryKind enum.
2046 unsigned X86TargetLowering::getJumpTableEncoding() const {
2047 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2049 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2050 Subtarget->isPICStyleGOT())
2051 return MachineJumpTableInfo::EK_Custom32;
2053 // Otherwise, use the normal jump table encoding heuristics.
2054 return TargetLowering::getJumpTableEncoding();
2057 bool X86TargetLowering::useSoftFloat() const {
2058 return Subtarget->useSoftFloat();
2062 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2063 const MachineBasicBlock *MBB,
2064 unsigned uid,MCContext &Ctx) const{
2065 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2066 Subtarget->isPICStyleGOT());
2067 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2069 return MCSymbolRefExpr::create(MBB->getSymbol(),
2070 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2073 /// Returns relocation base for the given PIC jumptable.
2074 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2075 SelectionDAG &DAG) const {
2076 if (!Subtarget->is64Bit())
2077 // This doesn't have SDLoc associated with it, but is not really the
2078 // same as a Register.
2079 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2080 getPointerTy(DAG.getDataLayout()));
2084 /// This returns the relocation base for the given PIC jumptable,
2085 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2086 const MCExpr *X86TargetLowering::
2087 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2088 MCContext &Ctx) const {
2089 // X86-64 uses RIP relative addressing based on the jump table label.
2090 if (Subtarget->isPICStyleRIPRel())
2091 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2093 // Otherwise, the reference is relative to the PIC base.
2094 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2097 std::pair<const TargetRegisterClass *, uint8_t>
2098 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2100 const TargetRegisterClass *RRC = nullptr;
2102 switch (VT.SimpleTy) {
2104 return TargetLowering::findRepresentativeClass(TRI, VT);
2105 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2106 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2109 RRC = &X86::VR64RegClass;
2111 case MVT::f32: case MVT::f64:
2112 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2113 case MVT::v4f32: case MVT::v2f64:
2114 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2116 RRC = &X86::VR128RegClass;
2119 return std::make_pair(RRC, Cost);
2122 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2123 unsigned &Offset) const {
2124 if (!Subtarget->isTargetLinux())
2127 if (Subtarget->is64Bit()) {
2128 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2130 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2142 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2143 if (!Subtarget->isTargetAndroid())
2144 return TargetLowering::getSafeStackPointerLocation(IRB);
2146 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2147 // definition of TLS_SLOT_SAFESTACK in
2148 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2149 unsigned AddressSpace, Offset;
2150 if (Subtarget->is64Bit()) {
2151 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2153 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2163 return ConstantExpr::getIntToPtr(
2164 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2165 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2168 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2169 unsigned DestAS) const {
2170 assert(SrcAS != DestAS && "Expected different address spaces!");
2172 return SrcAS < 256 && DestAS < 256;
2175 //===----------------------------------------------------------------------===//
2176 // Return Value Calling Convention Implementation
2177 //===----------------------------------------------------------------------===//
2179 #include "X86GenCallingConv.inc"
2181 bool X86TargetLowering::CanLowerReturn(
2182 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2183 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2184 SmallVector<CCValAssign, 16> RVLocs;
2185 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2186 return CCInfo.CheckReturn(Outs, RetCC_X86);
2189 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2190 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2195 X86TargetLowering::LowerReturn(SDValue Chain,
2196 CallingConv::ID CallConv, bool isVarArg,
2197 const SmallVectorImpl<ISD::OutputArg> &Outs,
2198 const SmallVectorImpl<SDValue> &OutVals,
2199 SDLoc dl, SelectionDAG &DAG) const {
2200 MachineFunction &MF = DAG.getMachineFunction();
2201 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2203 if (CallConv == CallingConv::X86_INTR && !Outs.empty())
2204 report_fatal_error("X86 interrupts may not return any value");
2206 SmallVector<CCValAssign, 16> RVLocs;
2207 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2208 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2211 SmallVector<SDValue, 6> RetOps;
2212 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2213 // Operand #1 = Bytes To Pop
2214 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2217 // Copy the result values into the output registers.
2218 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2219 CCValAssign &VA = RVLocs[i];
2220 assert(VA.isRegLoc() && "Can only return in registers!");
2221 SDValue ValToCopy = OutVals[i];
2222 EVT ValVT = ValToCopy.getValueType();
2224 // Promote values to the appropriate types.
2225 if (VA.getLocInfo() == CCValAssign::SExt)
2226 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2227 else if (VA.getLocInfo() == CCValAssign::ZExt)
2228 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2229 else if (VA.getLocInfo() == CCValAssign::AExt) {
2230 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2231 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2233 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2235 else if (VA.getLocInfo() == CCValAssign::BCvt)
2236 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2238 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2239 "Unexpected FP-extend for return value.");
2241 // If this is x86-64, and we disabled SSE, we can't return FP values,
2242 // or SSE or MMX vectors.
2243 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2244 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2245 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2246 report_fatal_error("SSE register return with SSE disabled");
2248 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2249 // llvm-gcc has never done it right and no one has noticed, so this
2250 // should be OK for now.
2251 if (ValVT == MVT::f64 &&
2252 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2253 report_fatal_error("SSE2 register return with SSE2 disabled");
2255 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2256 // the RET instruction and handled by the FP Stackifier.
2257 if (VA.getLocReg() == X86::FP0 ||
2258 VA.getLocReg() == X86::FP1) {
2259 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2260 // change the value to the FP stack register class.
2261 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2262 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2263 RetOps.push_back(ValToCopy);
2264 // Don't emit a copytoreg.
2268 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2269 // which is returned in RAX / RDX.
2270 if (Subtarget->is64Bit()) {
2271 if (ValVT == MVT::x86mmx) {
2272 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2273 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2274 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2276 // If we don't have SSE2 available, convert to v4f32 so the generated
2277 // register is legal.
2278 if (!Subtarget->hasSSE2())
2279 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2284 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2285 Flag = Chain.getValue(1);
2286 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2289 // All x86 ABIs require that for returning structs by value we copy
2290 // the sret argument into %rax/%eax (depending on ABI) for the return.
2291 // We saved the argument into a virtual register in the entry block,
2292 // so now we copy the value out and into %rax/%eax.
2294 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2295 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2296 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2297 // either case FuncInfo->setSRetReturnReg() will have been called.
2298 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2299 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2300 getPointerTy(MF.getDataLayout()));
2303 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2304 X86::RAX : X86::EAX;
2305 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2306 Flag = Chain.getValue(1);
2308 // RAX/EAX now acts like a return value.
2310 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2313 RetOps[0] = Chain; // Update chain.
2315 // Add the flag if we have it.
2317 RetOps.push_back(Flag);
2319 X86ISD::NodeType opcode = X86ISD::RET_FLAG;
2320 if (CallConv == CallingConv::X86_INTR)
2321 opcode = X86ISD::IRET;
2322 return DAG.getNode(opcode, dl, MVT::Other, RetOps);
2325 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2326 if (N->getNumValues() != 1)
2328 if (!N->hasNUsesOfValue(1, 0))
2331 SDValue TCChain = Chain;
2332 SDNode *Copy = *N->use_begin();
2333 if (Copy->getOpcode() == ISD::CopyToReg) {
2334 // If the copy has a glue operand, we conservatively assume it isn't safe to
2335 // perform a tail call.
2336 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2338 TCChain = Copy->getOperand(0);
2339 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2342 bool HasRet = false;
2343 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2345 if (UI->getOpcode() != X86ISD::RET_FLAG)
2347 // If we are returning more than one value, we can definitely
2348 // not make a tail call see PR19530
2349 if (UI->getNumOperands() > 4)
2351 if (UI->getNumOperands() == 4 &&
2352 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2365 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2366 ISD::NodeType ExtendKind) const {
2368 // TODO: Is this also valid on 32-bit?
2369 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2370 ReturnMVT = MVT::i8;
2372 ReturnMVT = MVT::i32;
2374 EVT MinVT = getRegisterType(Context, ReturnMVT);
2375 return VT.bitsLT(MinVT) ? MinVT : VT;
2378 /// Lower the result values of a call into the
2379 /// appropriate copies out of appropriate physical registers.
2382 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2383 CallingConv::ID CallConv, bool isVarArg,
2384 const SmallVectorImpl<ISD::InputArg> &Ins,
2385 SDLoc dl, SelectionDAG &DAG,
2386 SmallVectorImpl<SDValue> &InVals) const {
2388 // Assign locations to each value returned by this call.
2389 SmallVector<CCValAssign, 16> RVLocs;
2390 bool Is64Bit = Subtarget->is64Bit();
2391 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2393 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2395 // Copy all of the result registers out of their specified physreg.
2396 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2397 CCValAssign &VA = RVLocs[i];
2398 EVT CopyVT = VA.getLocVT();
2400 // If this is x86-64, and we disabled SSE, we can't return FP values
2401 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) &&
2402 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2403 report_fatal_error("SSE register return with SSE disabled");
2406 // If we prefer to use the value in xmm registers, copy it out as f80 and
2407 // use a truncate to move it from fp stack reg to xmm reg.
2408 bool RoundAfterCopy = false;
2409 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2410 isScalarFPTypeInSSEReg(VA.getValVT())) {
2412 RoundAfterCopy = (CopyVT != VA.getLocVT());
2415 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2416 CopyVT, InFlag).getValue(1);
2417 SDValue Val = Chain.getValue(0);
2420 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2421 // This truncation won't change the value.
2422 DAG.getIntPtrConstant(1, dl));
2424 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2425 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2427 InFlag = Chain.getValue(2);
2428 InVals.push_back(Val);
2434 //===----------------------------------------------------------------------===//
2435 // C & StdCall & Fast Calling Convention implementation
2436 //===----------------------------------------------------------------------===//
2437 // StdCall calling convention seems to be standard for many Windows' API
2438 // routines and around. It differs from C calling convention just a little:
2439 // callee should clean up the stack, not caller. Symbols should be also
2440 // decorated in some fancy way :) It doesn't support any vector arguments.
2441 // For info on fast calling convention see Fast Calling Convention (tail call)
2442 // implementation LowerX86_32FastCCCallTo.
2444 /// CallIsStructReturn - Determines whether a call uses struct return
2446 enum StructReturnType {
2451 static StructReturnType
2452 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsMCU) {
2454 return NotStructReturn;
2456 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2457 if (!Flags.isSRet())
2458 return NotStructReturn;
2459 if (Flags.isInReg() || IsMCU)
2460 return RegStructReturn;
2461 return StackStructReturn;
2464 /// Determines whether a function uses struct return semantics.
2465 static StructReturnType
2466 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsMCU) {
2468 return NotStructReturn;
2470 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2471 if (!Flags.isSRet())
2472 return NotStructReturn;
2473 if (Flags.isInReg() || IsMCU)
2474 return RegStructReturn;
2475 return StackStructReturn;
2478 /// Make a copy of an aggregate at address specified by "Src" to address
2479 /// "Dst" with size and alignment information specified by the specific
2480 /// parameter attribute. The copy will be passed as a byval function parameter.
2482 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2483 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2485 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2487 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2488 /*isVolatile*/false, /*AlwaysInline=*/true,
2489 /*isTailCall*/false,
2490 MachinePointerInfo(), MachinePointerInfo());
2493 /// Return true if the calling convention is one that we can guarantee TCO for.
2494 static bool canGuaranteeTCO(CallingConv::ID CC) {
2495 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2496 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2499 /// Return true if we might ever do TCO for calls with this calling convention.
2500 static bool mayTailCallThisCC(CallingConv::ID CC) {
2502 // C calling conventions:
2503 case CallingConv::C:
2504 case CallingConv::X86_64_Win64:
2505 case CallingConv::X86_64_SysV:
2506 // Callee pop conventions:
2507 case CallingConv::X86_ThisCall:
2508 case CallingConv::X86_StdCall:
2509 case CallingConv::X86_VectorCall:
2510 case CallingConv::X86_FastCall:
2513 return canGuaranteeTCO(CC);
2517 /// Return true if the function is being made into a tailcall target by
2518 /// changing its ABI.
2519 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2520 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2523 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2525 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2526 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2530 CallingConv::ID CalleeCC = CS.getCallingConv();
2531 if (!mayTailCallThisCC(CalleeCC))
2538 X86TargetLowering::LowerMemArgument(SDValue Chain,
2539 CallingConv::ID CallConv,
2540 const SmallVectorImpl<ISD::InputArg> &Ins,
2541 SDLoc dl, SelectionDAG &DAG,
2542 const CCValAssign &VA,
2543 MachineFrameInfo *MFI,
2545 // Create the nodes corresponding to a load from this parameter slot.
2546 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2547 bool AlwaysUseMutable = shouldGuaranteeTCO(
2548 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2549 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2552 // If value is passed by pointer we have address passed instead of the value
2554 bool ExtendedInMem = VA.isExtInLoc() &&
2555 VA.getValVT().getScalarType() == MVT::i1;
2557 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2558 ValVT = VA.getLocVT();
2560 ValVT = VA.getValVT();
2562 // Calculate SP offset of interrupt parameter, re-arrange the slot normally
2563 // taken by a return address.
2565 if (CallConv == CallingConv::X86_INTR) {
2566 const X86Subtarget& Subtarget =
2567 static_cast<const X86Subtarget&>(DAG.getSubtarget());
2568 // X86 interrupts may take one or two arguments.
2569 // On the stack there will be no return address as in regular call.
2570 // Offset of last argument need to be set to -4/-8 bytes.
2571 // Where offset of the first argument out of two, should be set to 0 bytes.
2572 Offset = (Subtarget.is64Bit() ? 8 : 4) * ((i + 1) % Ins.size() - 1);
2575 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2576 // changed with more analysis.
2577 // In case of tail call optimization mark all arguments mutable. Since they
2578 // could be overwritten by lowering of arguments in case of a tail call.
2579 if (Flags.isByVal()) {
2580 unsigned Bytes = Flags.getByValSize();
2581 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2582 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2583 // Adjust SP offset of interrupt parameter.
2584 if (CallConv == CallingConv::X86_INTR) {
2585 MFI->setObjectOffset(FI, Offset);
2587 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2589 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2590 VA.getLocMemOffset(), isImmutable);
2591 // Adjust SP offset of interrupt parameter.
2592 if (CallConv == CallingConv::X86_INTR) {
2593 MFI->setObjectOffset(FI, Offset);
2596 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2597 SDValue Val = DAG.getLoad(
2598 ValVT, dl, Chain, FIN,
2599 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2601 return ExtendedInMem ?
2602 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2606 // FIXME: Get this from tablegen.
2607 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2608 const X86Subtarget *Subtarget) {
2609 assert(Subtarget->is64Bit());
2611 if (Subtarget->isCallingConvWin64(CallConv)) {
2612 static const MCPhysReg GPR64ArgRegsWin64[] = {
2613 X86::RCX, X86::RDX, X86::R8, X86::R9
2615 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2618 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2619 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2621 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2624 // FIXME: Get this from tablegen.
2625 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2626 CallingConv::ID CallConv,
2627 const X86Subtarget *Subtarget) {
2628 assert(Subtarget->is64Bit());
2629 if (Subtarget->isCallingConvWin64(CallConv)) {
2630 // The XMM registers which might contain var arg parameters are shadowed
2631 // in their paired GPR. So we only need to save the GPR to their home
2633 // TODO: __vectorcall will change this.
2637 const Function *Fn = MF.getFunction();
2638 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2639 bool isSoftFloat = Subtarget->useSoftFloat();
2640 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2641 "SSE register cannot be used when SSE is disabled!");
2642 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2643 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2647 static const MCPhysReg XMMArgRegs64Bit[] = {
2648 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2649 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2651 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2654 SDValue X86TargetLowering::LowerFormalArguments(
2655 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2656 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2657 SmallVectorImpl<SDValue> &InVals) const {
2658 MachineFunction &MF = DAG.getMachineFunction();
2659 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2660 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2662 const Function* Fn = MF.getFunction();
2663 if (Fn->hasExternalLinkage() &&
2664 Subtarget->isTargetCygMing() &&
2665 Fn->getName() == "main")
2666 FuncInfo->setForceFramePointer(true);
2668 MachineFrameInfo *MFI = MF.getFrameInfo();
2669 bool Is64Bit = Subtarget->is64Bit();
2670 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2672 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2673 "Var args not supported with calling convention fastcc, ghc or hipe");
2675 if (CallConv == CallingConv::X86_INTR) {
2676 bool isLegal = Ins.size() == 1 ||
2677 (Ins.size() == 2 && ((Is64Bit && Ins[1].VT == MVT::i64) ||
2678 (!Is64Bit && Ins[1].VT == MVT::i32)));
2680 report_fatal_error("X86 interrupts may take one or two arguments");
2683 // Assign locations to all of the incoming arguments.
2684 SmallVector<CCValAssign, 16> ArgLocs;
2685 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2687 // Allocate shadow area for Win64
2689 CCInfo.AllocateStack(32, 8);
2691 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2693 unsigned LastVal = ~0U;
2695 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2696 CCValAssign &VA = ArgLocs[i];
2697 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2699 assert(VA.getValNo() != LastVal &&
2700 "Don't support value assigned to multiple locs yet");
2702 LastVal = VA.getValNo();
2704 if (VA.isRegLoc()) {
2705 EVT RegVT = VA.getLocVT();
2706 const TargetRegisterClass *RC;
2707 if (RegVT == MVT::i32)
2708 RC = &X86::GR32RegClass;
2709 else if (Is64Bit && RegVT == MVT::i64)
2710 RC = &X86::GR64RegClass;
2711 else if (RegVT == MVT::f32)
2712 RC = &X86::FR32RegClass;
2713 else if (RegVT == MVT::f64)
2714 RC = &X86::FR64RegClass;
2715 else if (RegVT == MVT::f128)
2716 RC = &X86::FR128RegClass;
2717 else if (RegVT.is512BitVector())
2718 RC = &X86::VR512RegClass;
2719 else if (RegVT.is256BitVector())
2720 RC = &X86::VR256RegClass;
2721 else if (RegVT.is128BitVector())
2722 RC = &X86::VR128RegClass;
2723 else if (RegVT == MVT::x86mmx)
2724 RC = &X86::VR64RegClass;
2725 else if (RegVT == MVT::i1)
2726 RC = &X86::VK1RegClass;
2727 else if (RegVT == MVT::v8i1)
2728 RC = &X86::VK8RegClass;
2729 else if (RegVT == MVT::v16i1)
2730 RC = &X86::VK16RegClass;
2731 else if (RegVT == MVT::v32i1)
2732 RC = &X86::VK32RegClass;
2733 else if (RegVT == MVT::v64i1)
2734 RC = &X86::VK64RegClass;
2736 llvm_unreachable("Unknown argument type!");
2738 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2739 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2741 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2742 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2744 if (VA.getLocInfo() == CCValAssign::SExt)
2745 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2746 DAG.getValueType(VA.getValVT()));
2747 else if (VA.getLocInfo() == CCValAssign::ZExt)
2748 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2749 DAG.getValueType(VA.getValVT()));
2750 else if (VA.getLocInfo() == CCValAssign::BCvt)
2751 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2753 if (VA.isExtInLoc()) {
2754 // Handle MMX values passed in XMM regs.
2755 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2756 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2758 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2761 assert(VA.isMemLoc());
2762 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2765 // If value is passed via pointer - do a load.
2766 if (VA.getLocInfo() == CCValAssign::Indirect)
2767 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2768 MachinePointerInfo(), false, false, false, 0);
2770 InVals.push_back(ArgValue);
2773 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2774 // All x86 ABIs require that for returning structs by value we copy the
2775 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2776 // the argument into a virtual register so that we can access it from the
2778 if (Ins[i].Flags.isSRet()) {
2779 unsigned Reg = FuncInfo->getSRetReturnReg();
2781 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2782 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2783 FuncInfo->setSRetReturnReg(Reg);
2785 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2786 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2791 unsigned StackSize = CCInfo.getNextStackOffset();
2792 // Align stack specially for tail calls.
2793 if (shouldGuaranteeTCO(CallConv,
2794 MF.getTarget().Options.GuaranteedTailCallOpt))
2795 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2797 // If the function takes variable number of arguments, make a frame index for
2798 // the start of the first vararg value... for expansion of llvm.va_start. We
2799 // can skip this if there are no va_start calls.
2800 if (MFI->hasVAStart() &&
2801 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2802 CallConv != CallingConv::X86_ThisCall))) {
2803 FuncInfo->setVarArgsFrameIndex(
2804 MFI->CreateFixedObject(1, StackSize, true));
2807 // Figure out if XMM registers are in use.
2808 assert(!(Subtarget->useSoftFloat() &&
2809 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2810 "SSE register cannot be used when SSE is disabled!");
2812 // 64-bit calling conventions support varargs and register parameters, so we
2813 // have to do extra work to spill them in the prologue.
2814 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2815 // Find the first unallocated argument registers.
2816 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2817 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2818 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2819 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2820 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2821 "SSE register cannot be used when SSE is disabled!");
2823 // Gather all the live in physical registers.
2824 SmallVector<SDValue, 6> LiveGPRs;
2825 SmallVector<SDValue, 8> LiveXMMRegs;
2827 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2828 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2830 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2832 if (!ArgXMMs.empty()) {
2833 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2834 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2835 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2836 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2837 LiveXMMRegs.push_back(
2838 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2843 // Get to the caller-allocated home save location. Add 8 to account
2844 // for the return address.
2845 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2846 FuncInfo->setRegSaveFrameIndex(
2847 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2848 // Fixup to set vararg frame on shadow area (4 x i64).
2850 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2852 // For X86-64, if there are vararg parameters that are passed via
2853 // registers, then we must store them to their spots on the stack so
2854 // they may be loaded by deferencing the result of va_next.
2855 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2856 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2857 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2858 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2861 // Store the integer parameter registers.
2862 SmallVector<SDValue, 8> MemOps;
2863 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2864 getPointerTy(DAG.getDataLayout()));
2865 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2866 for (SDValue Val : LiveGPRs) {
2867 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2868 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2870 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2871 MachinePointerInfo::getFixedStack(
2872 DAG.getMachineFunction(),
2873 FuncInfo->getRegSaveFrameIndex(), Offset),
2875 MemOps.push_back(Store);
2879 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2880 // Now store the XMM (fp + vector) parameter registers.
2881 SmallVector<SDValue, 12> SaveXMMOps;
2882 SaveXMMOps.push_back(Chain);
2883 SaveXMMOps.push_back(ALVal);
2884 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2885 FuncInfo->getRegSaveFrameIndex(), dl));
2886 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2887 FuncInfo->getVarArgsFPOffset(), dl));
2888 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2890 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2891 MVT::Other, SaveXMMOps));
2894 if (!MemOps.empty())
2895 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2898 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2899 // Find the largest legal vector type.
2900 MVT VecVT = MVT::Other;
2901 // FIXME: Only some x86_32 calling conventions support AVX512.
2902 if (Subtarget->hasAVX512() &&
2903 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2904 CallConv == CallingConv::Intel_OCL_BI)))
2905 VecVT = MVT::v16f32;
2906 else if (Subtarget->hasAVX())
2908 else if (Subtarget->hasSSE2())
2911 // We forward some GPRs and some vector types.
2912 SmallVector<MVT, 2> RegParmTypes;
2913 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2914 RegParmTypes.push_back(IntVT);
2915 if (VecVT != MVT::Other)
2916 RegParmTypes.push_back(VecVT);
2918 // Compute the set of forwarded registers. The rest are scratch.
2919 SmallVectorImpl<ForwardedRegister> &Forwards =
2920 FuncInfo->getForwardedMustTailRegParms();
2921 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2923 // Conservatively forward AL on x86_64, since it might be used for varargs.
2924 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2925 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2926 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2929 // Copy all forwards from physical to virtual registers.
2930 for (ForwardedRegister &F : Forwards) {
2931 // FIXME: Can we use a less constrained schedule?
2932 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2933 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2934 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2938 // Some CCs need callee pop.
2939 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2940 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2941 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2942 } else if (CallConv == CallingConv::X86_INTR && Ins.size() == 2) {
2943 // X86 interrupts must pop the error code if present
2944 FuncInfo->setBytesToPopOnReturn(Is64Bit ? 8 : 4);
2946 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2947 // If this is an sret function, the return should pop the hidden pointer.
2948 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2949 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2950 argsAreStructReturn(Ins, Subtarget->isTargetMCU()) == StackStructReturn)
2951 FuncInfo->setBytesToPopOnReturn(4);
2955 // RegSaveFrameIndex is X86-64 only.
2956 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2957 if (CallConv == CallingConv::X86_FastCall ||
2958 CallConv == CallingConv::X86_ThisCall)
2959 // fastcc functions can't have varargs.
2960 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2963 FuncInfo->setArgumentStackSize(StackSize);
2965 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2966 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2967 if (Personality == EHPersonality::CoreCLR) {
2969 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2970 // that we'd prefer this slot be allocated towards the bottom of the frame
2971 // (i.e. near the stack pointer after allocating the frame). Every
2972 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2973 // offset from the bottom of this and each funclet's frame must be the
2974 // same, so the size of funclets' (mostly empty) frames is dictated by
2975 // how far this slot is from the bottom (since they allocate just enough
2976 // space to accomodate holding this slot at the correct offset).
2977 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2978 EHInfo->PSPSymFrameIdx = PSPSymFI;
2986 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2987 SDValue StackPtr, SDValue Arg,
2988 SDLoc dl, SelectionDAG &DAG,
2989 const CCValAssign &VA,
2990 ISD::ArgFlagsTy Flags) const {
2991 unsigned LocMemOffset = VA.getLocMemOffset();
2992 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2993 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2995 if (Flags.isByVal())
2996 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2998 return DAG.getStore(
2999 Chain, dl, Arg, PtrOff,
3000 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
3004 /// Emit a load of return address if tail call
3005 /// optimization is performed and it is required.
3007 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
3008 SDValue &OutRetAddr, SDValue Chain,
3009 bool IsTailCall, bool Is64Bit,
3010 int FPDiff, SDLoc dl) const {
3011 // Adjust the Return address stack slot.
3012 EVT VT = getPointerTy(DAG.getDataLayout());
3013 OutRetAddr = getReturnAddressFrameIndex(DAG);
3015 // Load the "old" Return address.
3016 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
3017 false, false, false, 0);
3018 return SDValue(OutRetAddr.getNode(), 1);
3021 /// Emit a store of the return address if tail call
3022 /// optimization is performed and it is required (FPDiff!=0).
3023 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
3024 SDValue Chain, SDValue RetAddrFrIdx,
3025 EVT PtrVT, unsigned SlotSize,
3026 int FPDiff, SDLoc dl) {
3027 // Store the return address to the appropriate stack slot.
3028 if (!FPDiff) return Chain;
3029 // Calculate the new stack slot for the return address.
3030 int NewReturnAddrFI =
3031 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
3033 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
3034 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
3035 MachinePointerInfo::getFixedStack(
3036 DAG.getMachineFunction(), NewReturnAddrFI),
3041 /// Returns a vector_shuffle mask for an movs{s|d}, movd
3042 /// operation of specified width.
3043 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
3045 unsigned NumElems = VT.getVectorNumElements();
3046 SmallVector<int, 8> Mask;
3047 Mask.push_back(NumElems);
3048 for (unsigned i = 1; i != NumElems; ++i)
3050 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3054 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3055 SmallVectorImpl<SDValue> &InVals) const {
3056 SelectionDAG &DAG = CLI.DAG;
3058 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3059 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3060 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3061 SDValue Chain = CLI.Chain;
3062 SDValue Callee = CLI.Callee;
3063 CallingConv::ID CallConv = CLI.CallConv;
3064 bool &isTailCall = CLI.IsTailCall;
3065 bool isVarArg = CLI.IsVarArg;
3067 MachineFunction &MF = DAG.getMachineFunction();
3068 bool Is64Bit = Subtarget->is64Bit();
3069 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3070 StructReturnType SR = callIsStructReturn(Outs, Subtarget->isTargetMCU());
3071 bool IsSibcall = false;
3072 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3073 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3075 if (CallConv == CallingConv::X86_INTR)
3076 report_fatal_error("X86 interrupts may not be called directly");
3078 if (Attr.getValueAsString() == "true")
3081 if (Subtarget->isPICStyleGOT() &&
3082 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3083 // If we are using a GOT, disable tail calls to external symbols with
3084 // default visibility. Tail calling such a symbol requires using a GOT
3085 // relocation, which forces early binding of the symbol. This breaks code
3086 // that require lazy function symbol resolution. Using musttail or
3087 // GuaranteedTailCallOpt will override this.
3088 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3089 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3090 G->getGlobal()->hasDefaultVisibility()))
3094 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3096 // Force this to be a tail call. The verifier rules are enough to ensure
3097 // that we can lower this successfully without moving the return address
3100 } else if (isTailCall) {
3101 // Check if it's really possible to do a tail call.
3102 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3103 isVarArg, SR != NotStructReturn,
3104 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3105 Outs, OutVals, Ins, DAG);
3107 // Sibcalls are automatically detected tailcalls which do not require
3109 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3116 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3117 "Var args not supported with calling convention fastcc, ghc or hipe");
3119 // Analyze operands of the call, assigning locations to each operand.
3120 SmallVector<CCValAssign, 16> ArgLocs;
3121 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3123 // Allocate shadow area for Win64
3125 CCInfo.AllocateStack(32, 8);
3127 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3129 // Get a count of how many bytes are to be pushed on the stack.
3130 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3132 // This is a sibcall. The memory operands are available in caller's
3133 // own caller's stack.
3135 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3136 canGuaranteeTCO(CallConv))
3137 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3140 if (isTailCall && !IsSibcall && !IsMustTail) {
3141 // Lower arguments at fp - stackoffset + fpdiff.
3142 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3144 FPDiff = NumBytesCallerPushed - NumBytes;
3146 // Set the delta of movement of the returnaddr stackslot.
3147 // But only set if delta is greater than previous delta.
3148 if (FPDiff < X86Info->getTCReturnAddrDelta())
3149 X86Info->setTCReturnAddrDelta(FPDiff);
3152 unsigned NumBytesToPush = NumBytes;
3153 unsigned NumBytesToPop = NumBytes;
3155 // If we have an inalloca argument, all stack space has already been allocated
3156 // for us and be right at the top of the stack. We don't support multiple
3157 // arguments passed in memory when using inalloca.
3158 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3160 if (!ArgLocs.back().isMemLoc())
3161 report_fatal_error("cannot use inalloca attribute on a register "
3163 if (ArgLocs.back().getLocMemOffset() != 0)
3164 report_fatal_error("any parameter with the inalloca attribute must be "
3165 "the only memory argument");
3169 Chain = DAG.getCALLSEQ_START(
3170 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3172 SDValue RetAddrFrIdx;
3173 // Load return address for tail calls.
3174 if (isTailCall && FPDiff)
3175 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3176 Is64Bit, FPDiff, dl);
3178 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3179 SmallVector<SDValue, 8> MemOpChains;
3182 // Walk the register/memloc assignments, inserting copies/loads. In the case
3183 // of tail call optimization arguments are handle later.
3184 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3185 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3186 // Skip inalloca arguments, they have already been written.
3187 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3188 if (Flags.isInAlloca())
3191 CCValAssign &VA = ArgLocs[i];
3192 EVT RegVT = VA.getLocVT();
3193 SDValue Arg = OutVals[i];
3194 bool isByVal = Flags.isByVal();
3196 // Promote the value if needed.
3197 switch (VA.getLocInfo()) {
3198 default: llvm_unreachable("Unknown loc info!");
3199 case CCValAssign::Full: break;
3200 case CCValAssign::SExt:
3201 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3203 case CCValAssign::ZExt:
3204 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3206 case CCValAssign::AExt:
3207 if (Arg.getValueType().isVector() &&
3208 Arg.getValueType().getVectorElementType() == MVT::i1)
3209 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3210 else if (RegVT.is128BitVector()) {
3211 // Special case: passing MMX values in XMM registers.
3212 Arg = DAG.getBitcast(MVT::i64, Arg);
3213 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3214 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3216 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3218 case CCValAssign::BCvt:
3219 Arg = DAG.getBitcast(RegVT, Arg);
3221 case CCValAssign::Indirect: {
3222 // Store the argument.
3223 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3224 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3225 Chain = DAG.getStore(
3226 Chain, dl, Arg, SpillSlot,
3227 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3234 if (VA.isRegLoc()) {
3235 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3236 if (isVarArg && IsWin64) {
3237 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3238 // shadow reg if callee is a varargs function.
3239 unsigned ShadowReg = 0;
3240 switch (VA.getLocReg()) {
3241 case X86::XMM0: ShadowReg = X86::RCX; break;
3242 case X86::XMM1: ShadowReg = X86::RDX; break;
3243 case X86::XMM2: ShadowReg = X86::R8; break;
3244 case X86::XMM3: ShadowReg = X86::R9; break;
3247 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3249 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3250 assert(VA.isMemLoc());
3251 if (!StackPtr.getNode())
3252 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3253 getPointerTy(DAG.getDataLayout()));
3254 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3255 dl, DAG, VA, Flags));
3259 if (!MemOpChains.empty())
3260 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3262 if (Subtarget->isPICStyleGOT()) {
3263 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3266 RegsToPass.push_back(std::make_pair(
3267 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3268 getPointerTy(DAG.getDataLayout()))));
3270 // If we are tail calling and generating PIC/GOT style code load the
3271 // address of the callee into ECX. The value in ecx is used as target of
3272 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3273 // for tail calls on PIC/GOT architectures. Normally we would just put the
3274 // address of GOT into ebx and then call target@PLT. But for tail calls
3275 // ebx would be restored (since ebx is callee saved) before jumping to the
3278 // Note: The actual moving to ECX is done further down.
3279 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3280 if (G && !G->getGlobal()->hasLocalLinkage() &&
3281 G->getGlobal()->hasDefaultVisibility())
3282 Callee = LowerGlobalAddress(Callee, DAG);
3283 else if (isa<ExternalSymbolSDNode>(Callee))
3284 Callee = LowerExternalSymbol(Callee, DAG);
3288 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3289 // From AMD64 ABI document:
3290 // For calls that may call functions that use varargs or stdargs
3291 // (prototype-less calls or calls to functions containing ellipsis (...) in
3292 // the declaration) %al is used as hidden argument to specify the number
3293 // of SSE registers used. The contents of %al do not need to match exactly
3294 // the number of registers, but must be an ubound on the number of SSE
3295 // registers used and is in the range 0 - 8 inclusive.
3297 // Count the number of XMM registers allocated.
3298 static const MCPhysReg XMMArgRegs[] = {
3299 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3300 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3302 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3303 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3304 && "SSE registers cannot be used when SSE is disabled");
3306 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3307 DAG.getConstant(NumXMMRegs, dl,
3311 if (isVarArg && IsMustTail) {
3312 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3313 for (const auto &F : Forwards) {
3314 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3315 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3319 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3320 // don't need this because the eligibility check rejects calls that require
3321 // shuffling arguments passed in memory.
3322 if (!IsSibcall && isTailCall) {
3323 // Force all the incoming stack arguments to be loaded from the stack
3324 // before any new outgoing arguments are stored to the stack, because the
3325 // outgoing stack slots may alias the incoming argument stack slots, and
3326 // the alias isn't otherwise explicit. This is slightly more conservative
3327 // than necessary, because it means that each store effectively depends
3328 // on every argument instead of just those arguments it would clobber.
3329 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3331 SmallVector<SDValue, 8> MemOpChains2;
3334 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3335 CCValAssign &VA = ArgLocs[i];
3338 assert(VA.isMemLoc());
3339 SDValue Arg = OutVals[i];
3340 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3341 // Skip inalloca arguments. They don't require any work.
3342 if (Flags.isInAlloca())
3344 // Create frame index.
3345 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3346 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3347 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3348 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3350 if (Flags.isByVal()) {
3351 // Copy relative to framepointer.
3352 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3353 if (!StackPtr.getNode())
3354 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3355 getPointerTy(DAG.getDataLayout()));
3356 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3359 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3363 // Store relative to framepointer.
3364 MemOpChains2.push_back(DAG.getStore(
3365 ArgChain, dl, Arg, FIN,
3366 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3371 if (!MemOpChains2.empty())
3372 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3374 // Store the return address to the appropriate stack slot.
3375 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3376 getPointerTy(DAG.getDataLayout()),
3377 RegInfo->getSlotSize(), FPDiff, dl);
3380 // Build a sequence of copy-to-reg nodes chained together with token chain
3381 // and flag operands which copy the outgoing args into registers.
3383 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3384 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3385 RegsToPass[i].second, InFlag);
3386 InFlag = Chain.getValue(1);
3389 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3390 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3391 // In the 64-bit large code model, we have to make all calls
3392 // through a register, since the call instruction's 32-bit
3393 // pc-relative offset may not be large enough to hold the whole
3395 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3396 // If the callee is a GlobalAddress node (quite common, every direct call
3397 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3399 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3401 // We should use extra load for direct calls to dllimported functions in
3403 const GlobalValue *GV = G->getGlobal();
3404 if (!GV->hasDLLImportStorageClass()) {
3405 unsigned char OpFlags = 0;
3406 bool ExtraLoad = false;
3407 unsigned WrapperKind = ISD::DELETED_NODE;
3409 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3410 // external symbols most go through the PLT in PIC mode. If the symbol
3411 // has hidden or protected visibility, or if it is static or local, then
3412 // we don't need to use the PLT - we can directly call it.
3413 if (Subtarget->isTargetELF() &&
3414 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3415 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3416 OpFlags = X86II::MO_PLT;
3417 } else if (Subtarget->isPICStyleStubAny() &&
3418 !GV->isStrongDefinitionForLinker() &&
3419 (!Subtarget->getTargetTriple().isMacOSX() ||
3420 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3421 // PC-relative references to external symbols should go through $stub,
3422 // unless we're building with the leopard linker or later, which
3423 // automatically synthesizes these stubs.
3424 OpFlags = X86II::MO_DARWIN_STUB;
3425 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3426 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3427 // If the function is marked as non-lazy, generate an indirect call
3428 // which loads from the GOT directly. This avoids runtime overhead
3429 // at the cost of eager binding (and one extra byte of encoding).
3430 OpFlags = X86II::MO_GOTPCREL;
3431 WrapperKind = X86ISD::WrapperRIP;
3435 Callee = DAG.getTargetGlobalAddress(
3436 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3438 // Add a wrapper if needed.
3439 if (WrapperKind != ISD::DELETED_NODE)
3440 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3441 getPointerTy(DAG.getDataLayout()), Callee);
3442 // Add extra indirection if needed.
3444 Callee = DAG.getLoad(
3445 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3446 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3449 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3450 unsigned char OpFlags = 0;
3452 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3453 // external symbols should go through the PLT.
3454 if (Subtarget->isTargetELF() &&
3455 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3456 OpFlags = X86II::MO_PLT;
3457 } else if (Subtarget->isPICStyleStubAny() &&
3458 (!Subtarget->getTargetTriple().isMacOSX() ||
3459 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3460 // PC-relative references to external symbols should go through $stub,
3461 // unless we're building with the leopard linker or later, which
3462 // automatically synthesizes these stubs.
3463 OpFlags = X86II::MO_DARWIN_STUB;
3466 Callee = DAG.getTargetExternalSymbol(
3467 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3468 } else if (Subtarget->isTarget64BitILP32() &&
3469 Callee->getValueType(0) == MVT::i32) {
3470 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3471 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3474 // Returns a chain & a flag for retval copy to use.
3475 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3476 SmallVector<SDValue, 8> Ops;
3478 if (!IsSibcall && isTailCall) {
3479 Chain = DAG.getCALLSEQ_END(Chain,
3480 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3481 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3482 InFlag = Chain.getValue(1);
3485 Ops.push_back(Chain);
3486 Ops.push_back(Callee);
3489 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3491 // Add argument registers to the end of the list so that they are known live
3493 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3494 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3495 RegsToPass[i].second.getValueType()));
3497 // Add a register mask operand representing the call-preserved registers.
3498 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3499 assert(Mask && "Missing call preserved mask for calling convention");
3501 // If this is an invoke in a 32-bit function using a funclet-based
3502 // personality, assume the function clobbers all registers. If an exception
3503 // is thrown, the runtime will not restore CSRs.
3504 // FIXME: Model this more precisely so that we can register allocate across
3505 // the normal edge and spill and fill across the exceptional edge.
3506 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3507 const Function *CallerFn = MF.getFunction();
3508 EHPersonality Pers =
3509 CallerFn->hasPersonalityFn()
3510 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3511 : EHPersonality::Unknown;
3512 if (isFuncletEHPersonality(Pers))
3513 Mask = RegInfo->getNoPreservedMask();
3516 Ops.push_back(DAG.getRegisterMask(Mask));
3518 if (InFlag.getNode())
3519 Ops.push_back(InFlag);
3523 //// If this is the first return lowered for this function, add the regs
3524 //// to the liveout set for the function.
3525 // This isn't right, although it's probably harmless on x86; liveouts
3526 // should be computed from returns not tail calls. Consider a void
3527 // function making a tail call to a function returning int.
3528 MF.getFrameInfo()->setHasTailCall();
3529 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3532 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3533 InFlag = Chain.getValue(1);
3535 // Create the CALLSEQ_END node.
3536 unsigned NumBytesForCalleeToPop;
3537 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3538 DAG.getTarget().Options.GuaranteedTailCallOpt))
3539 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3540 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3541 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3542 SR == StackStructReturn)
3543 // If this is a call to a struct-return function, the callee
3544 // pops the hidden struct pointer, so we have to push it back.
3545 // This is common for Darwin/X86, Linux & Mingw32 targets.
3546 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3547 NumBytesForCalleeToPop = 4;
3549 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3551 // Returns a flag for retval copy to use.
3553 Chain = DAG.getCALLSEQ_END(Chain,
3554 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3555 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3558 InFlag = Chain.getValue(1);
3561 // Handle result values, copying them out of physregs into vregs that we
3563 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3564 Ins, dl, DAG, InVals);
3567 //===----------------------------------------------------------------------===//
3568 // Fast Calling Convention (tail call) implementation
3569 //===----------------------------------------------------------------------===//
3571 // Like std call, callee cleans arguments, convention except that ECX is
3572 // reserved for storing the tail called function address. Only 2 registers are
3573 // free for argument passing (inreg). Tail call optimization is performed
3575 // * tailcallopt is enabled
3576 // * caller/callee are fastcc
3577 // On X86_64 architecture with GOT-style position independent code only local
3578 // (within module) calls are supported at the moment.
3579 // To keep the stack aligned according to platform abi the function
3580 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3581 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3582 // If a tail called function callee has more arguments than the caller the
3583 // caller needs to make sure that there is room to move the RETADDR to. This is
3584 // achieved by reserving an area the size of the argument delta right after the
3585 // original RETADDR, but before the saved framepointer or the spilled registers
3586 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3598 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3601 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3602 SelectionDAG& DAG) const {
3603 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3604 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3605 unsigned StackAlignment = TFI.getStackAlignment();
3606 uint64_t AlignMask = StackAlignment - 1;
3607 int64_t Offset = StackSize;
3608 unsigned SlotSize = RegInfo->getSlotSize();
3609 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3610 // Number smaller than 12 so just add the difference.
3611 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3613 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3614 Offset = ((~AlignMask) & Offset) + StackAlignment +
3615 (StackAlignment-SlotSize);
3620 /// Return true if the given stack call argument is already available in the
3621 /// same position (relatively) of the caller's incoming argument stack.
3623 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3624 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3625 const X86InstrInfo *TII) {
3626 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3628 if (Arg.getOpcode() == ISD::CopyFromReg) {
3629 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3630 if (!TargetRegisterInfo::isVirtualRegister(VR))
3632 MachineInstr *Def = MRI->getVRegDef(VR);
3635 if (!Flags.isByVal()) {
3636 if (!TII->isLoadFromStackSlot(Def, FI))
3639 unsigned Opcode = Def->getOpcode();
3640 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3641 Opcode == X86::LEA64_32r) &&
3642 Def->getOperand(1).isFI()) {
3643 FI = Def->getOperand(1).getIndex();
3644 Bytes = Flags.getByValSize();
3648 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3649 if (Flags.isByVal())
3650 // ByVal argument is passed in as a pointer but it's now being
3651 // dereferenced. e.g.
3652 // define @foo(%struct.X* %A) {
3653 // tail call @bar(%struct.X* byval %A)
3656 SDValue Ptr = Ld->getBasePtr();
3657 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3660 FI = FINode->getIndex();
3661 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3662 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3663 FI = FINode->getIndex();
3664 Bytes = Flags.getByValSize();
3668 assert(FI != INT_MAX);
3669 if (!MFI->isFixedObjectIndex(FI))
3671 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3674 /// Check whether the call is eligible for tail call optimization. Targets
3675 /// that want to do tail call optimization should implement this function.
3676 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3677 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3678 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3679 const SmallVectorImpl<ISD::OutputArg> &Outs,
3680 const SmallVectorImpl<SDValue> &OutVals,
3681 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3682 if (!mayTailCallThisCC(CalleeCC))
3685 // If -tailcallopt is specified, make fastcc functions tail-callable.
3686 MachineFunction &MF = DAG.getMachineFunction();
3687 const Function *CallerF = MF.getFunction();
3689 // If the function return type is x86_fp80 and the callee return type is not,
3690 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3691 // perform a tailcall optimization here.
3692 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3695 CallingConv::ID CallerCC = CallerF->getCallingConv();
3696 bool CCMatch = CallerCC == CalleeCC;
3697 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3698 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3700 // Win64 functions have extra shadow space for argument homing. Don't do the
3701 // sibcall if the caller and callee have mismatched expectations for this
3703 if (IsCalleeWin64 != IsCallerWin64)
3706 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3707 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3712 // Look for obvious safe cases to perform tail call optimization that do not
3713 // require ABI changes. This is what gcc calls sibcall.
3715 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3716 // emit a special epilogue.
3717 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3718 if (RegInfo->needsStackRealignment(MF))
3721 // Also avoid sibcall optimization if either caller or callee uses struct
3722 // return semantics.
3723 if (isCalleeStructRet || isCallerStructRet)
3726 // Do not sibcall optimize vararg calls unless all arguments are passed via
3728 if (isVarArg && !Outs.empty()) {
3729 // Optimizing for varargs on Win64 is unlikely to be safe without
3730 // additional testing.
3731 if (IsCalleeWin64 || IsCallerWin64)
3734 SmallVector<CCValAssign, 16> ArgLocs;
3735 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3738 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3739 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3740 if (!ArgLocs[i].isRegLoc())
3744 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3745 // stack. Therefore, if it's not used by the call it is not safe to optimize
3746 // this into a sibcall.
3747 bool Unused = false;
3748 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3755 SmallVector<CCValAssign, 16> RVLocs;
3756 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3758 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3759 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3760 CCValAssign &VA = RVLocs[i];
3761 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3766 // If the calling conventions do not match, then we'd better make sure the
3767 // results are returned in the same way as what the caller expects.
3769 SmallVector<CCValAssign, 16> RVLocs1;
3770 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3772 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3774 SmallVector<CCValAssign, 16> RVLocs2;
3775 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3777 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3779 if (RVLocs1.size() != RVLocs2.size())
3781 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3782 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3784 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3786 if (RVLocs1[i].isRegLoc()) {
3787 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3790 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3796 unsigned StackArgsSize = 0;
3798 // If the callee takes no arguments then go on to check the results of the
3800 if (!Outs.empty()) {
3801 // Check if stack adjustment is needed. For now, do not do this if any
3802 // argument is passed on the stack.
3803 SmallVector<CCValAssign, 16> ArgLocs;
3804 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3807 // Allocate shadow area for Win64
3809 CCInfo.AllocateStack(32, 8);
3811 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3812 StackArgsSize = CCInfo.getNextStackOffset();
3814 if (CCInfo.getNextStackOffset()) {
3815 // Check if the arguments are already laid out in the right way as
3816 // the caller's fixed stack objects.
3817 MachineFrameInfo *MFI = MF.getFrameInfo();
3818 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3819 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3820 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3821 CCValAssign &VA = ArgLocs[i];
3822 SDValue Arg = OutVals[i];
3823 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3824 if (VA.getLocInfo() == CCValAssign::Indirect)
3826 if (!VA.isRegLoc()) {
3827 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3834 // If the tailcall address may be in a register, then make sure it's
3835 // possible to register allocate for it. In 32-bit, the call address can
3836 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3837 // callee-saved registers are restored. These happen to be the same
3838 // registers used to pass 'inreg' arguments so watch out for those.
3839 if (!Subtarget->is64Bit() &&
3840 ((!isa<GlobalAddressSDNode>(Callee) &&
3841 !isa<ExternalSymbolSDNode>(Callee)) ||
3842 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3843 unsigned NumInRegs = 0;
3844 // In PIC we need an extra register to formulate the address computation
3846 unsigned MaxInRegs =
3847 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3849 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3850 CCValAssign &VA = ArgLocs[i];
3853 unsigned Reg = VA.getLocReg();
3856 case X86::EAX: case X86::EDX: case X86::ECX:
3857 if (++NumInRegs == MaxInRegs)
3865 bool CalleeWillPop =
3866 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3867 MF.getTarget().Options.GuaranteedTailCallOpt);
3869 if (unsigned BytesToPop =
3870 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3871 // If we have bytes to pop, the callee must pop them.
3872 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3873 if (!CalleePopMatches)
3875 } else if (CalleeWillPop && StackArgsSize > 0) {
3876 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3884 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3885 const TargetLibraryInfo *libInfo) const {
3886 return X86::createFastISel(funcInfo, libInfo);
3889 //===----------------------------------------------------------------------===//
3890 // Other Lowering Hooks
3891 //===----------------------------------------------------------------------===//
3893 static bool MayFoldLoad(SDValue Op) {
3894 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3897 static bool MayFoldIntoStore(SDValue Op) {
3898 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3901 static bool isTargetShuffle(unsigned Opcode) {
3903 default: return false;
3904 case X86ISD::BLENDI:
3905 case X86ISD::PSHUFB:
3906 case X86ISD::PSHUFD:
3907 case X86ISD::PSHUFHW:
3908 case X86ISD::PSHUFLW:
3910 case X86ISD::INSERTPS:
3911 case X86ISD::PALIGNR:
3912 case X86ISD::MOVLHPS:
3913 case X86ISD::MOVLHPD:
3914 case X86ISD::MOVHLPS:
3915 case X86ISD::MOVLPS:
3916 case X86ISD::MOVLPD:
3917 case X86ISD::MOVSHDUP:
3918 case X86ISD::MOVSLDUP:
3919 case X86ISD::MOVDDUP:
3922 case X86ISD::UNPCKL:
3923 case X86ISD::UNPCKH:
3924 case X86ISD::VPERMILPI:
3925 case X86ISD::VPERM2X128:
3926 case X86ISD::VPERMI:
3927 case X86ISD::VPERMV:
3928 case X86ISD::VPERMV3:
3933 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3934 SDValue V1, unsigned TargetMask,
3935 SelectionDAG &DAG) {
3937 default: llvm_unreachable("Unknown x86 shuffle node");
3938 case X86ISD::PSHUFD:
3939 case X86ISD::PSHUFHW:
3940 case X86ISD::PSHUFLW:
3941 case X86ISD::VPERMILPI:
3942 case X86ISD::VPERMI:
3943 return DAG.getNode(Opc, dl, VT, V1,
3944 DAG.getConstant(TargetMask, dl, MVT::i8));
3948 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3949 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3951 default: llvm_unreachable("Unknown x86 shuffle node");
3952 case X86ISD::MOVLHPS:
3953 case X86ISD::MOVLHPD:
3954 case X86ISD::MOVHLPS:
3955 case X86ISD::MOVLPS:
3956 case X86ISD::MOVLPD:
3959 case X86ISD::UNPCKL:
3960 case X86ISD::UNPCKH:
3961 return DAG.getNode(Opc, dl, VT, V1, V2);
3965 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3966 MachineFunction &MF = DAG.getMachineFunction();
3967 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3968 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3969 int ReturnAddrIndex = FuncInfo->getRAIndex();
3971 if (ReturnAddrIndex == 0) {
3972 // Set up a frame object for the return address.
3973 unsigned SlotSize = RegInfo->getSlotSize();
3974 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3977 FuncInfo->setRAIndex(ReturnAddrIndex);
3980 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3983 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3984 bool hasSymbolicDisplacement) {
3985 // Offset should fit into 32 bit immediate field.
3986 if (!isInt<32>(Offset))
3989 // If we don't have a symbolic displacement - we don't have any extra
3991 if (!hasSymbolicDisplacement)
3994 // FIXME: Some tweaks might be needed for medium code model.
3995 if (M != CodeModel::Small && M != CodeModel::Kernel)
3998 // For small code model we assume that latest object is 16MB before end of 31
3999 // bits boundary. We may also accept pretty large negative constants knowing
4000 // that all objects are in the positive half of address space.
4001 if (M == CodeModel::Small && Offset < 16*1024*1024)
4004 // For kernel code model we know that all object resist in the negative half
4005 // of 32bits address space. We may not accept negative offsets, since they may
4006 // be just off and we may accept pretty large positive ones.
4007 if (M == CodeModel::Kernel && Offset >= 0)
4013 /// Determines whether the callee is required to pop its own arguments.
4014 /// Callee pop is necessary to support tail calls.
4015 bool X86::isCalleePop(CallingConv::ID CallingConv,
4016 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
4017 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
4018 // can guarantee TCO.
4019 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
4022 switch (CallingConv) {
4025 case CallingConv::X86_StdCall:
4026 case CallingConv::X86_FastCall:
4027 case CallingConv::X86_ThisCall:
4028 case CallingConv::X86_VectorCall:
4033 /// \brief Return true if the condition is an unsigned comparison operation.
4034 static bool isX86CCUnsigned(unsigned X86CC) {
4036 default: llvm_unreachable("Invalid integer condition!");
4037 case X86::COND_E: return true;
4038 case X86::COND_G: return false;
4039 case X86::COND_GE: return false;
4040 case X86::COND_L: return false;
4041 case X86::COND_LE: return false;
4042 case X86::COND_NE: return true;
4043 case X86::COND_B: return true;
4044 case X86::COND_A: return true;
4045 case X86::COND_BE: return true;
4046 case X86::COND_AE: return true;
4050 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
4051 switch (SetCCOpcode) {
4052 default: llvm_unreachable("Invalid integer condition!");
4053 case ISD::SETEQ: return X86::COND_E;
4054 case ISD::SETGT: return X86::COND_G;
4055 case ISD::SETGE: return X86::COND_GE;
4056 case ISD::SETLT: return X86::COND_L;
4057 case ISD::SETLE: return X86::COND_LE;
4058 case ISD::SETNE: return X86::COND_NE;
4059 case ISD::SETULT: return X86::COND_B;
4060 case ISD::SETUGT: return X86::COND_A;
4061 case ISD::SETULE: return X86::COND_BE;
4062 case ISD::SETUGE: return X86::COND_AE;
4066 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
4067 /// condition code, returning the condition code and the LHS/RHS of the
4068 /// comparison to make.
4069 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4070 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4072 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4073 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4074 // X > -1 -> X == 0, jump !sign.
4075 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4076 return X86::COND_NS;
4078 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4079 // X < 0 -> X == 0, jump on sign.
4082 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4084 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4085 return X86::COND_LE;
4089 return TranslateIntegerX86CC(SetCCOpcode);
4092 // First determine if it is required or is profitable to flip the operands.
4094 // If LHS is a foldable load, but RHS is not, flip the condition.
4095 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4096 !ISD::isNON_EXTLoad(RHS.getNode())) {
4097 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4098 std::swap(LHS, RHS);
4101 switch (SetCCOpcode) {
4107 std::swap(LHS, RHS);
4111 // On a floating point condition, the flags are set as follows:
4113 // 0 | 0 | 0 | X > Y
4114 // 0 | 0 | 1 | X < Y
4115 // 1 | 0 | 0 | X == Y
4116 // 1 | 1 | 1 | unordered
4117 switch (SetCCOpcode) {
4118 default: llvm_unreachable("Condcode should be pre-legalized away");
4120 case ISD::SETEQ: return X86::COND_E;
4121 case ISD::SETOLT: // flipped
4123 case ISD::SETGT: return X86::COND_A;
4124 case ISD::SETOLE: // flipped
4126 case ISD::SETGE: return X86::COND_AE;
4127 case ISD::SETUGT: // flipped
4129 case ISD::SETLT: return X86::COND_B;
4130 case ISD::SETUGE: // flipped
4132 case ISD::SETLE: return X86::COND_BE;
4134 case ISD::SETNE: return X86::COND_NE;
4135 case ISD::SETUO: return X86::COND_P;
4136 case ISD::SETO: return X86::COND_NP;
4138 case ISD::SETUNE: return X86::COND_INVALID;
4142 /// Is there a floating point cmov for the specific X86 condition code?
4143 /// Current x86 isa includes the following FP cmov instructions:
4144 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4145 static bool hasFPCMov(unsigned X86CC) {
4161 /// Returns true if the target can instruction select the
4162 /// specified FP immediate natively. If false, the legalizer will
4163 /// materialize the FP immediate as a load from a constant pool.
4164 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4165 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4166 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4172 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4173 ISD::LoadExtType ExtTy,
4175 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4176 // relocation target a movq or addq instruction: don't let the load shrink.
4177 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4178 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4179 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4180 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4184 /// \brief Returns true if it is beneficial to convert a load of a constant
4185 /// to just the constant itself.
4186 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4188 assert(Ty->isIntegerTy());
4190 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4191 if (BitSize == 0 || BitSize > 64)
4196 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4197 unsigned Index) const {
4198 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4201 return (Index == 0 || Index == ResVT.getVectorNumElements());
4204 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4205 // Speculate cttz only if we can directly use TZCNT.
4206 return Subtarget->hasBMI();
4209 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4210 // Speculate ctlz only if we can directly use LZCNT.
4211 return Subtarget->hasLZCNT();
4214 /// Return true if every element in Mask, beginning
4215 /// from position Pos and ending in Pos+Size is undef.
4216 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4217 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4223 /// Return true if Val is undef or if its value falls within the
4224 /// specified range (L, H].
4225 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4226 return (Val < 0) || (Val >= Low && Val < Hi);
4229 /// Val is either less than zero (undef) or equal to the specified value.
4230 static bool isUndefOrEqual(int Val, int CmpVal) {
4231 return (Val < 0 || Val == CmpVal);
4234 /// Return true if every element in Mask, beginning
4235 /// from position Pos and ending in Pos+Size, falls within the specified
4236 /// sequential range (Low, Low+Size]. or is undef.
4237 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4238 unsigned Pos, unsigned Size, int Low) {
4239 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4240 if (!isUndefOrEqual(Mask[i], Low))
4245 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4246 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4247 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4248 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4249 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4252 // The index should be aligned on a vecWidth-bit boundary.
4254 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4256 MVT VT = N->getSimpleValueType(0);
4257 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4258 bool Result = (Index * ElSize) % vecWidth == 0;
4263 /// Return true if the specified INSERT_SUBVECTOR
4264 /// operand specifies a subvector insert that is suitable for input to
4265 /// insertion of 128 or 256-bit subvectors
4266 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4267 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4268 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4270 // The index should be aligned on a vecWidth-bit boundary.
4272 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4274 MVT VT = N->getSimpleValueType(0);
4275 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4276 bool Result = (Index * ElSize) % vecWidth == 0;
4281 bool X86::isVINSERT128Index(SDNode *N) {
4282 return isVINSERTIndex(N, 128);
4285 bool X86::isVINSERT256Index(SDNode *N) {
4286 return isVINSERTIndex(N, 256);
4289 bool X86::isVEXTRACT128Index(SDNode *N) {
4290 return isVEXTRACTIndex(N, 128);
4293 bool X86::isVEXTRACT256Index(SDNode *N) {
4294 return isVEXTRACTIndex(N, 256);
4297 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4298 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4299 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4300 "Illegal extract subvector for VEXTRACT");
4303 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4305 MVT VecVT = N->getOperand(0).getSimpleValueType();
4306 MVT ElVT = VecVT.getVectorElementType();
4308 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4309 return Index / NumElemsPerChunk;
4312 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4313 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4314 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4315 "Illegal insert subvector for VINSERT");
4318 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4320 MVT VecVT = N->getSimpleValueType(0);
4321 MVT ElVT = VecVT.getVectorElementType();
4323 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4324 return Index / NumElemsPerChunk;
4327 /// Return the appropriate immediate to extract the specified
4328 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4329 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4330 return getExtractVEXTRACTImmediate(N, 128);
4333 /// Return the appropriate immediate to extract the specified
4334 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4335 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4336 return getExtractVEXTRACTImmediate(N, 256);
4339 /// Return the appropriate immediate to insert at the specified
4340 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4341 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4342 return getInsertVINSERTImmediate(N, 128);
4345 /// Return the appropriate immediate to insert at the specified
4346 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4347 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4348 return getInsertVINSERTImmediate(N, 256);
4351 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4352 bool X86::isZeroNode(SDValue Elt) {
4353 return isNullConstant(Elt) || isNullFPConstant(Elt);
4356 // Build a vector of constants
4357 // Use an UNDEF node if MaskElt == -1.
4358 // Spilt 64-bit constants in the 32-bit mode.
4359 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4361 SDLoc dl, bool IsMask = false) {
4363 SmallVector<SDValue, 32> Ops;
4366 MVT ConstVecVT = VT;
4367 unsigned NumElts = VT.getVectorNumElements();
4368 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4369 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4370 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4374 MVT EltVT = ConstVecVT.getVectorElementType();
4375 for (unsigned i = 0; i < NumElts; ++i) {
4376 bool IsUndef = Values[i] < 0 && IsMask;
4377 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4378 DAG.getConstant(Values[i], dl, EltVT);
4379 Ops.push_back(OpNode);
4381 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4382 DAG.getConstant(0, dl, EltVT));
4384 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4386 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4390 /// Returns a vector of specified type with all zero elements.
4391 static SDValue getZeroVector(MVT VT, const X86Subtarget *Subtarget,
4392 SelectionDAG &DAG, SDLoc dl) {
4393 assert(VT.isVector() && "Expected a vector type");
4395 // Always build SSE zero vectors as <4 x i32> bitcasted
4396 // to their dest type. This ensures they get CSE'd.
4398 if (VT.is128BitVector()) { // SSE
4399 if (Subtarget->hasSSE2()) { // SSE2
4400 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4401 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4403 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4404 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4406 } else if (VT.is256BitVector()) { // AVX
4407 if (Subtarget->hasInt256()) { // AVX2
4408 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4409 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4410 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4412 // 256-bit logic and arithmetic instructions in AVX are all
4413 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4414 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4415 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4416 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4418 } else if (VT.is512BitVector()) { // AVX-512
4419 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4420 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4421 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4422 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4423 } else if (VT.getVectorElementType() == MVT::i1) {
4425 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4426 && "Unexpected vector type");
4427 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4428 && "Unexpected vector type");
4429 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4430 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4431 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4433 llvm_unreachable("Unexpected vector type");
4435 return DAG.getBitcast(VT, Vec);
4438 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4439 SelectionDAG &DAG, SDLoc dl,
4440 unsigned vectorWidth) {
4441 assert((vectorWidth == 128 || vectorWidth == 256) &&
4442 "Unsupported vector width");
4443 EVT VT = Vec.getValueType();
4444 EVT ElVT = VT.getVectorElementType();
4445 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4446 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4447 VT.getVectorNumElements()/Factor);
4449 // Extract from UNDEF is UNDEF.
4450 if (Vec.getOpcode() == ISD::UNDEF)
4451 return DAG.getUNDEF(ResultVT);
4453 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4454 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4455 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4457 // This is the index of the first element of the vectorWidth-bit chunk
4458 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4459 IdxVal &= ~(ElemsPerChunk - 1);
4461 // If the input is a buildvector just emit a smaller one.
4462 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4463 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4464 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4466 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4467 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4470 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4471 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4472 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4473 /// instructions or a simple subregister reference. Idx is an index in the
4474 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4475 /// lowering EXTRACT_VECTOR_ELT operations easier.
4476 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4477 SelectionDAG &DAG, SDLoc dl) {
4478 assert((Vec.getValueType().is256BitVector() ||
4479 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4480 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4483 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4484 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4485 SelectionDAG &DAG, SDLoc dl) {
4486 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4487 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4490 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4491 unsigned IdxVal, SelectionDAG &DAG,
4492 SDLoc dl, unsigned vectorWidth) {
4493 assert((vectorWidth == 128 || vectorWidth == 256) &&
4494 "Unsupported vector width");
4495 // Inserting UNDEF is Result
4496 if (Vec.getOpcode() == ISD::UNDEF)
4498 EVT VT = Vec.getValueType();
4499 EVT ElVT = VT.getVectorElementType();
4500 EVT ResultVT = Result.getValueType();
4502 // Insert the relevant vectorWidth bits.
4503 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4504 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4506 // This is the index of the first element of the vectorWidth-bit chunk
4507 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4508 IdxVal &= ~(ElemsPerChunk - 1);
4510 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4511 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4514 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4515 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4516 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4517 /// simple superregister reference. Idx is an index in the 128 bits
4518 /// we want. It need not be aligned to a 128-bit boundary. That makes
4519 /// lowering INSERT_VECTOR_ELT operations easier.
4520 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4521 SelectionDAG &DAG, SDLoc dl) {
4522 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4524 // For insertion into the zero index (low half) of a 256-bit vector, it is
4525 // more efficient to generate a blend with immediate instead of an insert*128.
4526 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4527 // extend the subvector to the size of the result vector. Make sure that
4528 // we are not recursing on that node by checking for undef here.
4529 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4530 Result.getOpcode() != ISD::UNDEF) {
4531 EVT ResultVT = Result.getValueType();
4532 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4533 SDValue Undef = DAG.getUNDEF(ResultVT);
4534 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4537 // The blend instruction, and therefore its mask, depend on the data type.
4538 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4539 if (ScalarType.isFloatingPoint()) {
4540 // Choose either vblendps (float) or vblendpd (double).
4541 unsigned ScalarSize = ScalarType.getSizeInBits();
4542 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4543 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4544 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4545 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4548 const X86Subtarget &Subtarget =
4549 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4551 // AVX2 is needed for 256-bit integer blend support.
4552 // Integers must be cast to 32-bit because there is only vpblendd;
4553 // vpblendw can't be used for this because it has a handicapped mask.
4555 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4556 // is still more efficient than using the wrong domain vinsertf128 that
4557 // will be created by InsertSubVector().
4558 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4560 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4561 Result = DAG.getBitcast(CastVT, Result);
4562 Vec256 = DAG.getBitcast(CastVT, Vec256);
4563 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4564 return DAG.getBitcast(ResultVT, Vec256);
4567 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4570 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4571 SelectionDAG &DAG, SDLoc dl) {
4572 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4573 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4576 /// Insert i1-subvector to i1-vector.
4577 static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
4580 SDValue Vec = Op.getOperand(0);
4581 SDValue SubVec = Op.getOperand(1);
4582 SDValue Idx = Op.getOperand(2);
4584 if (!isa<ConstantSDNode>(Idx))
4587 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
4588 if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
4591 MVT OpVT = Op.getSimpleValueType();
4592 MVT SubVecVT = SubVec.getSimpleValueType();
4593 unsigned NumElems = OpVT.getVectorNumElements();
4594 unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
4596 assert(IdxVal + SubVecNumElems <= NumElems &&
4597 IdxVal % SubVecVT.getSizeInBits() == 0 &&
4598 "Unexpected index value in INSERT_SUBVECTOR");
4600 // There are 3 possible cases:
4601 // 1. Subvector should be inserted in the lower part (IdxVal == 0)
4602 // 2. Subvector should be inserted in the upper part
4603 // (IdxVal + SubVecNumElems == NumElems)
4604 // 3. Subvector should be inserted in the middle (for example v2i1
4605 // to v16i1, index 2)
4607 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
4608 SDValue Undef = DAG.getUNDEF(OpVT);
4609 SDValue WideSubVec =
4610 DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
4612 return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4613 DAG.getConstant(IdxVal, dl, MVT::i8));
4615 if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
4616 unsigned ShiftLeft = NumElems - SubVecNumElems;
4617 unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
4618 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4619 DAG.getConstant(ShiftLeft, dl, MVT::i8));
4620 return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
4621 DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
4625 // Zero lower bits of the Vec
4626 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4627 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4628 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4629 // Merge them together
4630 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4633 // Simple case when we put subvector in the upper part
4634 if (IdxVal + SubVecNumElems == NumElems) {
4635 // Zero upper bits of the Vec
4636 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
4637 DAG.getConstant(IdxVal, dl, MVT::i8));
4638 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4639 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4640 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4641 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4643 // Subvector should be inserted in the middle - use shuffle
4644 SmallVector<int, 64> Mask;
4645 for (unsigned i = 0; i < NumElems; ++i)
4646 Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
4648 return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
4651 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4652 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4653 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4654 /// large BUILD_VECTORS.
4655 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4656 unsigned NumElems, SelectionDAG &DAG,
4658 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4659 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4662 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4663 unsigned NumElems, SelectionDAG &DAG,
4665 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4666 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4669 /// Returns a vector of specified type with all bits set.
4670 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4671 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4672 /// Then bitcast to their original type, ensuring they get CSE'd.
4673 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4674 SelectionDAG &DAG, SDLoc dl) {
4675 assert(VT.isVector() && "Expected a vector type");
4677 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4679 if (VT.is512BitVector()) {
4680 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4681 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4682 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4683 } else if (VT.is256BitVector()) {
4684 if (Subtarget->hasInt256()) { // AVX2
4685 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4686 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4688 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4689 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4691 } else if (VT.is128BitVector()) {
4692 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4694 llvm_unreachable("Unexpected vector type");
4696 return DAG.getBitcast(VT, Vec);
4699 /// Returns a vector_shuffle node for an unpackl operation.
4700 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4702 unsigned NumElems = VT.getVectorNumElements();
4703 SmallVector<int, 8> Mask;
4704 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4706 Mask.push_back(i + NumElems);
4708 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4711 /// Returns a vector_shuffle node for an unpackh operation.
4712 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4714 unsigned NumElems = VT.getVectorNumElements();
4715 SmallVector<int, 8> Mask;
4716 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4717 Mask.push_back(i + Half);
4718 Mask.push_back(i + NumElems + Half);
4720 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4723 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4724 /// This produces a shuffle where the low element of V2 is swizzled into the
4725 /// zero/undef vector, landing at element Idx.
4726 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4727 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4729 const X86Subtarget *Subtarget,
4730 SelectionDAG &DAG) {
4731 MVT VT = V2.getSimpleValueType();
4733 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4734 unsigned NumElems = VT.getVectorNumElements();
4735 SmallVector<int, 16> MaskVec;
4736 for (unsigned i = 0; i != NumElems; ++i)
4737 // If this is the insertion idx, put the low elt of V2 here.
4738 MaskVec.push_back(i == Idx ? NumElems : i);
4739 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4742 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4743 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4744 /// uses one source. Note that this will set IsUnary for shuffles which use a
4745 /// single input multiple times, and in those cases it will
4746 /// adjust the mask to only have indices within that single input.
4747 static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
4748 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4749 unsigned NumElems = VT.getVectorNumElements();
4753 bool IsFakeUnary = false;
4754 switch(N->getOpcode()) {
4755 case X86ISD::BLENDI:
4756 ImmN = N->getOperand(N->getNumOperands()-1);
4757 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4760 ImmN = N->getOperand(N->getNumOperands()-1);
4761 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4762 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4764 case X86ISD::INSERTPS:
4765 ImmN = N->getOperand(N->getNumOperands()-1);
4766 DecodeINSERTPSMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4767 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4769 case X86ISD::UNPCKH:
4770 DecodeUNPCKHMask(VT, Mask);
4771 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4773 case X86ISD::UNPCKL:
4774 DecodeUNPCKLMask(VT, Mask);
4775 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4777 case X86ISD::MOVHLPS:
4778 DecodeMOVHLPSMask(NumElems, Mask);
4779 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4781 case X86ISD::MOVLHPS:
4782 DecodeMOVLHPSMask(NumElems, Mask);
4783 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4785 case X86ISD::PALIGNR:
4786 ImmN = N->getOperand(N->getNumOperands()-1);
4787 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4789 case X86ISD::PSHUFD:
4790 case X86ISD::VPERMILPI:
4791 ImmN = N->getOperand(N->getNumOperands()-1);
4792 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4795 case X86ISD::PSHUFHW:
4796 ImmN = N->getOperand(N->getNumOperands()-1);
4797 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4800 case X86ISD::PSHUFLW:
4801 ImmN = N->getOperand(N->getNumOperands()-1);
4802 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4805 case X86ISD::PSHUFB: {
4807 SDValue MaskNode = N->getOperand(1);
4808 while (MaskNode->getOpcode() == ISD::BITCAST)
4809 MaskNode = MaskNode->getOperand(0);
4811 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4812 // If we have a build-vector, then things are easy.
4813 MVT VT = MaskNode.getSimpleValueType();
4814 assert(VT.isVector() &&
4815 "Can't produce a non-vector with a build_vector!");
4816 if (!VT.isInteger())
4819 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4821 SmallVector<uint64_t, 32> RawMask;
4822 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4823 SDValue Op = MaskNode->getOperand(i);
4824 if (Op->getOpcode() == ISD::UNDEF) {
4825 RawMask.push_back((uint64_t)SM_SentinelUndef);
4828 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4831 APInt MaskElement = CN->getAPIntValue();
4833 // We now have to decode the element which could be any integer size and
4834 // extract each byte of it.
4835 for (int j = 0; j < NumBytesPerElement; ++j) {
4836 // Note that this is x86 and so always little endian: the low byte is
4837 // the first byte of the mask.
4838 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4839 MaskElement = MaskElement.lshr(8);
4842 DecodePSHUFBMask(RawMask, Mask);
4846 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4850 SDValue Ptr = MaskLoad->getBasePtr();
4851 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4852 Ptr->getOpcode() == X86ISD::WrapperRIP)
4853 Ptr = Ptr->getOperand(0);
4855 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4856 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4859 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4860 DecodePSHUFBMask(C, Mask);
4866 case X86ISD::VPERMI:
4867 ImmN = N->getOperand(N->getNumOperands()-1);
4868 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4873 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4875 case X86ISD::VPERM2X128:
4876 ImmN = N->getOperand(N->getNumOperands()-1);
4877 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4878 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4880 case X86ISD::MOVSLDUP:
4881 DecodeMOVSLDUPMask(VT, Mask);
4884 case X86ISD::MOVSHDUP:
4885 DecodeMOVSHDUPMask(VT, Mask);
4888 case X86ISD::MOVDDUP:
4889 DecodeMOVDDUPMask(VT, Mask);
4892 case X86ISD::MOVLHPD:
4893 case X86ISD::MOVLPD:
4894 case X86ISD::MOVLPS:
4895 // Not yet implemented
4897 case X86ISD::VPERMV: {
4899 SDValue MaskNode = N->getOperand(0);
4900 while (MaskNode->getOpcode() == ISD::BITCAST)
4901 MaskNode = MaskNode->getOperand(0);
4903 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4904 SmallVector<uint64_t, 32> RawMask;
4905 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4906 // If we have a build-vector, then things are easy.
4907 assert(MaskNode.getSimpleValueType().isInteger() &&
4908 MaskNode.getSimpleValueType().getVectorNumElements() ==
4909 VT.getVectorNumElements());
4911 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4912 SDValue Op = MaskNode->getOperand(i);
4913 if (Op->getOpcode() == ISD::UNDEF)
4914 RawMask.push_back((uint64_t)SM_SentinelUndef);
4915 else if (isa<ConstantSDNode>(Op)) {
4916 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4917 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4921 DecodeVPERMVMask(RawMask, Mask);
4924 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4925 unsigned NumEltsInMask = MaskNode->getNumOperands();
4926 MaskNode = MaskNode->getOperand(0);
4927 if (auto *CN = dyn_cast<ConstantSDNode>(MaskNode)) {
4928 APInt MaskEltValue = CN->getAPIntValue();
4929 for (unsigned i = 0; i < NumEltsInMask; ++i)
4930 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4931 DecodeVPERMVMask(RawMask, Mask);
4934 // It may be a scalar load
4937 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4941 SDValue Ptr = MaskLoad->getBasePtr();
4942 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4943 Ptr->getOpcode() == X86ISD::WrapperRIP)
4944 Ptr = Ptr->getOperand(0);
4946 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4947 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4950 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4951 DecodeVPERMVMask(C, VT, Mask);
4956 case X86ISD::VPERMV3: {
4958 SDValue MaskNode = N->getOperand(1);
4959 while (MaskNode->getOpcode() == ISD::BITCAST)
4960 MaskNode = MaskNode->getOperand(1);
4962 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4963 // If we have a build-vector, then things are easy.
4964 assert(MaskNode.getSimpleValueType().isInteger() &&
4965 MaskNode.getSimpleValueType().getVectorNumElements() ==
4966 VT.getVectorNumElements());
4968 SmallVector<uint64_t, 32> RawMask;
4969 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4971 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4972 SDValue Op = MaskNode->getOperand(i);
4973 if (Op->getOpcode() == ISD::UNDEF)
4974 RawMask.push_back((uint64_t)SM_SentinelUndef);
4976 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4979 APInt MaskElement = CN->getAPIntValue();
4980 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4983 DecodeVPERMV3Mask(RawMask, Mask);
4987 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4991 SDValue Ptr = MaskLoad->getBasePtr();
4992 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4993 Ptr->getOpcode() == X86ISD::WrapperRIP)
4994 Ptr = Ptr->getOperand(0);
4996 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4997 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5000 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5001 DecodeVPERMV3Mask(C, VT, Mask);
5006 default: llvm_unreachable("unknown target shuffle node");
5009 // Empty mask indicates the decode failed.
5013 // Check if we're getting a shuffle mask with zero'd elements.
5014 if (!AllowSentinelZero)
5015 if (std::any_of(Mask.begin(), Mask.end(),
5016 [](int M){ return M == SM_SentinelZero; }))
5019 // If we have a fake unary shuffle, the shuffle mask is spread across two
5020 // inputs that are actually the same node. Re-map the mask to always point
5021 // into the first input.
5024 if (M >= (int)Mask.size())
5030 /// Returns the scalar element that will make up the ith
5031 /// element of the result of the vector shuffle.
5032 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5035 return SDValue(); // Limit search depth.
5037 SDValue V = SDValue(N, 0);
5038 EVT VT = V.getValueType();
5039 unsigned Opcode = V.getOpcode();
5041 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5042 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5043 int Elt = SV->getMaskElt(Index);
5046 return DAG.getUNDEF(VT.getVectorElementType());
5048 unsigned NumElems = VT.getVectorNumElements();
5049 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5050 : SV->getOperand(1);
5051 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5054 // Recurse into target specific vector shuffles to find scalars.
5055 if (isTargetShuffle(Opcode)) {
5056 MVT ShufVT = V.getSimpleValueType();
5057 int NumElems = (int)ShufVT.getVectorNumElements();
5058 SmallVector<int, 16> ShuffleMask;
5061 if (!getTargetShuffleMask(N, ShufVT, false, ShuffleMask, IsUnary))
5064 int Elt = ShuffleMask[Index];
5065 if (Elt == SM_SentinelUndef)
5066 return DAG.getUNDEF(ShufVT.getVectorElementType());
5068 assert(0 <= Elt && Elt < (2*NumElems) && "Shuffle index out of range");
5069 SDValue NewV = (Elt < NumElems) ? N->getOperand(0) : N->getOperand(1);
5070 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5074 // Actual nodes that may contain scalar elements
5075 if (Opcode == ISD::BITCAST) {
5076 V = V.getOperand(0);
5077 EVT SrcVT = V.getValueType();
5078 unsigned NumElems = VT.getVectorNumElements();
5080 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5084 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5085 return (Index == 0) ? V.getOperand(0)
5086 : DAG.getUNDEF(VT.getVectorElementType());
5088 if (V.getOpcode() == ISD::BUILD_VECTOR)
5089 return V.getOperand(Index);
5094 /// Custom lower build_vector of v16i8.
5095 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5096 unsigned NumNonZero, unsigned NumZero,
5098 const X86Subtarget* Subtarget,
5099 const TargetLowering &TLI) {
5107 // SSE4.1 - use PINSRB to insert each byte directly.
5108 if (Subtarget->hasSSE41()) {
5109 for (unsigned i = 0; i < 16; ++i) {
5110 bool isNonZero = (NonZeros & (1 << i)) != 0;
5114 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
5116 V = DAG.getUNDEF(MVT::v16i8);
5119 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5120 MVT::v16i8, V, Op.getOperand(i),
5121 DAG.getIntPtrConstant(i, dl));
5128 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5129 for (unsigned i = 0; i < 16; ++i) {
5130 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5131 if (ThisIsNonZero && First) {
5133 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5135 V = DAG.getUNDEF(MVT::v8i16);
5140 SDValue ThisElt, LastElt;
5141 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5142 if (LastIsNonZero) {
5143 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5144 MVT::i16, Op.getOperand(i-1));
5146 if (ThisIsNonZero) {
5147 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5148 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5149 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5151 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5155 if (ThisElt.getNode())
5156 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5157 DAG.getIntPtrConstant(i/2, dl));
5161 return DAG.getBitcast(MVT::v16i8, V);
5164 /// Custom lower build_vector of v8i16.
5165 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5166 unsigned NumNonZero, unsigned NumZero,
5168 const X86Subtarget* Subtarget,
5169 const TargetLowering &TLI) {
5176 for (unsigned i = 0; i < 8; ++i) {
5177 bool isNonZero = (NonZeros & (1 << i)) != 0;
5181 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5183 V = DAG.getUNDEF(MVT::v8i16);
5186 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5187 MVT::v8i16, V, Op.getOperand(i),
5188 DAG.getIntPtrConstant(i, dl));
5195 /// Custom lower build_vector of v4i32 or v4f32.
5196 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5197 const X86Subtarget *Subtarget,
5198 const TargetLowering &TLI) {
5199 // Find all zeroable elements.
5200 std::bitset<4> Zeroable;
5201 for (int i=0; i < 4; ++i) {
5202 SDValue Elt = Op->getOperand(i);
5203 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5205 assert(Zeroable.size() - Zeroable.count() > 1 &&
5206 "We expect at least two non-zero elements!");
5208 // We only know how to deal with build_vector nodes where elements are either
5209 // zeroable or extract_vector_elt with constant index.
5210 SDValue FirstNonZero;
5211 unsigned FirstNonZeroIdx;
5212 for (unsigned i=0; i < 4; ++i) {
5215 SDValue Elt = Op->getOperand(i);
5216 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5217 !isa<ConstantSDNode>(Elt.getOperand(1)))
5219 // Make sure that this node is extracting from a 128-bit vector.
5220 MVT VT = Elt.getOperand(0).getSimpleValueType();
5221 if (!VT.is128BitVector())
5223 if (!FirstNonZero.getNode()) {
5225 FirstNonZeroIdx = i;
5229 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5230 SDValue V1 = FirstNonZero.getOperand(0);
5231 MVT VT = V1.getSimpleValueType();
5233 // See if this build_vector can be lowered as a blend with zero.
5235 unsigned EltMaskIdx, EltIdx;
5237 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5238 if (Zeroable[EltIdx]) {
5239 // The zero vector will be on the right hand side.
5240 Mask[EltIdx] = EltIdx+4;
5244 Elt = Op->getOperand(EltIdx);
5245 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5246 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5247 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5249 Mask[EltIdx] = EltIdx;
5253 // Let the shuffle legalizer deal with blend operations.
5254 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5255 if (V1.getSimpleValueType() != VT)
5256 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5257 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5260 // See if we can lower this build_vector to a INSERTPS.
5261 if (!Subtarget->hasSSE41())
5264 SDValue V2 = Elt.getOperand(0);
5265 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5268 bool CanFold = true;
5269 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5273 SDValue Current = Op->getOperand(i);
5274 SDValue SrcVector = Current->getOperand(0);
5277 CanFold = SrcVector == V1 &&
5278 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5284 assert(V1.getNode() && "Expected at least two non-zero elements!");
5285 if (V1.getSimpleValueType() != MVT::v4f32)
5286 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5287 if (V2.getSimpleValueType() != MVT::v4f32)
5288 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5290 // Ok, we can emit an INSERTPS instruction.
5291 unsigned ZMask = Zeroable.to_ulong();
5293 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5294 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5296 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5297 DAG.getIntPtrConstant(InsertPSMask, DL));
5298 return DAG.getBitcast(VT, Result);
5301 /// Return a vector logical shift node.
5302 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5303 unsigned NumBits, SelectionDAG &DAG,
5304 const TargetLowering &TLI, SDLoc dl) {
5305 assert(VT.is128BitVector() && "Unknown type for VShift");
5306 MVT ShVT = MVT::v2i64;
5307 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5308 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5309 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5310 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5311 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5312 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5316 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5318 // Check if the scalar load can be widened into a vector load. And if
5319 // the address is "base + cst" see if the cst can be "absorbed" into
5320 // the shuffle mask.
5321 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5322 SDValue Ptr = LD->getBasePtr();
5323 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5325 EVT PVT = LD->getValueType(0);
5326 if (PVT != MVT::i32 && PVT != MVT::f32)
5331 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5332 FI = FINode->getIndex();
5334 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5335 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5336 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5337 Offset = Ptr.getConstantOperandVal(1);
5338 Ptr = Ptr.getOperand(0);
5343 // FIXME: 256-bit vector instructions don't require a strict alignment,
5344 // improve this code to support it better.
5345 unsigned RequiredAlign = VT.getSizeInBits()/8;
5346 SDValue Chain = LD->getChain();
5347 // Make sure the stack object alignment is at least 16 or 32.
5348 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5349 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5350 if (MFI->isFixedObjectIndex(FI)) {
5351 // Can't change the alignment. FIXME: It's possible to compute
5352 // the exact stack offset and reference FI + adjust offset instead.
5353 // If someone *really* cares about this. That's the way to implement it.
5356 MFI->setObjectAlignment(FI, RequiredAlign);
5360 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5361 // Ptr + (Offset & ~15).
5364 if ((Offset % RequiredAlign) & 3)
5366 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5369 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5370 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5373 int EltNo = (Offset - StartOffset) >> 2;
5374 unsigned NumElems = VT.getVectorNumElements();
5376 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5377 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5378 LD->getPointerInfo().getWithOffset(StartOffset),
5379 false, false, false, 0);
5381 SmallVector<int, 8> Mask(NumElems, EltNo);
5383 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5389 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5390 /// elements can be replaced by a single large load which has the same value as
5391 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5393 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5395 /// FIXME: we'd also like to handle the case where the last elements are zero
5396 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5397 /// There's even a handy isZeroNode for that purpose.
5398 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5399 SDLoc &DL, SelectionDAG &DAG,
5400 bool isAfterLegalize) {
5401 unsigned NumElems = Elts.size();
5403 LoadSDNode *LDBase = nullptr;
5404 unsigned LastLoadedElt = -1U;
5406 // For each element in the initializer, see if we've found a load or an undef.
5407 // If we don't find an initial load element, or later load elements are
5408 // non-consecutive, bail out.
5409 for (unsigned i = 0; i < NumElems; ++i) {
5410 SDValue Elt = Elts[i];
5411 // Look through a bitcast.
5412 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5413 Elt = Elt.getOperand(0);
5414 if (!Elt.getNode() ||
5415 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5418 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5420 LDBase = cast<LoadSDNode>(Elt.getNode());
5424 if (Elt.getOpcode() == ISD::UNDEF)
5427 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5428 EVT LdVT = Elt.getValueType();
5429 // Each loaded element must be the correct fractional portion of the
5430 // requested vector load.
5431 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5433 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5438 // If we have found an entire vector of loads and undefs, then return a large
5439 // load of the entire vector width starting at the base pointer. If we found
5440 // consecutive loads for the low half, generate a vzext_load node.
5441 if (LastLoadedElt == NumElems - 1) {
5442 assert(LDBase && "Did not find base load for merging consecutive loads");
5443 EVT EltVT = LDBase->getValueType(0);
5444 // Ensure that the input vector size for the merged loads matches the
5445 // cumulative size of the input elements.
5446 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5449 if (isAfterLegalize &&
5450 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5453 SDValue NewLd = SDValue();
5455 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5456 LDBase->getPointerInfo(), LDBase->isVolatile(),
5457 LDBase->isNonTemporal(), LDBase->isInvariant(),
5458 LDBase->getAlignment());
5460 if (LDBase->hasAnyUseOfValue(1)) {
5461 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5463 SDValue(NewLd.getNode(), 1));
5464 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5465 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5466 SDValue(NewLd.getNode(), 1));
5472 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5473 //of a v4i32 / v4f32. It's probably worth generalizing.
5474 EVT EltVT = VT.getVectorElementType();
5475 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5476 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5477 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5478 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5480 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5481 LDBase->getPointerInfo(),
5482 LDBase->getAlignment(),
5483 false/*isVolatile*/, true/*ReadMem*/,
5486 // Make sure the newly-created LOAD is in the same position as LDBase in
5487 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5488 // update uses of LDBase's output chain to use the TokenFactor.
5489 if (LDBase->hasAnyUseOfValue(1)) {
5490 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5491 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5492 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5493 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5494 SDValue(ResNode.getNode(), 1));
5497 return DAG.getBitcast(VT, ResNode);
5502 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5503 /// to generate a splat value for the following cases:
5504 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5505 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5506 /// a scalar load, or a constant.
5507 /// The VBROADCAST node is returned when a pattern is found,
5508 /// or SDValue() otherwise.
5509 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5510 SelectionDAG &DAG) {
5511 // VBROADCAST requires AVX.
5512 // TODO: Splats could be generated for non-AVX CPUs using SSE
5513 // instructions, but there's less potential gain for only 128-bit vectors.
5514 if (!Subtarget->hasAVX())
5517 MVT VT = Op.getSimpleValueType();
5520 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5521 "Unsupported vector type for broadcast.");
5526 switch (Op.getOpcode()) {
5528 // Unknown pattern found.
5531 case ISD::BUILD_VECTOR: {
5532 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5533 BitVector UndefElements;
5534 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5536 // We need a splat of a single value to use broadcast, and it doesn't
5537 // make any sense if the value is only in one element of the vector.
5538 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5542 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5543 Ld.getOpcode() == ISD::ConstantFP);
5545 // Make sure that all of the users of a non-constant load are from the
5546 // BUILD_VECTOR node.
5547 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5552 case ISD::VECTOR_SHUFFLE: {
5553 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5555 // Shuffles must have a splat mask where the first element is
5557 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5560 SDValue Sc = Op.getOperand(0);
5561 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5562 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5564 if (!Subtarget->hasInt256())
5567 // Use the register form of the broadcast instruction available on AVX2.
5568 if (VT.getSizeInBits() >= 256)
5569 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5570 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5573 Ld = Sc.getOperand(0);
5574 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5575 Ld.getOpcode() == ISD::ConstantFP);
5577 // The scalar_to_vector node and the suspected
5578 // load node must have exactly one user.
5579 // Constants may have multiple users.
5581 // AVX-512 has register version of the broadcast
5582 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5583 Ld.getValueType().getSizeInBits() >= 32;
5584 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5591 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5592 bool IsGE256 = (VT.getSizeInBits() >= 256);
5594 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5595 // instruction to save 8 or more bytes of constant pool data.
5596 // TODO: If multiple splats are generated to load the same constant,
5597 // it may be detrimental to overall size. There needs to be a way to detect
5598 // that condition to know if this is truly a size win.
5599 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5601 // Handle broadcasting a single constant scalar from the constant pool
5603 // On Sandybridge (no AVX2), it is still better to load a constant vector
5604 // from the constant pool and not to broadcast it from a scalar.
5605 // But override that restriction when optimizing for size.
5606 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5607 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5608 EVT CVT = Ld.getValueType();
5609 assert(!CVT.isVector() && "Must not broadcast a vector type");
5611 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5612 // For size optimization, also splat v2f64 and v2i64, and for size opt
5613 // with AVX2, also splat i8 and i16.
5614 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5615 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5616 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5617 const Constant *C = nullptr;
5618 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5619 C = CI->getConstantIntValue();
5620 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5621 C = CF->getConstantFPValue();
5623 assert(C && "Invalid constant type");
5625 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5627 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5628 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5630 CVT, dl, DAG.getEntryNode(), CP,
5631 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5632 false, false, Alignment);
5634 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5638 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5640 // Handle AVX2 in-register broadcasts.
5641 if (!IsLoad && Subtarget->hasInt256() &&
5642 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5643 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5645 // The scalar source must be a normal load.
5649 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5650 (Subtarget->hasVLX() && ScalarSize == 64))
5651 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5653 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5654 // double since there is no vbroadcastsd xmm
5655 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5656 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5657 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5660 // Unsupported broadcast.
5664 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5665 /// underlying vector and index.
5667 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5669 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5671 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5672 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5675 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5677 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5679 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5680 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5683 // In this case the vector is the extract_subvector expression and the index
5684 // is 2, as specified by the shuffle.
5685 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5686 SDValue ShuffleVec = SVOp->getOperand(0);
5687 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5688 assert(ShuffleVecVT.getVectorElementType() ==
5689 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5691 int ShuffleIdx = SVOp->getMaskElt(Idx);
5692 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5693 ExtractedFromVec = ShuffleVec;
5699 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5700 MVT VT = Op.getSimpleValueType();
5702 // Skip if insert_vec_elt is not supported.
5703 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5704 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5708 unsigned NumElems = Op.getNumOperands();
5712 SmallVector<unsigned, 4> InsertIndices;
5713 SmallVector<int, 8> Mask(NumElems, -1);
5715 for (unsigned i = 0; i != NumElems; ++i) {
5716 unsigned Opc = Op.getOperand(i).getOpcode();
5718 if (Opc == ISD::UNDEF)
5721 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5722 // Quit if more than 1 elements need inserting.
5723 if (InsertIndices.size() > 1)
5726 InsertIndices.push_back(i);
5730 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5731 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5732 // Quit if non-constant index.
5733 if (!isa<ConstantSDNode>(ExtIdx))
5735 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5737 // Quit if extracted from vector of different type.
5738 if (ExtractedFromVec.getValueType() != VT)
5741 if (!VecIn1.getNode())
5742 VecIn1 = ExtractedFromVec;
5743 else if (VecIn1 != ExtractedFromVec) {
5744 if (!VecIn2.getNode())
5745 VecIn2 = ExtractedFromVec;
5746 else if (VecIn2 != ExtractedFromVec)
5747 // Quit if more than 2 vectors to shuffle
5751 if (ExtractedFromVec == VecIn1)
5753 else if (ExtractedFromVec == VecIn2)
5754 Mask[i] = Idx + NumElems;
5757 if (!VecIn1.getNode())
5760 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5761 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5762 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5763 unsigned Idx = InsertIndices[i];
5764 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5765 DAG.getIntPtrConstant(Idx, DL));
5771 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5772 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5773 Op.getScalarValueSizeInBits() == 1 &&
5774 "Can not convert non-constant vector");
5775 uint64_t Immediate = 0;
5776 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5777 SDValue In = Op.getOperand(idx);
5778 if (In.getOpcode() != ISD::UNDEF)
5779 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5783 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5784 return DAG.getConstant(Immediate, dl, VT);
5786 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5788 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5790 MVT VT = Op.getSimpleValueType();
5791 assert((VT.getVectorElementType() == MVT::i1) &&
5792 "Unexpected type in LowerBUILD_VECTORvXi1!");
5795 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5796 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5797 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5798 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5801 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5802 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5803 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5804 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5807 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5808 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5809 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5810 return DAG.getBitcast(VT, Imm);
5811 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5812 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5813 DAG.getIntPtrConstant(0, dl));
5816 // Vector has one or more non-const elements
5817 uint64_t Immediate = 0;
5818 SmallVector<unsigned, 16> NonConstIdx;
5819 bool IsSplat = true;
5820 bool HasConstElts = false;
5822 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5823 SDValue In = Op.getOperand(idx);
5824 if (In.getOpcode() == ISD::UNDEF)
5826 if (!isa<ConstantSDNode>(In))
5827 NonConstIdx.push_back(idx);
5829 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5830 HasConstElts = true;
5834 else if (In != Op.getOperand(SplatIdx))
5838 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5840 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5841 DAG.getConstant(1, dl, VT),
5842 DAG.getConstant(0, dl, VT));
5844 // insert elements one by one
5848 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5849 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5851 else if (HasConstElts)
5852 Imm = DAG.getConstant(0, dl, VT);
5854 Imm = DAG.getUNDEF(VT);
5855 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5856 DstVec = DAG.getBitcast(VT, Imm);
5858 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5859 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5860 DAG.getIntPtrConstant(0, dl));
5863 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5864 unsigned InsertIdx = NonConstIdx[i];
5865 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5866 Op.getOperand(InsertIdx),
5867 DAG.getIntPtrConstant(InsertIdx, dl));
5872 /// \brief Return true if \p N implements a horizontal binop and return the
5873 /// operands for the horizontal binop into V0 and V1.
5875 /// This is a helper function of LowerToHorizontalOp().
5876 /// This function checks that the build_vector \p N in input implements a
5877 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5878 /// operation to match.
5879 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5880 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5881 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5884 /// This function only analyzes elements of \p N whose indices are
5885 /// in range [BaseIdx, LastIdx).
5886 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5888 unsigned BaseIdx, unsigned LastIdx,
5889 SDValue &V0, SDValue &V1) {
5890 EVT VT = N->getValueType(0);
5892 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5893 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5894 "Invalid Vector in input!");
5896 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5897 bool CanFold = true;
5898 unsigned ExpectedVExtractIdx = BaseIdx;
5899 unsigned NumElts = LastIdx - BaseIdx;
5900 V0 = DAG.getUNDEF(VT);
5901 V1 = DAG.getUNDEF(VT);
5903 // Check if N implements a horizontal binop.
5904 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5905 SDValue Op = N->getOperand(i + BaseIdx);
5908 if (Op->getOpcode() == ISD::UNDEF) {
5909 // Update the expected vector extract index.
5910 if (i * 2 == NumElts)
5911 ExpectedVExtractIdx = BaseIdx;
5912 ExpectedVExtractIdx += 2;
5916 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5921 SDValue Op0 = Op.getOperand(0);
5922 SDValue Op1 = Op.getOperand(1);
5924 // Try to match the following pattern:
5925 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5926 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5927 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5928 Op0.getOperand(0) == Op1.getOperand(0) &&
5929 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5930 isa<ConstantSDNode>(Op1.getOperand(1)));
5934 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5935 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5937 if (i * 2 < NumElts) {
5938 if (V0.getOpcode() == ISD::UNDEF) {
5939 V0 = Op0.getOperand(0);
5940 if (V0.getValueType() != VT)
5944 if (V1.getOpcode() == ISD::UNDEF) {
5945 V1 = Op0.getOperand(0);
5946 if (V1.getValueType() != VT)
5949 if (i * 2 == NumElts)
5950 ExpectedVExtractIdx = BaseIdx;
5953 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5954 if (I0 == ExpectedVExtractIdx)
5955 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5956 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5957 // Try to match the following dag sequence:
5958 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5959 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5963 ExpectedVExtractIdx += 2;
5969 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5970 /// a concat_vector.
5972 /// This is a helper function of LowerToHorizontalOp().
5973 /// This function expects two 256-bit vectors called V0 and V1.
5974 /// At first, each vector is split into two separate 128-bit vectors.
5975 /// Then, the resulting 128-bit vectors are used to implement two
5976 /// horizontal binary operations.
5978 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5980 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5981 /// the two new horizontal binop.
5982 /// When Mode is set, the first horizontal binop dag node would take as input
5983 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5984 /// horizontal binop dag node would take as input the lower 128-bit of V1
5985 /// and the upper 128-bit of V1.
5987 /// HADD V0_LO, V0_HI
5988 /// HADD V1_LO, V1_HI
5990 /// Otherwise, the first horizontal binop dag node takes as input the lower
5991 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5992 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5994 /// HADD V0_LO, V1_LO
5995 /// HADD V0_HI, V1_HI
5997 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5998 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5999 /// the upper 128-bits of the result.
6000 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6001 SDLoc DL, SelectionDAG &DAG,
6002 unsigned X86Opcode, bool Mode,
6003 bool isUndefLO, bool isUndefHI) {
6004 EVT VT = V0.getValueType();
6005 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6006 "Invalid nodes in input!");
6008 unsigned NumElts = VT.getVectorNumElements();
6009 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6010 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6011 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6012 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6013 EVT NewVT = V0_LO.getValueType();
6015 SDValue LO = DAG.getUNDEF(NewVT);
6016 SDValue HI = DAG.getUNDEF(NewVT);
6019 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6020 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6021 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6022 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6023 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6025 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6026 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6027 V1_LO->getOpcode() != ISD::UNDEF))
6028 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6030 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6031 V1_HI->getOpcode() != ISD::UNDEF))
6032 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6035 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6038 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
6040 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
6041 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6042 MVT VT = BV->getSimpleValueType(0);
6043 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
6044 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
6048 unsigned NumElts = VT.getVectorNumElements();
6049 SDValue InVec0 = DAG.getUNDEF(VT);
6050 SDValue InVec1 = DAG.getUNDEF(VT);
6052 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6053 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6055 // Odd-numbered elements in the input build vector are obtained from
6056 // adding two integer/float elements.
6057 // Even-numbered elements in the input build vector are obtained from
6058 // subtracting two integer/float elements.
6059 unsigned ExpectedOpcode = ISD::FSUB;
6060 unsigned NextExpectedOpcode = ISD::FADD;
6061 bool AddFound = false;
6062 bool SubFound = false;
6064 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6065 SDValue Op = BV->getOperand(i);
6067 // Skip 'undef' values.
6068 unsigned Opcode = Op.getOpcode();
6069 if (Opcode == ISD::UNDEF) {
6070 std::swap(ExpectedOpcode, NextExpectedOpcode);
6074 // Early exit if we found an unexpected opcode.
6075 if (Opcode != ExpectedOpcode)
6078 SDValue Op0 = Op.getOperand(0);
6079 SDValue Op1 = Op.getOperand(1);
6081 // Try to match the following pattern:
6082 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6083 // Early exit if we cannot match that sequence.
6084 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6085 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6086 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6087 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6088 Op0.getOperand(1) != Op1.getOperand(1))
6091 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6095 // We found a valid add/sub node. Update the information accordingly.
6101 // Update InVec0 and InVec1.
6102 if (InVec0.getOpcode() == ISD::UNDEF) {
6103 InVec0 = Op0.getOperand(0);
6104 if (InVec0.getSimpleValueType() != VT)
6107 if (InVec1.getOpcode() == ISD::UNDEF) {
6108 InVec1 = Op1.getOperand(0);
6109 if (InVec1.getSimpleValueType() != VT)
6113 // Make sure that operands in input to each add/sub node always
6114 // come from a same pair of vectors.
6115 if (InVec0 != Op0.getOperand(0)) {
6116 if (ExpectedOpcode == ISD::FSUB)
6119 // FADD is commutable. Try to commute the operands
6120 // and then test again.
6121 std::swap(Op0, Op1);
6122 if (InVec0 != Op0.getOperand(0))
6126 if (InVec1 != Op1.getOperand(0))
6129 // Update the pair of expected opcodes.
6130 std::swap(ExpectedOpcode, NextExpectedOpcode);
6133 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6134 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6135 InVec1.getOpcode() != ISD::UNDEF)
6136 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6141 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6142 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6143 const X86Subtarget *Subtarget,
6144 SelectionDAG &DAG) {
6145 MVT VT = BV->getSimpleValueType(0);
6146 unsigned NumElts = VT.getVectorNumElements();
6147 unsigned NumUndefsLO = 0;
6148 unsigned NumUndefsHI = 0;
6149 unsigned Half = NumElts/2;
6151 // Count the number of UNDEF operands in the build_vector in input.
6152 for (unsigned i = 0, e = Half; i != e; ++i)
6153 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6156 for (unsigned i = Half, e = NumElts; i != e; ++i)
6157 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6160 // Early exit if this is either a build_vector of all UNDEFs or all the
6161 // operands but one are UNDEF.
6162 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6166 SDValue InVec0, InVec1;
6167 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6168 // Try to match an SSE3 float HADD/HSUB.
6169 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6170 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6172 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6173 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6174 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6175 // Try to match an SSSE3 integer HADD/HSUB.
6176 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6177 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6179 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6180 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6183 if (!Subtarget->hasAVX())
6186 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6187 // Try to match an AVX horizontal add/sub of packed single/double
6188 // precision floating point values from 256-bit vectors.
6189 SDValue InVec2, InVec3;
6190 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6191 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6192 ((InVec0.getOpcode() == ISD::UNDEF ||
6193 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6194 ((InVec1.getOpcode() == ISD::UNDEF ||
6195 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6196 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6198 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6199 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6200 ((InVec0.getOpcode() == ISD::UNDEF ||
6201 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6202 ((InVec1.getOpcode() == ISD::UNDEF ||
6203 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6204 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6205 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6206 // Try to match an AVX2 horizontal add/sub of signed integers.
6207 SDValue InVec2, InVec3;
6209 bool CanFold = true;
6211 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6212 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6213 ((InVec0.getOpcode() == ISD::UNDEF ||
6214 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6215 ((InVec1.getOpcode() == ISD::UNDEF ||
6216 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6217 X86Opcode = X86ISD::HADD;
6218 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6219 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6220 ((InVec0.getOpcode() == ISD::UNDEF ||
6221 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6222 ((InVec1.getOpcode() == ISD::UNDEF ||
6223 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6224 X86Opcode = X86ISD::HSUB;
6229 // Fold this build_vector into a single horizontal add/sub.
6230 // Do this only if the target has AVX2.
6231 if (Subtarget->hasAVX2())
6232 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6234 // Do not try to expand this build_vector into a pair of horizontal
6235 // add/sub if we can emit a pair of scalar add/sub.
6236 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6239 // Convert this build_vector into a pair of horizontal binop followed by
6241 bool isUndefLO = NumUndefsLO == Half;
6242 bool isUndefHI = NumUndefsHI == Half;
6243 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6244 isUndefLO, isUndefHI);
6248 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6249 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6251 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6252 X86Opcode = X86ISD::HADD;
6253 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6254 X86Opcode = X86ISD::HSUB;
6255 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6256 X86Opcode = X86ISD::FHADD;
6257 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6258 X86Opcode = X86ISD::FHSUB;
6262 // Don't try to expand this build_vector into a pair of horizontal add/sub
6263 // if we can simply emit a pair of scalar add/sub.
6264 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6267 // Convert this build_vector into two horizontal add/sub followed by
6269 bool isUndefLO = NumUndefsLO == Half;
6270 bool isUndefHI = NumUndefsHI == Half;
6271 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6272 isUndefLO, isUndefHI);
6279 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6282 MVT VT = Op.getSimpleValueType();
6283 MVT ExtVT = VT.getVectorElementType();
6284 unsigned NumElems = Op.getNumOperands();
6286 // Generate vectors for predicate vectors.
6287 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6288 return LowerBUILD_VECTORvXi1(Op, DAG);
6290 // Vectors containing all zeros can be matched by pxor and xorps later
6291 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6292 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6293 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6294 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6297 return getZeroVector(VT, Subtarget, DAG, dl);
6300 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6301 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6302 // vpcmpeqd on 256-bit vectors.
6303 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6304 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6307 if (!VT.is512BitVector())
6308 return getOnesVector(VT, Subtarget, DAG, dl);
6311 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6312 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6314 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6315 return HorizontalOp;
6316 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6319 unsigned EVTBits = ExtVT.getSizeInBits();
6321 unsigned NumZero = 0;
6322 unsigned NumNonZero = 0;
6323 uint64_t NonZeros = 0;
6324 bool IsAllConstants = true;
6325 SmallSet<SDValue, 8> Values;
6326 for (unsigned i = 0; i < NumElems; ++i) {
6327 SDValue Elt = Op.getOperand(i);
6328 if (Elt.getOpcode() == ISD::UNDEF)
6331 if (Elt.getOpcode() != ISD::Constant &&
6332 Elt.getOpcode() != ISD::ConstantFP)
6333 IsAllConstants = false;
6334 if (X86::isZeroNode(Elt))
6337 assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
6338 NonZeros |= ((uint64_t)1 << i);
6343 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6344 if (NumNonZero == 0)
6345 return DAG.getUNDEF(VT);
6347 // Special case for single non-zero, non-undef, element.
6348 if (NumNonZero == 1) {
6349 unsigned Idx = countTrailingZeros(NonZeros);
6350 SDValue Item = Op.getOperand(Idx);
6352 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6353 // the value are obviously zero, truncate the value to i32 and do the
6354 // insertion that way. Only do this if the value is non-constant or if the
6355 // value is a constant being inserted into element 0. It is cheaper to do
6356 // a constant pool load than it is to do a movd + shuffle.
6357 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6358 (!IsAllConstants || Idx == 0)) {
6359 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6361 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6362 MVT VecVT = MVT::v4i32;
6364 // Truncate the value (which may itself be a constant) to i32, and
6365 // convert it to a vector with movd (S2V+shuffle to zero extend).
6366 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6367 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6368 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6369 Item, Idx * 2, true, Subtarget, DAG));
6373 // If we have a constant or non-constant insertion into the low element of
6374 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6375 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6376 // depending on what the source datatype is.
6379 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6381 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6382 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6383 if (VT.is512BitVector()) {
6384 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6385 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6386 Item, DAG.getIntPtrConstant(0, dl));
6388 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6389 "Expected an SSE value type!");
6390 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6391 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6392 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6395 // We can't directly insert an i8 or i16 into a vector, so zero extend
6397 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6398 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6399 if (VT.is256BitVector()) {
6400 if (Subtarget->hasAVX()) {
6401 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6402 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6404 // Without AVX, we need to extend to a 128-bit vector and then
6405 // insert into the 256-bit vector.
6406 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6407 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6408 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6411 assert(VT.is128BitVector() && "Expected an SSE value type!");
6412 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6413 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6415 return DAG.getBitcast(VT, Item);
6419 // Is it a vector logical left shift?
6420 if (NumElems == 2 && Idx == 1 &&
6421 X86::isZeroNode(Op.getOperand(0)) &&
6422 !X86::isZeroNode(Op.getOperand(1))) {
6423 unsigned NumBits = VT.getSizeInBits();
6424 return getVShift(true, VT,
6425 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6426 VT, Op.getOperand(1)),
6427 NumBits/2, DAG, *this, dl);
6430 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6433 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6434 // is a non-constant being inserted into an element other than the low one,
6435 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6436 // movd/movss) to move this into the low element, then shuffle it into
6438 if (EVTBits == 32) {
6439 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6440 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6444 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6445 if (Values.size() == 1) {
6446 if (EVTBits == 32) {
6447 // Instead of a shuffle like this:
6448 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6449 // Check if it's possible to issue this instead.
6450 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6451 unsigned Idx = countTrailingZeros(NonZeros);
6452 SDValue Item = Op.getOperand(Idx);
6453 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6454 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6459 // A vector full of immediates; various special cases are already
6460 // handled, so this is best done with a single constant-pool load.
6464 // For AVX-length vectors, see if we can use a vector load to get all of the
6465 // elements, otherwise build the individual 128-bit pieces and use
6466 // shuffles to put them in place.
6467 if (VT.is256BitVector() || VT.is512BitVector()) {
6468 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6470 // Check for a build vector of consecutive loads.
6471 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6474 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6476 // Build both the lower and upper subvector.
6477 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6478 makeArrayRef(&V[0], NumElems/2));
6479 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6480 makeArrayRef(&V[NumElems / 2], NumElems/2));
6482 // Recreate the wider vector with the lower and upper part.
6483 if (VT.is256BitVector())
6484 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6485 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6488 // Let legalizer expand 2-wide build_vectors.
6489 if (EVTBits == 64) {
6490 if (NumNonZero == 1) {
6491 // One half is zero or undef.
6492 unsigned Idx = countTrailingZeros(NonZeros);
6493 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6494 Op.getOperand(Idx));
6495 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6500 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6501 if (EVTBits == 8 && NumElems == 16)
6502 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
6503 DAG, Subtarget, *this))
6506 if (EVTBits == 16 && NumElems == 8)
6507 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
6508 DAG, Subtarget, *this))
6511 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6512 if (EVTBits == 32 && NumElems == 4)
6513 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6516 // If element VT is == 32 bits, turn it into a number of shuffles.
6517 SmallVector<SDValue, 8> V(NumElems);
6518 if (NumElems == 4 && NumZero > 0) {
6519 for (unsigned i = 0; i < 4; ++i) {
6520 bool isZero = !(NonZeros & (1ULL << i));
6522 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6524 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6527 for (unsigned i = 0; i < 2; ++i) {
6528 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6531 V[i] = V[i*2]; // Must be a zero vector.
6534 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6537 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6540 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6545 bool Reverse1 = (NonZeros & 0x3) == 2;
6546 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6550 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6551 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6553 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6556 if (Values.size() > 1 && VT.is128BitVector()) {
6557 // Check for a build vector of consecutive loads.
6558 for (unsigned i = 0; i < NumElems; ++i)
6559 V[i] = Op.getOperand(i);
6561 // Check for elements which are consecutive loads.
6562 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6565 // Check for a build vector from mostly shuffle plus few inserting.
6566 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6569 // For SSE 4.1, use insertps to put the high elements into the low element.
6570 if (Subtarget->hasSSE41()) {
6572 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6573 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6575 Result = DAG.getUNDEF(VT);
6577 for (unsigned i = 1; i < NumElems; ++i) {
6578 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6579 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6580 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6585 // Otherwise, expand into a number of unpckl*, start by extending each of
6586 // our (non-undef) elements to the full vector width with the element in the
6587 // bottom slot of the vector (which generates no code for SSE).
6588 for (unsigned i = 0; i < NumElems; ++i) {
6589 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6590 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6592 V[i] = DAG.getUNDEF(VT);
6595 // Next, we iteratively mix elements, e.g. for v4f32:
6596 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6597 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6598 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6599 unsigned EltStride = NumElems >> 1;
6600 while (EltStride != 0) {
6601 for (unsigned i = 0; i < EltStride; ++i) {
6602 // If V[i+EltStride] is undef and this is the first round of mixing,
6603 // then it is safe to just drop this shuffle: V[i] is already in the
6604 // right place, the one element (since it's the first round) being
6605 // inserted as undef can be dropped. This isn't safe for successive
6606 // rounds because they will permute elements within both vectors.
6607 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6608 EltStride == NumElems/2)
6611 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6620 // 256-bit AVX can use the vinsertf128 instruction
6621 // to create 256-bit vectors from two other 128-bit ones.
6622 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6624 MVT ResVT = Op.getSimpleValueType();
6626 assert((ResVT.is256BitVector() ||
6627 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6629 SDValue V1 = Op.getOperand(0);
6630 SDValue V2 = Op.getOperand(1);
6631 unsigned NumElems = ResVT.getVectorNumElements();
6632 if (ResVT.is256BitVector())
6633 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6635 if (Op.getNumOperands() == 4) {
6636 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6637 ResVT.getVectorNumElements()/2);
6638 SDValue V3 = Op.getOperand(2);
6639 SDValue V4 = Op.getOperand(3);
6640 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6641 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6643 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6646 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6647 const X86Subtarget *Subtarget,
6648 SelectionDAG & DAG) {
6650 MVT ResVT = Op.getSimpleValueType();
6651 unsigned NumOfOperands = Op.getNumOperands();
6653 assert(isPowerOf2_32(NumOfOperands) &&
6654 "Unexpected number of operands in CONCAT_VECTORS");
6656 SDValue Undef = DAG.getUNDEF(ResVT);
6657 if (NumOfOperands > 2) {
6658 // Specialize the cases when all, or all but one, of the operands are undef.
6659 unsigned NumOfDefinedOps = 0;
6661 for (unsigned i = 0; i < NumOfOperands; i++)
6662 if (!Op.getOperand(i).isUndef()) {
6666 if (NumOfDefinedOps == 0)
6668 if (NumOfDefinedOps == 1) {
6669 unsigned SubVecNumElts =
6670 Op.getOperand(OpIdx).getValueType().getVectorNumElements();
6671 SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
6672 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
6673 Op.getOperand(OpIdx), IdxVal);
6676 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6677 ResVT.getVectorNumElements()/2);
6678 SmallVector<SDValue, 2> Ops;
6679 for (unsigned i = 0; i < NumOfOperands/2; i++)
6680 Ops.push_back(Op.getOperand(i));
6681 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6683 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6684 Ops.push_back(Op.getOperand(i));
6685 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6686 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6690 SDValue V1 = Op.getOperand(0);
6691 SDValue V2 = Op.getOperand(1);
6692 unsigned NumElems = ResVT.getVectorNumElements();
6693 assert(V1.getValueType() == V2.getValueType() &&
6694 V1.getValueType().getVectorNumElements() == NumElems/2 &&
6695 "Unexpected operands in CONCAT_VECTORS");
6697 if (ResVT.getSizeInBits() >= 16)
6698 return Op; // The operation is legal with KUNPCK
6700 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6701 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6702 SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
6703 if (IsZeroV1 && IsZeroV2)
6706 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6708 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6710 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
6712 SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
6714 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
6717 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
6719 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6720 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
6723 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6724 const X86Subtarget *Subtarget,
6725 SelectionDAG &DAG) {
6726 MVT VT = Op.getSimpleValueType();
6727 if (VT.getVectorElementType() == MVT::i1)
6728 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6730 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6731 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6732 Op.getNumOperands() == 4)));
6734 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6735 // from two other 128-bit ones.
6737 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6738 return LowerAVXCONCAT_VECTORS(Op, DAG);
6741 //===----------------------------------------------------------------------===//
6742 // Vector shuffle lowering
6744 // This is an experimental code path for lowering vector shuffles on x86. It is
6745 // designed to handle arbitrary vector shuffles and blends, gracefully
6746 // degrading performance as necessary. It works hard to recognize idiomatic
6747 // shuffles and lower them to optimal instruction patterns without leaving
6748 // a framework that allows reasonably efficient handling of all vector shuffle
6750 //===----------------------------------------------------------------------===//
6752 /// \brief Tiny helper function to identify a no-op mask.
6754 /// This is a somewhat boring predicate function. It checks whether the mask
6755 /// array input, which is assumed to be a single-input shuffle mask of the kind
6756 /// used by the X86 shuffle instructions (not a fully general
6757 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6758 /// in-place shuffle are 'no-op's.
6759 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6760 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6761 if (Mask[i] != -1 && Mask[i] != i)
6766 /// \brief Helper function to classify a mask as a single-input mask.
6768 /// This isn't a generic single-input test because in the vector shuffle
6769 /// lowering we canonicalize single inputs to be the first input operand. This
6770 /// means we can more quickly test for a single input by only checking whether
6771 /// an input from the second operand exists. We also assume that the size of
6772 /// mask corresponds to the size of the input vectors which isn't true in the
6773 /// fully general case.
6774 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6776 if (M >= (int)Mask.size())
6781 /// \brief Test whether there are elements crossing 128-bit lanes in this
6784 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6785 /// and we routinely test for these.
6786 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6787 int LaneSize = 128 / VT.getScalarSizeInBits();
6788 int Size = Mask.size();
6789 for (int i = 0; i < Size; ++i)
6790 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6795 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6797 /// This checks a shuffle mask to see if it is performing the same
6798 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6799 /// that it is also not lane-crossing. It may however involve a blend from the
6800 /// same lane of a second vector.
6802 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6803 /// non-trivial to compute in the face of undef lanes. The representation is
6804 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6805 /// entries from both V1 and V2 inputs to the wider mask.
6807 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6808 SmallVectorImpl<int> &RepeatedMask) {
6809 int LaneSize = 128 / VT.getScalarSizeInBits();
6810 RepeatedMask.resize(LaneSize, -1);
6811 int Size = Mask.size();
6812 for (int i = 0; i < Size; ++i) {
6815 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6816 // This entry crosses lanes, so there is no way to model this shuffle.
6819 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6820 if (RepeatedMask[i % LaneSize] == -1)
6821 // This is the first non-undef entry in this slot of a 128-bit lane.
6822 RepeatedMask[i % LaneSize] =
6823 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6824 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6825 // Found a mismatch with the repeated mask.
6831 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6834 /// This is a fast way to test a shuffle mask against a fixed pattern:
6836 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6838 /// It returns true if the mask is exactly as wide as the argument list, and
6839 /// each element of the mask is either -1 (signifying undef) or the value given
6840 /// in the argument.
6841 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6842 ArrayRef<int> ExpectedMask) {
6843 if (Mask.size() != ExpectedMask.size())
6846 int Size = Mask.size();
6848 // If the values are build vectors, we can look through them to find
6849 // equivalent inputs that make the shuffles equivalent.
6850 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6851 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6853 for (int i = 0; i < Size; ++i)
6854 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6855 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6856 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6857 if (!MaskBV || !ExpectedBV ||
6858 MaskBV->getOperand(Mask[i] % Size) !=
6859 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6866 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6868 /// This helper function produces an 8-bit shuffle immediate corresponding to
6869 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6870 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6873 /// NB: We rely heavily on "undef" masks preserving the input lane.
6874 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6875 SelectionDAG &DAG) {
6876 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6877 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6878 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6879 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6880 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6883 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6884 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6885 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6886 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6887 return DAG.getConstant(Imm, DL, MVT::i8);
6890 /// \brief Compute whether each element of a shuffle is zeroable.
6892 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6893 /// Either it is an undef element in the shuffle mask, the element of the input
6894 /// referenced is undef, or the element of the input referenced is known to be
6895 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6896 /// as many lanes with this technique as possible to simplify the remaining
6898 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6899 SDValue V1, SDValue V2) {
6900 SmallBitVector Zeroable(Mask.size(), false);
6902 while (V1.getOpcode() == ISD::BITCAST)
6903 V1 = V1->getOperand(0);
6904 while (V2.getOpcode() == ISD::BITCAST)
6905 V2 = V2->getOperand(0);
6907 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6908 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6910 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6912 // Handle the easy cases.
6913 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6918 // If this is an index into a build_vector node (which has the same number
6919 // of elements), dig out the input value and use it.
6920 SDValue V = M < Size ? V1 : V2;
6921 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6924 SDValue Input = V.getOperand(M % Size);
6925 // The UNDEF opcode check really should be dead code here, but not quite
6926 // worth asserting on (it isn't invalid, just unexpected).
6927 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6934 // X86 has dedicated unpack instructions that can handle specific blend
6935 // operations: UNPCKH and UNPCKL.
6936 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6937 SDValue V1, SDValue V2,
6938 SelectionDAG &DAG) {
6939 int NumElts = VT.getVectorNumElements();
6940 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6941 SmallVector<int, 8> Unpckl;
6942 SmallVector<int, 8> Unpckh;
6944 for (int i = 0; i < NumElts; ++i) {
6945 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6946 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6947 int HiPos = LoPos + NumEltsInLane / 2;
6948 Unpckl.push_back(LoPos);
6949 Unpckh.push_back(HiPos);
6952 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6953 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6954 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6955 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6957 // Commute and try again.
6958 ShuffleVectorSDNode::commuteMask(Unpckl);
6959 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6960 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6962 ShuffleVectorSDNode::commuteMask(Unpckh);
6963 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6964 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6969 /// \brief Try to emit a bitmask instruction for a shuffle.
6971 /// This handles cases where we can model a blend exactly as a bitmask due to
6972 /// one of the inputs being zeroable.
6973 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6974 SDValue V2, ArrayRef<int> Mask,
6975 SelectionDAG &DAG) {
6976 MVT EltVT = VT.getVectorElementType();
6977 int NumEltBits = EltVT.getSizeInBits();
6978 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6979 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6980 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6982 if (EltVT.isFloatingPoint()) {
6983 Zero = DAG.getBitcast(EltVT, Zero);
6984 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6986 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6987 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6989 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6992 if (Mask[i] % Size != i)
6993 return SDValue(); // Not a blend.
6995 V = Mask[i] < Size ? V1 : V2;
6996 else if (V != (Mask[i] < Size ? V1 : V2))
6997 return SDValue(); // Can only let one input through the mask.
6999 VMaskOps[i] = AllOnes;
7002 return SDValue(); // No non-zeroable elements!
7004 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
7005 V = DAG.getNode(VT.isFloatingPoint()
7006 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
7011 /// \brief Try to emit a blend instruction for a shuffle using bit math.
7013 /// This is used as a fallback approach when first class blend instructions are
7014 /// unavailable. Currently it is only suitable for integer vectors, but could
7015 /// be generalized for floating point vectors if desirable.
7016 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
7017 SDValue V2, ArrayRef<int> Mask,
7018 SelectionDAG &DAG) {
7019 assert(VT.isInteger() && "Only supports integer vector types!");
7020 MVT EltVT = VT.getVectorElementType();
7021 int NumEltBits = EltVT.getSizeInBits();
7022 SDValue Zero = DAG.getConstant(0, DL, EltVT);
7023 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
7025 SmallVector<SDValue, 16> MaskOps;
7026 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7027 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
7028 return SDValue(); // Shuffled input!
7029 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
7032 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
7033 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
7034 // We have to cast V2 around.
7035 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
7036 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
7037 DAG.getBitcast(MaskVT, V1Mask),
7038 DAG.getBitcast(MaskVT, V2)));
7039 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
7042 /// \brief Try to emit a blend instruction for a shuffle.
7044 /// This doesn't do any checks for the availability of instructions for blending
7045 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7046 /// be matched in the backend with the type given. What it does check for is
7047 /// that the shuffle mask is a blend, or convertible into a blend with zero.
7048 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7049 SDValue V2, ArrayRef<int> Original,
7050 const X86Subtarget *Subtarget,
7051 SelectionDAG &DAG) {
7052 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7053 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7054 SmallVector<int, 8> Mask(Original.begin(), Original.end());
7055 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7056 bool ForceV1Zero = false, ForceV2Zero = false;
7058 // Attempt to generate the binary blend mask. If an input is zero then
7059 // we can use any lane.
7060 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
7061 unsigned BlendMask = 0;
7062 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7068 if (M == i + Size) {
7069 BlendMask |= 1u << i;
7080 BlendMask |= 1u << i;
7085 return SDValue(); // Shuffled input!
7088 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
7090 V1 = getZeroVector(VT, Subtarget, DAG, DL);
7092 V2 = getZeroVector(VT, Subtarget, DAG, DL);
7094 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
7095 unsigned ScaledMask = 0;
7096 for (int i = 0; i != Size; ++i)
7097 if (BlendMask & (1u << i))
7098 for (int j = 0; j != Scale; ++j)
7099 ScaledMask |= 1u << (i * Scale + j);
7103 switch (VT.SimpleTy) {
7108 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7109 DAG.getConstant(BlendMask, DL, MVT::i8));
7113 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7117 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7118 // that instruction.
7119 if (Subtarget->hasAVX2()) {
7120 // Scale the blend by the number of 32-bit dwords per element.
7121 int Scale = VT.getScalarSizeInBits() / 32;
7122 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7123 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7124 V1 = DAG.getBitcast(BlendVT, V1);
7125 V2 = DAG.getBitcast(BlendVT, V2);
7126 return DAG.getBitcast(
7127 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7128 DAG.getConstant(BlendMask, DL, MVT::i8)));
7132 // For integer shuffles we need to expand the mask and cast the inputs to
7133 // v8i16s prior to blending.
7134 int Scale = 8 / VT.getVectorNumElements();
7135 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7136 V1 = DAG.getBitcast(MVT::v8i16, V1);
7137 V2 = DAG.getBitcast(MVT::v8i16, V2);
7138 return DAG.getBitcast(VT,
7139 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7140 DAG.getConstant(BlendMask, DL, MVT::i8)));
7144 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7145 SmallVector<int, 8> RepeatedMask;
7146 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7147 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7148 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7150 for (int i = 0; i < 8; ++i)
7151 if (RepeatedMask[i] >= 16)
7152 BlendMask |= 1u << i;
7153 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7154 DAG.getConstant(BlendMask, DL, MVT::i8));
7160 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
7161 "256-bit byte-blends require AVX2 support!");
7163 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
7164 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
7167 // Scale the blend by the number of bytes per element.
7168 int Scale = VT.getScalarSizeInBits() / 8;
7170 // This form of blend is always done on bytes. Compute the byte vector
7172 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7174 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7175 // mix of LLVM's code generator and the x86 backend. We tell the code
7176 // generator that boolean values in the elements of an x86 vector register
7177 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7178 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7179 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7180 // of the element (the remaining are ignored) and 0 in that high bit would
7181 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7182 // the LLVM model for boolean values in vector elements gets the relevant
7183 // bit set, it is set backwards and over constrained relative to x86's
7185 SmallVector<SDValue, 32> VSELECTMask;
7186 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7187 for (int j = 0; j < Scale; ++j)
7188 VSELECTMask.push_back(
7189 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7190 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7193 V1 = DAG.getBitcast(BlendVT, V1);
7194 V2 = DAG.getBitcast(BlendVT, V2);
7195 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7196 DAG.getNode(ISD::BUILD_VECTOR, DL,
7197 BlendVT, VSELECTMask),
7202 llvm_unreachable("Not a supported integer vector type!");
7206 /// \brief Try to lower as a blend of elements from two inputs followed by
7207 /// a single-input permutation.
7209 /// This matches the pattern where we can blend elements from two inputs and
7210 /// then reduce the shuffle to a single-input permutation.
7211 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7214 SelectionDAG &DAG) {
7215 // We build up the blend mask while checking whether a blend is a viable way
7216 // to reduce the shuffle.
7217 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7218 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7220 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7224 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7226 if (BlendMask[Mask[i] % Size] == -1)
7227 BlendMask[Mask[i] % Size] = Mask[i];
7228 else if (BlendMask[Mask[i] % Size] != Mask[i])
7229 return SDValue(); // Can't blend in the needed input!
7231 PermuteMask[i] = Mask[i] % Size;
7234 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7235 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7238 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7239 /// blends and permutes.
7241 /// This matches the extremely common pattern for handling combined
7242 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7243 /// operations. It will try to pick the best arrangement of shuffles and
7245 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7249 SelectionDAG &DAG) {
7250 // Shuffle the input elements into the desired positions in V1 and V2 and
7251 // blend them together.
7252 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7253 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7254 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7255 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7256 if (Mask[i] >= 0 && Mask[i] < Size) {
7257 V1Mask[i] = Mask[i];
7259 } else if (Mask[i] >= Size) {
7260 V2Mask[i] = Mask[i] - Size;
7261 BlendMask[i] = i + Size;
7264 // Try to lower with the simpler initial blend strategy unless one of the
7265 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7266 // shuffle may be able to fold with a load or other benefit. However, when
7267 // we'll have to do 2x as many shuffles in order to achieve this, blending
7268 // first is a better strategy.
7269 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7270 if (SDValue BlendPerm =
7271 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7274 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7275 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7276 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7279 /// \brief Try to lower a vector shuffle as a byte rotation.
7281 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7282 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7283 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7284 /// try to generically lower a vector shuffle through such an pattern. It
7285 /// does not check for the profitability of lowering either as PALIGNR or
7286 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7287 /// This matches shuffle vectors that look like:
7289 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7291 /// Essentially it concatenates V1 and V2, shifts right by some number of
7292 /// elements, and takes the low elements as the result. Note that while this is
7293 /// specified as a *right shift* because x86 is little-endian, it is a *left
7294 /// rotate* of the vector lanes.
7295 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7298 const X86Subtarget *Subtarget,
7299 SelectionDAG &DAG) {
7300 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7302 int NumElts = Mask.size();
7303 int NumLanes = VT.getSizeInBits() / 128;
7304 int NumLaneElts = NumElts / NumLanes;
7306 // We need to detect various ways of spelling a rotation:
7307 // [11, 12, 13, 14, 15, 0, 1, 2]
7308 // [-1, 12, 13, 14, -1, -1, 1, -1]
7309 // [-1, -1, -1, -1, -1, -1, 1, 2]
7310 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7311 // [-1, 4, 5, 6, -1, -1, 9, -1]
7312 // [-1, 4, 5, 6, -1, -1, -1, -1]
7315 for (int l = 0; l < NumElts; l += NumLaneElts) {
7316 for (int i = 0; i < NumLaneElts; ++i) {
7317 if (Mask[l + i] == -1)
7319 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7321 // Get the mod-Size index and lane correct it.
7322 int LaneIdx = (Mask[l + i] % NumElts) - l;
7323 // Make sure it was in this lane.
7324 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7327 // Determine where a rotated vector would have started.
7328 int StartIdx = i - LaneIdx;
7330 // The identity rotation isn't interesting, stop.
7333 // If we found the tail of a vector the rotation must be the missing
7334 // front. If we found the head of a vector, it must be how much of the
7336 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7339 Rotation = CandidateRotation;
7340 else if (Rotation != CandidateRotation)
7341 // The rotations don't match, so we can't match this mask.
7344 // Compute which value this mask is pointing at.
7345 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7347 // Compute which of the two target values this index should be assigned
7348 // to. This reflects whether the high elements are remaining or the low
7349 // elements are remaining.
7350 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7352 // Either set up this value if we've not encountered it before, or check
7353 // that it remains consistent.
7356 else if (TargetV != MaskV)
7357 // This may be a rotation, but it pulls from the inputs in some
7358 // unsupported interleaving.
7363 // Check that we successfully analyzed the mask, and normalize the results.
7364 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7365 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7371 // The actual rotate instruction rotates bytes, so we need to scale the
7372 // rotation based on how many bytes are in the vector lane.
7373 int Scale = 16 / NumLaneElts;
7375 // SSSE3 targets can use the palignr instruction.
7376 if (Subtarget->hasSSSE3()) {
7377 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7378 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7379 Lo = DAG.getBitcast(AlignVT, Lo);
7380 Hi = DAG.getBitcast(AlignVT, Hi);
7382 return DAG.getBitcast(
7383 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7384 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7387 assert(VT.is128BitVector() &&
7388 "Rotate-based lowering only supports 128-bit lowering!");
7389 assert(Mask.size() <= 16 &&
7390 "Can shuffle at most 16 bytes in a 128-bit vector!");
7392 // Default SSE2 implementation
7393 int LoByteShift = 16 - Rotation * Scale;
7394 int HiByteShift = Rotation * Scale;
7396 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7397 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7398 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7400 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7401 DAG.getConstant(LoByteShift, DL, MVT::i8));
7402 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7403 DAG.getConstant(HiByteShift, DL, MVT::i8));
7404 return DAG.getBitcast(VT,
7405 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7408 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7410 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7411 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7412 /// matches elements from one of the input vectors shuffled to the left or
7413 /// right with zeroable elements 'shifted in'. It handles both the strictly
7414 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7417 /// PSHL : (little-endian) left bit shift.
7418 /// [ zz, 0, zz, 2 ]
7419 /// [ -1, 4, zz, -1 ]
7420 /// PSRL : (little-endian) right bit shift.
7422 /// [ -1, -1, 7, zz]
7423 /// PSLLDQ : (little-endian) left byte shift
7424 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7425 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7426 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7427 /// PSRLDQ : (little-endian) right byte shift
7428 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7429 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7430 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7431 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7432 SDValue V2, ArrayRef<int> Mask,
7433 SelectionDAG &DAG) {
7434 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7436 int Size = Mask.size();
7437 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7439 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7440 for (int i = 0; i < Size; i += Scale)
7441 for (int j = 0; j < Shift; ++j)
7442 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7448 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7449 for (int i = 0; i != Size; i += Scale) {
7450 unsigned Pos = Left ? i + Shift : i;
7451 unsigned Low = Left ? i : i + Shift;
7452 unsigned Len = Scale - Shift;
7453 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7454 Low + (V == V1 ? 0 : Size)))
7458 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7459 bool ByteShift = ShiftEltBits > 64;
7460 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7461 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7462 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7464 // Normalize the scale for byte shifts to still produce an i64 element
7466 Scale = ByteShift ? Scale / 2 : Scale;
7468 // We need to round trip through the appropriate type for the shift.
7469 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7470 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7471 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7472 "Illegal integer vector type");
7473 V = DAG.getBitcast(ShiftVT, V);
7475 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7476 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7477 return DAG.getBitcast(VT, V);
7480 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7481 // keep doubling the size of the integer elements up to that. We can
7482 // then shift the elements of the integer vector by whole multiples of
7483 // their width within the elements of the larger integer vector. Test each
7484 // multiple to see if we can find a match with the moved element indices
7485 // and that the shifted in elements are all zeroable.
7486 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7487 for (int Shift = 1; Shift != Scale; ++Shift)
7488 for (bool Left : {true, false})
7489 if (CheckZeros(Shift, Scale, Left))
7490 for (SDValue V : {V1, V2})
7491 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7498 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7499 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7500 SDValue V2, ArrayRef<int> Mask,
7501 SelectionDAG &DAG) {
7502 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7503 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7505 int Size = Mask.size();
7506 int HalfSize = Size / 2;
7507 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7509 // Upper half must be undefined.
7510 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7513 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7514 // Remainder of lower half result is zero and upper half is all undef.
7515 auto LowerAsEXTRQ = [&]() {
7516 // Determine the extraction length from the part of the
7517 // lower half that isn't zeroable.
7519 for (; Len > 0; --Len)
7520 if (!Zeroable[Len - 1])
7522 assert(Len > 0 && "Zeroable shuffle mask");
7524 // Attempt to match first Len sequential elements from the lower half.
7527 for (int i = 0; i != Len; ++i) {
7531 SDValue &V = (M < Size ? V1 : V2);
7534 // The extracted elements must start at a valid index and all mask
7535 // elements must be in the lower half.
7536 if (i > M || M >= HalfSize)
7539 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7550 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7551 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7552 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7553 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7554 DAG.getConstant(BitLen, DL, MVT::i8),
7555 DAG.getConstant(BitIdx, DL, MVT::i8));
7558 if (SDValue ExtrQ = LowerAsEXTRQ())
7561 // INSERTQ: Extract lowest Len elements from lower half of second source and
7562 // insert over first source, starting at Idx.
7563 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7564 auto LowerAsInsertQ = [&]() {
7565 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7568 // Attempt to match first source from mask before insertion point.
7569 if (isUndefInRange(Mask, 0, Idx)) {
7571 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7573 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7579 // Extend the extraction length looking to match both the insertion of
7580 // the second source and the remaining elements of the first.
7581 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7586 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7588 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7594 // Match the remaining elements of the lower half.
7595 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7597 } else if ((!Base || (Base == V1)) &&
7598 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7600 } else if ((!Base || (Base == V2)) &&
7601 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7608 // We may not have a base (first source) - this can safely be undefined.
7610 Base = DAG.getUNDEF(VT);
7612 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7613 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7614 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7615 DAG.getConstant(BitLen, DL, MVT::i8),
7616 DAG.getConstant(BitIdx, DL, MVT::i8));
7623 if (SDValue InsertQ = LowerAsInsertQ())
7629 /// \brief Lower a vector shuffle as a zero or any extension.
7631 /// Given a specific number of elements, element bit width, and extension
7632 /// stride, produce either a zero or any extension based on the available
7633 /// features of the subtarget. The extended elements are consecutive and
7634 /// begin and can start from an offseted element index in the input; to
7635 /// avoid excess shuffling the offset must either being in the bottom lane
7636 /// or at the start of a higher lane. All extended elements must be from
7638 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7639 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7640 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7641 assert(Scale > 1 && "Need a scale to extend.");
7642 int EltBits = VT.getScalarSizeInBits();
7643 int NumElements = VT.getVectorNumElements();
7644 int NumEltsPerLane = 128 / EltBits;
7645 int OffsetLane = Offset / NumEltsPerLane;
7646 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7647 "Only 8, 16, and 32 bit elements can be extended.");
7648 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7649 assert(0 <= Offset && "Extension offset must be positive.");
7650 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7651 "Extension offset must be in the first lane or start an upper lane.");
7653 // Check that an index is in same lane as the base offset.
7654 auto SafeOffset = [&](int Idx) {
7655 return OffsetLane == (Idx / NumEltsPerLane);
7658 // Shift along an input so that the offset base moves to the first element.
7659 auto ShuffleOffset = [&](SDValue V) {
7663 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7664 for (int i = 0; i * Scale < NumElements; ++i) {
7665 int SrcIdx = i + Offset;
7666 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7668 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7671 // Found a valid zext mask! Try various lowering strategies based on the
7672 // input type and available ISA extensions.
7673 if (Subtarget->hasSSE41()) {
7674 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7675 // PUNPCK will catch this in a later shuffle match.
7676 if (Offset && Scale == 2 && VT.is128BitVector())
7678 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7679 NumElements / Scale);
7680 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7681 return DAG.getBitcast(VT, InputV);
7684 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7686 // For any extends we can cheat for larger element sizes and use shuffle
7687 // instructions that can fold with a load and/or copy.
7688 if (AnyExt && EltBits == 32) {
7689 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7691 return DAG.getBitcast(
7692 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7693 DAG.getBitcast(MVT::v4i32, InputV),
7694 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7696 if (AnyExt && EltBits == 16 && Scale > 2) {
7697 int PSHUFDMask[4] = {Offset / 2, -1,
7698 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7699 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7700 DAG.getBitcast(MVT::v4i32, InputV),
7701 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7702 int PSHUFWMask[4] = {1, -1, -1, -1};
7703 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7704 return DAG.getBitcast(
7705 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7706 DAG.getBitcast(MVT::v8i16, InputV),
7707 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7710 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7712 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7713 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7714 assert(VT.is128BitVector() && "Unexpected vector width!");
7716 int LoIdx = Offset * EltBits;
7717 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7718 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7719 DAG.getConstant(EltBits, DL, MVT::i8),
7720 DAG.getConstant(LoIdx, DL, MVT::i8)));
7722 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7723 !SafeOffset(Offset + 1))
7724 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7726 int HiIdx = (Offset + 1) * EltBits;
7727 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7728 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7729 DAG.getConstant(EltBits, DL, MVT::i8),
7730 DAG.getConstant(HiIdx, DL, MVT::i8)));
7731 return DAG.getNode(ISD::BITCAST, DL, VT,
7732 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7735 // If this would require more than 2 unpack instructions to expand, use
7736 // pshufb when available. We can only use more than 2 unpack instructions
7737 // when zero extending i8 elements which also makes it easier to use pshufb.
7738 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7739 assert(NumElements == 16 && "Unexpected byte vector width!");
7740 SDValue PSHUFBMask[16];
7741 for (int i = 0; i < 16; ++i) {
7742 int Idx = Offset + (i / Scale);
7743 PSHUFBMask[i] = DAG.getConstant(
7744 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7746 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7747 return DAG.getBitcast(VT,
7748 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7749 DAG.getNode(ISD::BUILD_VECTOR, DL,
7750 MVT::v16i8, PSHUFBMask)));
7753 // If we are extending from an offset, ensure we start on a boundary that
7754 // we can unpack from.
7755 int AlignToUnpack = Offset % (NumElements / Scale);
7756 if (AlignToUnpack) {
7757 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7758 for (int i = AlignToUnpack; i < NumElements; ++i)
7759 ShMask[i - AlignToUnpack] = i;
7760 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7761 Offset -= AlignToUnpack;
7764 // Otherwise emit a sequence of unpacks.
7766 unsigned UnpackLoHi = X86ISD::UNPCKL;
7767 if (Offset >= (NumElements / 2)) {
7768 UnpackLoHi = X86ISD::UNPCKH;
7769 Offset -= (NumElements / 2);
7772 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7773 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7774 : getZeroVector(InputVT, Subtarget, DAG, DL);
7775 InputV = DAG.getBitcast(InputVT, InputV);
7776 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7780 } while (Scale > 1);
7781 return DAG.getBitcast(VT, InputV);
7784 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7786 /// This routine will try to do everything in its power to cleverly lower
7787 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7788 /// check for the profitability of this lowering, it tries to aggressively
7789 /// match this pattern. It will use all of the micro-architectural details it
7790 /// can to emit an efficient lowering. It handles both blends with all-zero
7791 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7792 /// masking out later).
7794 /// The reason we have dedicated lowering for zext-style shuffles is that they
7795 /// are both incredibly common and often quite performance sensitive.
7796 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7797 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7798 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7799 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7801 int Bits = VT.getSizeInBits();
7802 int NumLanes = Bits / 128;
7803 int NumElements = VT.getVectorNumElements();
7804 int NumEltsPerLane = NumElements / NumLanes;
7805 assert(VT.getScalarSizeInBits() <= 32 &&
7806 "Exceeds 32-bit integer zero extension limit");
7807 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7809 // Define a helper function to check a particular ext-scale and lower to it if
7811 auto Lower = [&](int Scale) -> SDValue {
7816 for (int i = 0; i < NumElements; ++i) {
7819 continue; // Valid anywhere but doesn't tell us anything.
7820 if (i % Scale != 0) {
7821 // Each of the extended elements need to be zeroable.
7825 // We no longer are in the anyext case.
7830 // Each of the base elements needs to be consecutive indices into the
7831 // same input vector.
7832 SDValue V = M < NumElements ? V1 : V2;
7833 M = M % NumElements;
7836 Offset = M - (i / Scale);
7837 } else if (InputV != V)
7838 return SDValue(); // Flip-flopping inputs.
7840 // Offset must start in the lowest 128-bit lane or at the start of an
7842 // FIXME: Is it ever worth allowing a negative base offset?
7843 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7844 (Offset % NumEltsPerLane) == 0))
7847 // If we are offsetting, all referenced entries must come from the same
7849 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7852 if ((M % NumElements) != (Offset + (i / Scale)))
7853 return SDValue(); // Non-consecutive strided elements.
7857 // If we fail to find an input, we have a zero-shuffle which should always
7858 // have already been handled.
7859 // FIXME: Maybe handle this here in case during blending we end up with one?
7863 // If we are offsetting, don't extend if we only match a single input, we
7864 // can always do better by using a basic PSHUF or PUNPCK.
7865 if (Offset != 0 && Matches < 2)
7868 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7869 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7872 // The widest scale possible for extending is to a 64-bit integer.
7873 assert(Bits % 64 == 0 &&
7874 "The number of bits in a vector must be divisible by 64 on x86!");
7875 int NumExtElements = Bits / 64;
7877 // Each iteration, try extending the elements half as much, but into twice as
7879 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7880 assert(NumElements % NumExtElements == 0 &&
7881 "The input vector size must be divisible by the extended size.");
7882 if (SDValue V = Lower(NumElements / NumExtElements))
7886 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7890 // Returns one of the source operands if the shuffle can be reduced to a
7891 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7892 auto CanZExtLowHalf = [&]() {
7893 for (int i = NumElements / 2; i != NumElements; ++i)
7896 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7898 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7903 if (SDValue V = CanZExtLowHalf()) {
7904 V = DAG.getBitcast(MVT::v2i64, V);
7905 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7906 return DAG.getBitcast(VT, V);
7909 // No viable ext lowering found.
7913 /// \brief Try to get a scalar value for a specific element of a vector.
7915 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7916 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7917 SelectionDAG &DAG) {
7918 MVT VT = V.getSimpleValueType();
7919 MVT EltVT = VT.getVectorElementType();
7920 while (V.getOpcode() == ISD::BITCAST)
7921 V = V.getOperand(0);
7922 // If the bitcasts shift the element size, we can't extract an equivalent
7924 MVT NewVT = V.getSimpleValueType();
7925 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7928 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7929 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7930 // Ensure the scalar operand is the same size as the destination.
7931 // FIXME: Add support for scalar truncation where possible.
7932 SDValue S = V.getOperand(Idx);
7933 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7934 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7940 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7942 /// This is particularly important because the set of instructions varies
7943 /// significantly based on whether the operand is a load or not.
7944 static bool isShuffleFoldableLoad(SDValue V) {
7945 while (V.getOpcode() == ISD::BITCAST)
7946 V = V.getOperand(0);
7948 return ISD::isNON_EXTLoad(V.getNode());
7951 /// \brief Try to lower insertion of a single element into a zero vector.
7953 /// This is a common pattern that we have especially efficient patterns to lower
7954 /// across all subtarget feature sets.
7955 static SDValue lowerVectorShuffleAsElementInsertion(
7956 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7957 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7958 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7960 MVT EltVT = VT.getVectorElementType();
7962 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7963 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7965 bool IsV1Zeroable = true;
7966 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7967 if (i != V2Index && !Zeroable[i]) {
7968 IsV1Zeroable = false;
7972 // Check for a single input from a SCALAR_TO_VECTOR node.
7973 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7974 // all the smarts here sunk into that routine. However, the current
7975 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7976 // vector shuffle lowering is dead.
7977 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7979 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7980 // We need to zext the scalar if it is smaller than an i32.
7981 V2S = DAG.getBitcast(EltVT, V2S);
7982 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7983 // Using zext to expand a narrow element won't work for non-zero
7988 // Zero-extend directly to i32.
7990 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7992 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7993 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7994 EltVT == MVT::i16) {
7995 // Either not inserting from the low element of the input or the input
7996 // element size is too small to use VZEXT_MOVL to clear the high bits.
8000 if (!IsV1Zeroable) {
8001 // If V1 can't be treated as a zero vector we have fewer options to lower
8002 // this. We can't support integer vectors or non-zero targets cheaply, and
8003 // the V1 elements can't be permuted in any way.
8004 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8005 if (!VT.isFloatingPoint() || V2Index != 0)
8007 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8008 V1Mask[V2Index] = -1;
8009 if (!isNoopShuffleMask(V1Mask))
8011 // This is essentially a special case blend operation, but if we have
8012 // general purpose blend operations, they are always faster. Bail and let
8013 // the rest of the lowering handle these as blends.
8014 if (Subtarget->hasSSE41())
8017 // Otherwise, use MOVSD or MOVSS.
8018 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8019 "Only two types of floating point element types to handle!");
8020 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8024 // This lowering only works for the low element with floating point vectors.
8025 if (VT.isFloatingPoint() && V2Index != 0)
8028 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8030 V2 = DAG.getBitcast(VT, V2);
8033 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8034 // the desired position. Otherwise it is more efficient to do a vector
8035 // shift left. We know that we can do a vector shift left because all
8036 // the inputs are zero.
8037 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8038 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8039 V2Shuffle[V2Index] = 0;
8040 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8042 V2 = DAG.getBitcast(MVT::v2i64, V2);
8044 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8045 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
8046 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
8047 DAG.getDataLayout(), VT)));
8048 V2 = DAG.getBitcast(VT, V2);
8054 /// \brief Try to lower broadcast of a single - truncated - integer element,
8055 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
8057 /// This assumes we have AVX2.
8058 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
8060 const X86Subtarget *Subtarget,
8061 SelectionDAG &DAG) {
8062 assert(Subtarget->hasAVX2() &&
8063 "We can only lower integer broadcasts with AVX2!");
8065 EVT EltVT = VT.getVectorElementType();
8066 EVT V0VT = V0.getValueType();
8068 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
8069 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
8071 EVT V0EltVT = V0VT.getVectorElementType();
8072 if (!V0EltVT.isInteger())
8075 const unsigned EltSize = EltVT.getSizeInBits();
8076 const unsigned V0EltSize = V0EltVT.getSizeInBits();
8078 // This is only a truncation if the original element type is larger.
8079 if (V0EltSize <= EltSize)
8082 assert(((V0EltSize % EltSize) == 0) &&
8083 "Scalar type sizes must all be powers of 2 on x86!");
8085 const unsigned V0Opc = V0.getOpcode();
8086 const unsigned Scale = V0EltSize / EltSize;
8087 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
8089 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
8090 V0Opc != ISD::BUILD_VECTOR)
8093 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
8095 // If we're extracting non-least-significant bits, shift so we can truncate.
8096 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
8097 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
8098 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
8099 if (const int OffsetIdx = BroadcastIdx % Scale)
8100 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
8101 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
8103 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
8104 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
8107 /// \brief Try to lower broadcast of a single element.
8109 /// For convenience, this code also bundles all of the subtarget feature set
8110 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8111 /// a convenient way to factor it out.
8112 /// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
8113 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
8115 const X86Subtarget *Subtarget,
8116 SelectionDAG &DAG) {
8117 if (!Subtarget->hasAVX())
8119 if (VT.isInteger() && !Subtarget->hasAVX2())
8122 // Check that the mask is a broadcast.
8123 int BroadcastIdx = -1;
8125 if (M >= 0 && BroadcastIdx == -1)
8127 else if (M >= 0 && M != BroadcastIdx)
8130 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8131 "a sorted mask where the broadcast "
8134 // Go up the chain of (vector) values to find a scalar load that we can
8135 // combine with the broadcast.
8137 switch (V.getOpcode()) {
8138 case ISD::CONCAT_VECTORS: {
8139 int OperandSize = Mask.size() / V.getNumOperands();
8140 V = V.getOperand(BroadcastIdx / OperandSize);
8141 BroadcastIdx %= OperandSize;
8145 case ISD::INSERT_SUBVECTOR: {
8146 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8147 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8151 int BeginIdx = (int)ConstantIdx->getZExtValue();
8153 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
8154 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8155 BroadcastIdx -= BeginIdx;
8166 // Check if this is a broadcast of a scalar. We special case lowering
8167 // for scalars so that we can more effectively fold with loads.
8168 // First, look through bitcast: if the original value has a larger element
8169 // type than the shuffle, the broadcast element is in essence truncated.
8170 // Make that explicit to ease folding.
8171 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
8172 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
8173 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
8174 return TruncBroadcast;
8176 MVT BroadcastVT = VT;
8178 // Also check the simpler case, where we can directly reuse the scalar.
8179 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8180 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8181 V = V.getOperand(BroadcastIdx);
8183 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8184 // Only AVX2 has register broadcasts.
8185 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8187 } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
8188 // 32-bit targets need to load i64 as a f64 and then bitcast the result.
8189 if (!Subtarget->is64Bit() && VT.getScalarType() == MVT::i64)
8190 BroadcastVT = MVT::getVectorVT(MVT::f64, VT.getVectorNumElements());
8192 // If we are broadcasting a load that is only used by the shuffle
8193 // then we can reduce the vector load to the broadcasted scalar load.
8194 LoadSDNode *Ld = cast<LoadSDNode>(V);
8195 SDValue BaseAddr = Ld->getOperand(1);
8196 EVT AddrVT = BaseAddr.getValueType();
8197 EVT SVT = BroadcastVT.getScalarType();
8198 unsigned Offset = BroadcastIdx * SVT.getStoreSize();
8199 SDValue NewAddr = DAG.getNode(
8200 ISD::ADD, DL, AddrVT, BaseAddr,
8201 DAG.getConstant(Offset, DL, AddrVT));
8202 V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
8203 DAG.getMachineFunction().getMachineMemOperand(
8204 Ld->getMemOperand(), Offset, SVT.getStoreSize()));
8205 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8206 // We can't broadcast from a vector register without AVX2, and we can only
8207 // broadcast from the zero-element of a vector register.
8211 V = DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, V);
8212 return DAG.getBitcast(VT, V);
8215 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8216 // INSERTPS when the V1 elements are already in the correct locations
8217 // because otherwise we can just always use two SHUFPS instructions which
8218 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8219 // perform INSERTPS if a single V1 element is out of place and all V2
8220 // elements are zeroable.
8221 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8223 SelectionDAG &DAG) {
8224 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8225 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8226 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8227 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8229 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8232 int V1DstIndex = -1;
8233 int V2DstIndex = -1;
8234 bool V1UsedInPlace = false;
8236 for (int i = 0; i < 4; ++i) {
8237 // Synthesize a zero mask from the zeroable elements (includes undefs).
8243 // Flag if we use any V1 inputs in place.
8245 V1UsedInPlace = true;
8249 // We can only insert a single non-zeroable element.
8250 if (V1DstIndex != -1 || V2DstIndex != -1)
8254 // V1 input out of place for insertion.
8257 // V2 input for insertion.
8262 // Don't bother if we have no (non-zeroable) element for insertion.
8263 if (V1DstIndex == -1 && V2DstIndex == -1)
8266 // Determine element insertion src/dst indices. The src index is from the
8267 // start of the inserted vector, not the start of the concatenated vector.
8268 unsigned V2SrcIndex = 0;
8269 if (V1DstIndex != -1) {
8270 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8271 // and don't use the original V2 at all.
8272 V2SrcIndex = Mask[V1DstIndex];
8273 V2DstIndex = V1DstIndex;
8276 V2SrcIndex = Mask[V2DstIndex] - 4;
8279 // If no V1 inputs are used in place, then the result is created only from
8280 // the zero mask and the V2 insertion - so remove V1 dependency.
8282 V1 = DAG.getUNDEF(MVT::v4f32);
8284 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8285 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8287 // Insert the V2 element into the desired position.
8289 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8290 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8293 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8294 /// UNPCK instruction.
8296 /// This specifically targets cases where we end up with alternating between
8297 /// the two inputs, and so can permute them into something that feeds a single
8298 /// UNPCK instruction. Note that this routine only targets integer vectors
8299 /// because for floating point vectors we have a generalized SHUFPS lowering
8300 /// strategy that handles everything that doesn't *exactly* match an unpack,
8301 /// making this clever lowering unnecessary.
8302 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8303 SDValue V1, SDValue V2,
8305 SelectionDAG &DAG) {
8306 assert(!VT.isFloatingPoint() &&
8307 "This routine only supports integer vectors.");
8308 assert(!isSingleInputShuffleMask(Mask) &&
8309 "This routine should only be used when blending two inputs.");
8310 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8312 int Size = Mask.size();
8314 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8315 return M >= 0 && M % Size < Size / 2;
8317 int NumHiInputs = std::count_if(
8318 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8320 bool UnpackLo = NumLoInputs >= NumHiInputs;
8322 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8323 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8324 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8326 for (int i = 0; i < Size; ++i) {
8330 // Each element of the unpack contains Scale elements from this mask.
8331 int UnpackIdx = i / Scale;
8333 // We only handle the case where V1 feeds the first slots of the unpack.
8334 // We rely on canonicalization to ensure this is the case.
8335 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8338 // Setup the mask for this input. The indexing is tricky as we have to
8339 // handle the unpack stride.
8340 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8341 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8345 // If we will have to shuffle both inputs to use the unpack, check whether
8346 // we can just unpack first and shuffle the result. If so, skip this unpack.
8347 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8348 !isNoopShuffleMask(V2Mask))
8351 // Shuffle the inputs into place.
8352 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8353 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8355 // Cast the inputs to the type we will use to unpack them.
8356 V1 = DAG.getBitcast(UnpackVT, V1);
8357 V2 = DAG.getBitcast(UnpackVT, V2);
8359 // Unpack the inputs and cast the result back to the desired type.
8360 return DAG.getBitcast(
8361 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8365 // We try each unpack from the largest to the smallest to try and find one
8366 // that fits this mask.
8367 int OrigNumElements = VT.getVectorNumElements();
8368 int OrigScalarSize = VT.getScalarSizeInBits();
8369 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8370 int Scale = ScalarSize / OrigScalarSize;
8371 int NumElements = OrigNumElements / Scale;
8372 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8373 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8377 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8379 if (NumLoInputs == 0 || NumHiInputs == 0) {
8380 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8381 "We have to have *some* inputs!");
8382 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8384 // FIXME: We could consider the total complexity of the permute of each
8385 // possible unpacking. Or at the least we should consider how many
8386 // half-crossings are created.
8387 // FIXME: We could consider commuting the unpacks.
8389 SmallVector<int, 32> PermMask;
8390 PermMask.assign(Size, -1);
8391 for (int i = 0; i < Size; ++i) {
8395 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8398 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8400 return DAG.getVectorShuffle(
8401 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8403 DAG.getUNDEF(VT), PermMask);
8409 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8411 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8412 /// support for floating point shuffles but not integer shuffles. These
8413 /// instructions will incur a domain crossing penalty on some chips though so
8414 /// it is better to avoid lowering through this for integer vectors where
8416 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8417 const X86Subtarget *Subtarget,
8418 SelectionDAG &DAG) {
8420 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8421 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8422 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8423 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8424 ArrayRef<int> Mask = SVOp->getMask();
8425 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8427 if (isSingleInputShuffleMask(Mask)) {
8428 // Use low duplicate instructions for masks that match their pattern.
8429 if (Subtarget->hasSSE3())
8430 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8431 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8433 // Straight shuffle of a single input vector. Simulate this by using the
8434 // single input as both of the "inputs" to this instruction..
8435 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8437 if (Subtarget->hasAVX()) {
8438 // If we have AVX, we can use VPERMILPS which will allow folding a load
8439 // into the shuffle.
8440 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8441 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8444 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8445 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8447 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8448 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8450 // If we have a single input, insert that into V1 if we can do so cheaply.
8451 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8452 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8453 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8455 // Try inverting the insertion since for v2 masks it is easy to do and we
8456 // can't reliably sort the mask one way or the other.
8457 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8458 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8459 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8460 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8464 // Try to use one of the special instruction patterns to handle two common
8465 // blend patterns if a zero-blend above didn't work.
8466 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8467 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8468 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8469 // We can either use a special instruction to load over the low double or
8470 // to move just the low double.
8472 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8474 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8476 if (Subtarget->hasSSE41())
8477 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8481 // Use dedicated unpack instructions for masks that match their pattern.
8483 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8486 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8487 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8488 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8491 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8493 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8494 /// the integer unit to minimize domain crossing penalties. However, for blends
8495 /// it falls back to the floating point shuffle operation with appropriate bit
8497 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8498 const X86Subtarget *Subtarget,
8499 SelectionDAG &DAG) {
8501 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8502 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8503 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8504 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8505 ArrayRef<int> Mask = SVOp->getMask();
8506 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8508 if (isSingleInputShuffleMask(Mask)) {
8509 // Check for being able to broadcast a single element.
8510 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8511 Mask, Subtarget, DAG))
8514 // Straight shuffle of a single input vector. For everything from SSE2
8515 // onward this has a single fast instruction with no scary immediates.
8516 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8517 V1 = DAG.getBitcast(MVT::v4i32, V1);
8518 int WidenedMask[4] = {
8519 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8520 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8521 return DAG.getBitcast(
8523 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8524 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8526 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8527 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8528 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8529 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8531 // If we have a blend of two PACKUS operations an the blend aligns with the
8532 // low and half halves, we can just merge the PACKUS operations. This is
8533 // particularly important as it lets us merge shuffles that this routine itself
8535 auto GetPackNode = [](SDValue V) {
8536 while (V.getOpcode() == ISD::BITCAST)
8537 V = V.getOperand(0);
8539 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8541 if (SDValue V1Pack = GetPackNode(V1))
8542 if (SDValue V2Pack = GetPackNode(V2))
8543 return DAG.getBitcast(MVT::v2i64,
8544 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8545 Mask[0] == 0 ? V1Pack.getOperand(0)
8546 : V1Pack.getOperand(1),
8547 Mask[1] == 2 ? V2Pack.getOperand(0)
8548 : V2Pack.getOperand(1)));
8550 // Try to use shift instructions.
8552 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8555 // When loading a scalar and then shuffling it into a vector we can often do
8556 // the insertion cheaply.
8557 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8558 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8560 // Try inverting the insertion since for v2 masks it is easy to do and we
8561 // can't reliably sort the mask one way or the other.
8562 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8563 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8564 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8567 // We have different paths for blend lowering, but they all must use the
8568 // *exact* same predicate.
8569 bool IsBlendSupported = Subtarget->hasSSE41();
8570 if (IsBlendSupported)
8571 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8575 // Use dedicated unpack instructions for masks that match their pattern.
8577 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8580 // Try to use byte rotation instructions.
8581 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8582 if (Subtarget->hasSSSE3())
8583 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8584 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8587 // If we have direct support for blends, we should lower by decomposing into
8588 // a permute. That will be faster than the domain cross.
8589 if (IsBlendSupported)
8590 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8593 // We implement this with SHUFPD which is pretty lame because it will likely
8594 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8595 // However, all the alternatives are still more cycles and newer chips don't
8596 // have this problem. It would be really nice if x86 had better shuffles here.
8597 V1 = DAG.getBitcast(MVT::v2f64, V1);
8598 V2 = DAG.getBitcast(MVT::v2f64, V2);
8599 return DAG.getBitcast(MVT::v2i64,
8600 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8603 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8605 /// This is used to disable more specialized lowerings when the shufps lowering
8606 /// will happen to be efficient.
8607 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8608 // This routine only handles 128-bit shufps.
8609 assert(Mask.size() == 4 && "Unsupported mask size!");
8611 // To lower with a single SHUFPS we need to have the low half and high half
8612 // each requiring a single input.
8613 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8615 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8621 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8623 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8624 /// It makes no assumptions about whether this is the *best* lowering, it simply
8626 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8627 ArrayRef<int> Mask, SDValue V1,
8628 SDValue V2, SelectionDAG &DAG) {
8629 SDValue LowV = V1, HighV = V2;
8630 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8633 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8635 if (NumV2Elements == 1) {
8637 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8640 // Compute the index adjacent to V2Index and in the same half by toggling
8642 int V2AdjIndex = V2Index ^ 1;
8644 if (Mask[V2AdjIndex] == -1) {
8645 // Handles all the cases where we have a single V2 element and an undef.
8646 // This will only ever happen in the high lanes because we commute the
8647 // vector otherwise.
8649 std::swap(LowV, HighV);
8650 NewMask[V2Index] -= 4;
8652 // Handle the case where the V2 element ends up adjacent to a V1 element.
8653 // To make this work, blend them together as the first step.
8654 int V1Index = V2AdjIndex;
8655 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8656 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8657 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8659 // Now proceed to reconstruct the final blend as we have the necessary
8660 // high or low half formed.
8667 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8668 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8670 } else if (NumV2Elements == 2) {
8671 if (Mask[0] < 4 && Mask[1] < 4) {
8672 // Handle the easy case where we have V1 in the low lanes and V2 in the
8676 } else if (Mask[2] < 4 && Mask[3] < 4) {
8677 // We also handle the reversed case because this utility may get called
8678 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8679 // arrange things in the right direction.
8685 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8686 // trying to place elements directly, just blend them and set up the final
8687 // shuffle to place them.
8689 // The first two blend mask elements are for V1, the second two are for
8691 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8692 Mask[2] < 4 ? Mask[2] : Mask[3],
8693 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8694 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8695 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8696 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8698 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8701 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8702 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8703 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8704 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8707 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8708 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8711 /// \brief Lower 4-lane 32-bit floating point shuffles.
8713 /// Uses instructions exclusively from the floating point unit to minimize
8714 /// domain crossing penalties, as these are sufficient to implement all v4f32
8716 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8717 const X86Subtarget *Subtarget,
8718 SelectionDAG &DAG) {
8720 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8721 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8722 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8723 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8724 ArrayRef<int> Mask = SVOp->getMask();
8725 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8728 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8730 if (NumV2Elements == 0) {
8731 // Check for being able to broadcast a single element.
8732 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8733 Mask, Subtarget, DAG))
8736 // Use even/odd duplicate instructions for masks that match their pattern.
8737 if (Subtarget->hasSSE3()) {
8738 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8739 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8740 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8741 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8744 if (Subtarget->hasAVX()) {
8745 // If we have AVX, we can use VPERMILPS which will allow folding a load
8746 // into the shuffle.
8747 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8748 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8751 // Otherwise, use a straight shuffle of a single input vector. We pass the
8752 // input vector to both operands to simulate this with a SHUFPS.
8753 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8754 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8757 // There are special ways we can lower some single-element blends. However, we
8758 // have custom ways we can lower more complex single-element blends below that
8759 // we defer to if both this and BLENDPS fail to match, so restrict this to
8760 // when the V2 input is targeting element 0 of the mask -- that is the fast
8762 if (NumV2Elements == 1 && Mask[0] >= 4)
8763 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8764 Mask, Subtarget, DAG))
8767 if (Subtarget->hasSSE41()) {
8768 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8772 // Use INSERTPS if we can complete the shuffle efficiently.
8773 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8776 if (!isSingleSHUFPSMask(Mask))
8777 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8778 DL, MVT::v4f32, V1, V2, Mask, DAG))
8782 // Use dedicated unpack instructions for masks that match their pattern.
8784 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8787 // Otherwise fall back to a SHUFPS lowering strategy.
8788 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8791 /// \brief Lower 4-lane i32 vector shuffles.
8793 /// We try to handle these with integer-domain shuffles where we can, but for
8794 /// blends we use the floating point domain blend instructions.
8795 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8796 const X86Subtarget *Subtarget,
8797 SelectionDAG &DAG) {
8799 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8800 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8801 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8802 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8803 ArrayRef<int> Mask = SVOp->getMask();
8804 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8806 // Whenever we can lower this as a zext, that instruction is strictly faster
8807 // than any alternative. It also allows us to fold memory operands into the
8808 // shuffle in many cases.
8809 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8810 Mask, Subtarget, DAG))
8814 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8816 if (NumV2Elements == 0) {
8817 // Check for being able to broadcast a single element.
8818 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8819 Mask, Subtarget, DAG))
8822 // Straight shuffle of a single input vector. For everything from SSE2
8823 // onward this has a single fast instruction with no scary immediates.
8824 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8825 // but we aren't actually going to use the UNPCK instruction because doing
8826 // so prevents folding a load into this instruction or making a copy.
8827 const int UnpackLoMask[] = {0, 0, 1, 1};
8828 const int UnpackHiMask[] = {2, 2, 3, 3};
8829 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8830 Mask = UnpackLoMask;
8831 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8832 Mask = UnpackHiMask;
8834 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8835 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8838 // Try to use shift instructions.
8840 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8843 // There are special ways we can lower some single-element blends.
8844 if (NumV2Elements == 1)
8845 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8846 Mask, Subtarget, DAG))
8849 // We have different paths for blend lowering, but they all must use the
8850 // *exact* same predicate.
8851 bool IsBlendSupported = Subtarget->hasSSE41();
8852 if (IsBlendSupported)
8853 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8857 if (SDValue Masked =
8858 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8861 // Use dedicated unpack instructions for masks that match their pattern.
8863 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8866 // Try to use byte rotation instructions.
8867 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8868 if (Subtarget->hasSSSE3())
8869 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8870 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8873 // If we have direct support for blends, we should lower by decomposing into
8874 // a permute. That will be faster than the domain cross.
8875 if (IsBlendSupported)
8876 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8879 // Try to lower by permuting the inputs into an unpack instruction.
8880 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8884 // We implement this with SHUFPS because it can blend from two vectors.
8885 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8886 // up the inputs, bypassing domain shift penalties that we would encur if we
8887 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8889 return DAG.getBitcast(
8891 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8892 DAG.getBitcast(MVT::v4f32, V2), Mask));
8895 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8896 /// shuffle lowering, and the most complex part.
8898 /// The lowering strategy is to try to form pairs of input lanes which are
8899 /// targeted at the same half of the final vector, and then use a dword shuffle
8900 /// to place them onto the right half, and finally unpack the paired lanes into
8901 /// their final position.
8903 /// The exact breakdown of how to form these dword pairs and align them on the
8904 /// correct sides is really tricky. See the comments within the function for
8905 /// more of the details.
8907 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8908 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8909 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8910 /// vector, form the analogous 128-bit 8-element Mask.
8911 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8912 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8913 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8914 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8915 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8917 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8918 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8919 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8921 SmallVector<int, 4> LoInputs;
8922 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8923 [](int M) { return M >= 0; });
8924 std::sort(LoInputs.begin(), LoInputs.end());
8925 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8926 SmallVector<int, 4> HiInputs;
8927 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8928 [](int M) { return M >= 0; });
8929 std::sort(HiInputs.begin(), HiInputs.end());
8930 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8932 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8933 int NumHToL = LoInputs.size() - NumLToL;
8935 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8936 int NumHToH = HiInputs.size() - NumLToH;
8937 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8938 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8939 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8940 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8942 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8943 // such inputs we can swap two of the dwords across the half mark and end up
8944 // with <=2 inputs to each half in each half. Once there, we can fall through
8945 // to the generic code below. For example:
8947 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8948 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8950 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8951 // and an existing 2-into-2 on the other half. In this case we may have to
8952 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8953 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8954 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8955 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8956 // half than the one we target for fixing) will be fixed when we re-enter this
8957 // path. We will also combine away any sequence of PSHUFD instructions that
8958 // result into a single instruction. Here is an example of the tricky case:
8960 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8961 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8963 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8965 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8966 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8968 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8969 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8971 // The result is fine to be handled by the generic logic.
8972 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8973 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8974 int AOffset, int BOffset) {
8975 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8976 "Must call this with A having 3 or 1 inputs from the A half.");
8977 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8978 "Must call this with B having 1 or 3 inputs from the B half.");
8979 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8980 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8982 bool ThreeAInputs = AToAInputs.size() == 3;
8984 // Compute the index of dword with only one word among the three inputs in
8985 // a half by taking the sum of the half with three inputs and subtracting
8986 // the sum of the actual three inputs. The difference is the remaining
8989 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8990 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8991 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8992 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8993 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8994 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8995 int TripleNonInputIdx =
8996 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8997 TripleDWord = TripleNonInputIdx / 2;
8999 // We use xor with one to compute the adjacent DWord to whichever one the
9001 OneInputDWord = (OneInput / 2) ^ 1;
9003 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
9004 // and BToA inputs. If there is also such a problem with the BToB and AToB
9005 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
9006 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
9007 // is essential that we don't *create* a 3<-1 as then we might oscillate.
9008 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
9009 // Compute how many inputs will be flipped by swapping these DWords. We
9011 // to balance this to ensure we don't form a 3-1 shuffle in the other
9013 int NumFlippedAToBInputs =
9014 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
9015 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
9016 int NumFlippedBToBInputs =
9017 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
9018 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
9019 if ((NumFlippedAToBInputs == 1 &&
9020 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
9021 (NumFlippedBToBInputs == 1 &&
9022 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
9023 // We choose whether to fix the A half or B half based on whether that
9024 // half has zero flipped inputs. At zero, we may not be able to fix it
9025 // with that half. We also bias towards fixing the B half because that
9026 // will more commonly be the high half, and we have to bias one way.
9027 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
9028 ArrayRef<int> Inputs) {
9029 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
9030 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
9031 PinnedIdx ^ 1) != Inputs.end();
9032 // Determine whether the free index is in the flipped dword or the
9033 // unflipped dword based on where the pinned index is. We use this bit
9034 // in an xor to conditionally select the adjacent dword.
9035 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
9036 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9037 FixFreeIdx) != Inputs.end();
9038 if (IsFixIdxInput == IsFixFreeIdxInput)
9040 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9041 FixFreeIdx) != Inputs.end();
9042 assert(IsFixIdxInput != IsFixFreeIdxInput &&
9043 "We need to be changing the number of flipped inputs!");
9044 int PSHUFHalfMask[] = {0, 1, 2, 3};
9045 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
9046 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
9048 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
9051 if (M != -1 && M == FixIdx)
9053 else if (M != -1 && M == FixFreeIdx)
9056 if (NumFlippedBToBInputs != 0) {
9058 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9059 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
9061 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
9062 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
9063 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9068 int PSHUFDMask[] = {0, 1, 2, 3};
9069 PSHUFDMask[ADWord] = BDWord;
9070 PSHUFDMask[BDWord] = ADWord;
9073 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9074 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9076 // Adjust the mask to match the new locations of A and B.
9078 if (M != -1 && M/2 == ADWord)
9079 M = 2 * BDWord + M % 2;
9080 else if (M != -1 && M/2 == BDWord)
9081 M = 2 * ADWord + M % 2;
9083 // Recurse back into this routine to re-compute state now that this isn't
9084 // a 3 and 1 problem.
9085 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
9088 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9089 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9090 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9091 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9093 // At this point there are at most two inputs to the low and high halves from
9094 // each half. That means the inputs can always be grouped into dwords and
9095 // those dwords can then be moved to the correct half with a dword shuffle.
9096 // We use at most one low and one high word shuffle to collect these paired
9097 // inputs into dwords, and finally a dword shuffle to place them.
9098 int PSHUFLMask[4] = {-1, -1, -1, -1};
9099 int PSHUFHMask[4] = {-1, -1, -1, -1};
9100 int PSHUFDMask[4] = {-1, -1, -1, -1};
9102 // First fix the masks for all the inputs that are staying in their
9103 // original halves. This will then dictate the targets of the cross-half
9105 auto fixInPlaceInputs =
9106 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9107 MutableArrayRef<int> SourceHalfMask,
9108 MutableArrayRef<int> HalfMask, int HalfOffset) {
9109 if (InPlaceInputs.empty())
9111 if (InPlaceInputs.size() == 1) {
9112 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9113 InPlaceInputs[0] - HalfOffset;
9114 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9117 if (IncomingInputs.empty()) {
9118 // Just fix all of the in place inputs.
9119 for (int Input : InPlaceInputs) {
9120 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9121 PSHUFDMask[Input / 2] = Input / 2;
9126 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9127 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9128 InPlaceInputs[0] - HalfOffset;
9129 // Put the second input next to the first so that they are packed into
9130 // a dword. We find the adjacent index by toggling the low bit.
9131 int AdjIndex = InPlaceInputs[0] ^ 1;
9132 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9133 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9134 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9136 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9137 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9139 // Now gather the cross-half inputs and place them into a free dword of
9140 // their target half.
9141 // FIXME: This operation could almost certainly be simplified dramatically to
9142 // look more like the 3-1 fixing operation.
9143 auto moveInputsToRightHalf = [&PSHUFDMask](
9144 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9145 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9146 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9148 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9149 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9151 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9153 int LowWord = Word & ~1;
9154 int HighWord = Word | 1;
9155 return isWordClobbered(SourceHalfMask, LowWord) ||
9156 isWordClobbered(SourceHalfMask, HighWord);
9159 if (IncomingInputs.empty())
9162 if (ExistingInputs.empty()) {
9163 // Map any dwords with inputs from them into the right half.
9164 for (int Input : IncomingInputs) {
9165 // If the source half mask maps over the inputs, turn those into
9166 // swaps and use the swapped lane.
9167 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9168 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9169 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9170 Input - SourceOffset;
9171 // We have to swap the uses in our half mask in one sweep.
9172 for (int &M : HalfMask)
9173 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9175 else if (M == Input)
9176 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9178 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9179 Input - SourceOffset &&
9180 "Previous placement doesn't match!");
9182 // Note that this correctly re-maps both when we do a swap and when
9183 // we observe the other side of the swap above. We rely on that to
9184 // avoid swapping the members of the input list directly.
9185 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9188 // Map the input's dword into the correct half.
9189 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9190 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9192 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9194 "Previous placement doesn't match!");
9197 // And just directly shift any other-half mask elements to be same-half
9198 // as we will have mirrored the dword containing the element into the
9199 // same position within that half.
9200 for (int &M : HalfMask)
9201 if (M >= SourceOffset && M < SourceOffset + 4) {
9202 M = M - SourceOffset + DestOffset;
9203 assert(M >= 0 && "This should never wrap below zero!");
9208 // Ensure we have the input in a viable dword of its current half. This
9209 // is particularly tricky because the original position may be clobbered
9210 // by inputs being moved and *staying* in that half.
9211 if (IncomingInputs.size() == 1) {
9212 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9213 int InputFixed = std::find(std::begin(SourceHalfMask),
9214 std::end(SourceHalfMask), -1) -
9215 std::begin(SourceHalfMask) + SourceOffset;
9216 SourceHalfMask[InputFixed - SourceOffset] =
9217 IncomingInputs[0] - SourceOffset;
9218 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9220 IncomingInputs[0] = InputFixed;
9222 } else if (IncomingInputs.size() == 2) {
9223 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9224 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9225 // We have two non-adjacent or clobbered inputs we need to extract from
9226 // the source half. To do this, we need to map them into some adjacent
9227 // dword slot in the source mask.
9228 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9229 IncomingInputs[1] - SourceOffset};
9231 // If there is a free slot in the source half mask adjacent to one of
9232 // the inputs, place the other input in it. We use (Index XOR 1) to
9233 // compute an adjacent index.
9234 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9235 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9236 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9237 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9238 InputsFixed[1] = InputsFixed[0] ^ 1;
9239 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9240 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9241 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9242 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9243 InputsFixed[0] = InputsFixed[1] ^ 1;
9244 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9245 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9246 // The two inputs are in the same DWord but it is clobbered and the
9247 // adjacent DWord isn't used at all. Move both inputs to the free
9249 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9250 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9251 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9252 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9254 // The only way we hit this point is if there is no clobbering
9255 // (because there are no off-half inputs to this half) and there is no
9256 // free slot adjacent to one of the inputs. In this case, we have to
9257 // swap an input with a non-input.
9258 for (int i = 0; i < 4; ++i)
9259 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9260 "We can't handle any clobbers here!");
9261 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9262 "Cannot have adjacent inputs here!");
9264 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9265 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9267 // We also have to update the final source mask in this case because
9268 // it may need to undo the above swap.
9269 for (int &M : FinalSourceHalfMask)
9270 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9271 M = InputsFixed[1] + SourceOffset;
9272 else if (M == InputsFixed[1] + SourceOffset)
9273 M = (InputsFixed[0] ^ 1) + SourceOffset;
9275 InputsFixed[1] = InputsFixed[0] ^ 1;
9278 // Point everything at the fixed inputs.
9279 for (int &M : HalfMask)
9280 if (M == IncomingInputs[0])
9281 M = InputsFixed[0] + SourceOffset;
9282 else if (M == IncomingInputs[1])
9283 M = InputsFixed[1] + SourceOffset;
9285 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9286 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9289 llvm_unreachable("Unhandled input size!");
9292 // Now hoist the DWord down to the right half.
9293 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9294 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9295 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9296 for (int &M : HalfMask)
9297 for (int Input : IncomingInputs)
9299 M = FreeDWord * 2 + Input % 2;
9301 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9302 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9303 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9304 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9306 // Now enact all the shuffles we've computed to move the inputs into their
9308 if (!isNoopShuffleMask(PSHUFLMask))
9309 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9310 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9311 if (!isNoopShuffleMask(PSHUFHMask))
9312 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9313 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9314 if (!isNoopShuffleMask(PSHUFDMask))
9317 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9318 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9320 // At this point, each half should contain all its inputs, and we can then
9321 // just shuffle them into their final position.
9322 assert(std::count_if(LoMask.begin(), LoMask.end(),
9323 [](int M) { return M >= 4; }) == 0 &&
9324 "Failed to lift all the high half inputs to the low mask!");
9325 assert(std::count_if(HiMask.begin(), HiMask.end(),
9326 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9327 "Failed to lift all the low half inputs to the high mask!");
9329 // Do a half shuffle for the low mask.
9330 if (!isNoopShuffleMask(LoMask))
9331 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9332 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9334 // Do a half shuffle with the high mask after shifting its values down.
9335 for (int &M : HiMask)
9338 if (!isNoopShuffleMask(HiMask))
9339 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9340 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9345 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9346 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9347 SDValue V2, ArrayRef<int> Mask,
9348 SelectionDAG &DAG, bool &V1InUse,
9350 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9356 int Size = Mask.size();
9357 int Scale = 16 / Size;
9358 for (int i = 0; i < 16; ++i) {
9359 if (Mask[i / Scale] == -1) {
9360 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9362 const int ZeroMask = 0x80;
9363 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9365 int V2Idx = Mask[i / Scale] < Size
9367 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9368 if (Zeroable[i / Scale])
9369 V1Idx = V2Idx = ZeroMask;
9370 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9371 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9372 V1InUse |= (ZeroMask != V1Idx);
9373 V2InUse |= (ZeroMask != V2Idx);
9378 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9379 DAG.getBitcast(MVT::v16i8, V1),
9380 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9382 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9383 DAG.getBitcast(MVT::v16i8, V2),
9384 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9386 // If we need shuffled inputs from both, blend the two.
9388 if (V1InUse && V2InUse)
9389 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9391 V = V1InUse ? V1 : V2;
9393 // Cast the result back to the correct type.
9394 return DAG.getBitcast(VT, V);
9397 /// \brief Generic lowering of 8-lane i16 shuffles.
9399 /// This handles both single-input shuffles and combined shuffle/blends with
9400 /// two inputs. The single input shuffles are immediately delegated to
9401 /// a dedicated lowering routine.
9403 /// The blends are lowered in one of three fundamental ways. If there are few
9404 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9405 /// of the input is significantly cheaper when lowered as an interleaving of
9406 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9407 /// halves of the inputs separately (making them have relatively few inputs)
9408 /// and then concatenate them.
9409 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9410 const X86Subtarget *Subtarget,
9411 SelectionDAG &DAG) {
9413 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9414 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9415 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9416 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9417 ArrayRef<int> OrigMask = SVOp->getMask();
9418 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9419 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9420 MutableArrayRef<int> Mask(MaskStorage);
9422 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9424 // Whenever we can lower this as a zext, that instruction is strictly faster
9425 // than any alternative.
9426 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9427 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9430 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9432 auto isV2 = [](int M) { return M >= 8; };
9434 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9436 if (NumV2Inputs == 0) {
9437 // Check for being able to broadcast a single element.
9438 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9439 Mask, Subtarget, DAG))
9442 // Try to use shift instructions.
9444 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9447 // Use dedicated unpack instructions for masks that match their pattern.
9449 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9452 // Try to use byte rotation instructions.
9453 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9454 Mask, Subtarget, DAG))
9457 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9461 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9462 "All single-input shuffles should be canonicalized to be V1-input "
9465 // Try to use shift instructions.
9467 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9470 // See if we can use SSE4A Extraction / Insertion.
9471 if (Subtarget->hasSSE4A())
9472 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9475 // There are special ways we can lower some single-element blends.
9476 if (NumV2Inputs == 1)
9477 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9478 Mask, Subtarget, DAG))
9481 // We have different paths for blend lowering, but they all must use the
9482 // *exact* same predicate.
9483 bool IsBlendSupported = Subtarget->hasSSE41();
9484 if (IsBlendSupported)
9485 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9489 if (SDValue Masked =
9490 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9493 // Use dedicated unpack instructions for masks that match their pattern.
9495 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9498 // Try to use byte rotation instructions.
9499 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9500 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9503 if (SDValue BitBlend =
9504 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9507 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9511 // If we can't directly blend but can use PSHUFB, that will be better as it
9512 // can both shuffle and set up the inefficient blend.
9513 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9514 bool V1InUse, V2InUse;
9515 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9519 // We can always bit-blend if we have to so the fallback strategy is to
9520 // decompose into single-input permutes and blends.
9521 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9525 /// \brief Check whether a compaction lowering can be done by dropping even
9526 /// elements and compute how many times even elements must be dropped.
9528 /// This handles shuffles which take every Nth element where N is a power of
9529 /// two. Example shuffle masks:
9531 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9532 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9533 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9534 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9535 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9536 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9538 /// Any of these lanes can of course be undef.
9540 /// This routine only supports N <= 3.
9541 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9544 /// \returns N above, or the number of times even elements must be dropped if
9545 /// there is such a number. Otherwise returns zero.
9546 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9547 // Figure out whether we're looping over two inputs or just one.
9548 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9550 // The modulus for the shuffle vector entries is based on whether this is
9551 // a single input or not.
9552 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9553 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9554 "We should only be called with masks with a power-of-2 size!");
9556 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9558 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9559 // and 2^3 simultaneously. This is because we may have ambiguity with
9560 // partially undef inputs.
9561 bool ViableForN[3] = {true, true, true};
9563 for (int i = 0, e = Mask.size(); i < e; ++i) {
9564 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9569 bool IsAnyViable = false;
9570 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9571 if (ViableForN[j]) {
9574 // The shuffle mask must be equal to (i * 2^N) % M.
9575 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9578 ViableForN[j] = false;
9580 // Early exit if we exhaust the possible powers of two.
9585 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9589 // Return 0 as there is no viable power of two.
9593 /// \brief Generic lowering of v16i8 shuffles.
9595 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9596 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9597 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9598 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9600 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9601 const X86Subtarget *Subtarget,
9602 SelectionDAG &DAG) {
9604 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9605 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9606 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9607 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9608 ArrayRef<int> Mask = SVOp->getMask();
9609 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9611 // Try to use shift instructions.
9613 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9616 // Try to use byte rotation instructions.
9617 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9618 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9621 // Try to use a zext lowering.
9622 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9623 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9626 // See if we can use SSE4A Extraction / Insertion.
9627 if (Subtarget->hasSSE4A())
9628 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9632 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9634 // For single-input shuffles, there are some nicer lowering tricks we can use.
9635 if (NumV2Elements == 0) {
9636 // Check for being able to broadcast a single element.
9637 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9638 Mask, Subtarget, DAG))
9641 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9642 // Notably, this handles splat and partial-splat shuffles more efficiently.
9643 // However, it only makes sense if the pre-duplication shuffle simplifies
9644 // things significantly. Currently, this means we need to be able to
9645 // express the pre-duplication shuffle as an i16 shuffle.
9647 // FIXME: We should check for other patterns which can be widened into an
9648 // i16 shuffle as well.
9649 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9650 for (int i = 0; i < 16; i += 2)
9651 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9656 auto tryToWidenViaDuplication = [&]() -> SDValue {
9657 if (!canWidenViaDuplication(Mask))
9659 SmallVector<int, 4> LoInputs;
9660 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9661 [](int M) { return M >= 0 && M < 8; });
9662 std::sort(LoInputs.begin(), LoInputs.end());
9663 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9665 SmallVector<int, 4> HiInputs;
9666 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9667 [](int M) { return M >= 8; });
9668 std::sort(HiInputs.begin(), HiInputs.end());
9669 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9672 bool TargetLo = LoInputs.size() >= HiInputs.size();
9673 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9674 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9676 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9677 SmallDenseMap<int, int, 8> LaneMap;
9678 for (int I : InPlaceInputs) {
9679 PreDupI16Shuffle[I/2] = I/2;
9682 int j = TargetLo ? 0 : 4, je = j + 4;
9683 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9684 // Check if j is already a shuffle of this input. This happens when
9685 // there are two adjacent bytes after we move the low one.
9686 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9687 // If we haven't yet mapped the input, search for a slot into which
9689 while (j < je && PreDupI16Shuffle[j] != -1)
9693 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9696 // Map this input with the i16 shuffle.
9697 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9700 // Update the lane map based on the mapping we ended up with.
9701 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9703 V1 = DAG.getBitcast(
9705 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9706 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9708 // Unpack the bytes to form the i16s that will be shuffled into place.
9709 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9710 MVT::v16i8, V1, V1);
9712 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9713 for (int i = 0; i < 16; ++i)
9714 if (Mask[i] != -1) {
9715 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9716 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9717 if (PostDupI16Shuffle[i / 2] == -1)
9718 PostDupI16Shuffle[i / 2] = MappedMask;
9720 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9721 "Conflicting entrties in the original shuffle!");
9723 return DAG.getBitcast(
9725 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9726 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9728 if (SDValue V = tryToWidenViaDuplication())
9732 if (SDValue Masked =
9733 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9736 // Use dedicated unpack instructions for masks that match their pattern.
9738 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9741 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9742 // with PSHUFB. It is important to do this before we attempt to generate any
9743 // blends but after all of the single-input lowerings. If the single input
9744 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9745 // want to preserve that and we can DAG combine any longer sequences into
9746 // a PSHUFB in the end. But once we start blending from multiple inputs,
9747 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9748 // and there are *very* few patterns that would actually be faster than the
9749 // PSHUFB approach because of its ability to zero lanes.
9751 // FIXME: The only exceptions to the above are blends which are exact
9752 // interleavings with direct instructions supporting them. We currently don't
9753 // handle those well here.
9754 if (Subtarget->hasSSSE3()) {
9755 bool V1InUse = false;
9756 bool V2InUse = false;
9758 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9759 DAG, V1InUse, V2InUse);
9761 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9762 // do so. This avoids using them to handle blends-with-zero which is
9763 // important as a single pshufb is significantly faster for that.
9764 if (V1InUse && V2InUse) {
9765 if (Subtarget->hasSSE41())
9766 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9767 Mask, Subtarget, DAG))
9770 // We can use an unpack to do the blending rather than an or in some
9771 // cases. Even though the or may be (very minorly) more efficient, we
9772 // preference this lowering because there are common cases where part of
9773 // the complexity of the shuffles goes away when we do the final blend as
9775 // FIXME: It might be worth trying to detect if the unpack-feeding
9776 // shuffles will both be pshufb, in which case we shouldn't bother with
9778 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9779 DL, MVT::v16i8, V1, V2, Mask, DAG))
9786 // There are special ways we can lower some single-element blends.
9787 if (NumV2Elements == 1)
9788 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9789 Mask, Subtarget, DAG))
9792 if (SDValue BitBlend =
9793 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9796 // Check whether a compaction lowering can be done. This handles shuffles
9797 // which take every Nth element for some even N. See the helper function for
9800 // We special case these as they can be particularly efficiently handled with
9801 // the PACKUSB instruction on x86 and they show up in common patterns of
9802 // rearranging bytes to truncate wide elements.
9803 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9804 // NumEvenDrops is the power of two stride of the elements. Another way of
9805 // thinking about it is that we need to drop the even elements this many
9806 // times to get the original input.
9807 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9809 // First we need to zero all the dropped bytes.
9810 assert(NumEvenDrops <= 3 &&
9811 "No support for dropping even elements more than 3 times.");
9812 // We use the mask type to pick which bytes are preserved based on how many
9813 // elements are dropped.
9814 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9815 SDValue ByteClearMask = DAG.getBitcast(
9816 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9817 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9819 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9821 // Now pack things back together.
9822 V1 = DAG.getBitcast(MVT::v8i16, V1);
9823 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9824 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9825 for (int i = 1; i < NumEvenDrops; ++i) {
9826 Result = DAG.getBitcast(MVT::v8i16, Result);
9827 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9833 // Handle multi-input cases by blending single-input shuffles.
9834 if (NumV2Elements > 0)
9835 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9838 // The fallback path for single-input shuffles widens this into two v8i16
9839 // vectors with unpacks, shuffles those, and then pulls them back together
9843 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9844 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9845 for (int i = 0; i < 16; ++i)
9847 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9849 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9851 SDValue VLoHalf, VHiHalf;
9852 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9853 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9855 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9856 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9857 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9858 [](int M) { return M >= 0 && M % 2 == 1; })) {
9859 // Use a mask to drop the high bytes.
9860 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9861 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9862 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9864 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9865 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9867 // Squash the masks to point directly into VLoHalf.
9868 for (int &M : LoBlendMask)
9871 for (int &M : HiBlendMask)
9875 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9876 // VHiHalf so that we can blend them as i16s.
9877 VLoHalf = DAG.getBitcast(
9878 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9879 VHiHalf = DAG.getBitcast(
9880 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9883 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9884 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9886 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9889 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9891 /// This routine breaks down the specific type of 128-bit shuffle and
9892 /// dispatches to the lowering routines accordingly.
9893 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9894 MVT VT, const X86Subtarget *Subtarget,
9895 SelectionDAG &DAG) {
9896 switch (VT.SimpleTy) {
9898 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9900 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9902 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9904 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9906 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9908 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9911 llvm_unreachable("Unimplemented!");
9915 /// \brief Helper function to test whether a shuffle mask could be
9916 /// simplified by widening the elements being shuffled.
9918 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9919 /// leaves it in an unspecified state.
9921 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9922 /// shuffle masks. The latter have the special property of a '-2' representing
9923 /// a zero-ed lane of a vector.
9924 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9925 SmallVectorImpl<int> &WidenedMask) {
9926 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9927 // If both elements are undef, its trivial.
9928 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9929 WidenedMask.push_back(SM_SentinelUndef);
9933 // Check for an undef mask and a mask value properly aligned to fit with
9934 // a pair of values. If we find such a case, use the non-undef mask's value.
9935 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9936 WidenedMask.push_back(Mask[i + 1] / 2);
9939 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9940 WidenedMask.push_back(Mask[i] / 2);
9944 // When zeroing, we need to spread the zeroing across both lanes to widen.
9945 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9946 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9947 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9948 WidenedMask.push_back(SM_SentinelZero);
9954 // Finally check if the two mask values are adjacent and aligned with
9956 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9957 WidenedMask.push_back(Mask[i] / 2);
9961 // Otherwise we can't safely widen the elements used in this shuffle.
9964 assert(WidenedMask.size() == Mask.size() / 2 &&
9965 "Incorrect size of mask after widening the elements!");
9970 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9972 /// This routine just extracts two subvectors, shuffles them independently, and
9973 /// then concatenates them back together. This should work effectively with all
9974 /// AVX vector shuffle types.
9975 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9976 SDValue V2, ArrayRef<int> Mask,
9977 SelectionDAG &DAG) {
9978 assert(VT.getSizeInBits() >= 256 &&
9979 "Only for 256-bit or wider vector shuffles!");
9980 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9981 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9983 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9984 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9986 int NumElements = VT.getVectorNumElements();
9987 int SplitNumElements = NumElements / 2;
9988 MVT ScalarVT = VT.getVectorElementType();
9989 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9991 // Rather than splitting build-vectors, just build two narrower build
9992 // vectors. This helps shuffling with splats and zeros.
9993 auto SplitVector = [&](SDValue V) {
9994 while (V.getOpcode() == ISD::BITCAST)
9995 V = V->getOperand(0);
9997 MVT OrigVT = V.getSimpleValueType();
9998 int OrigNumElements = OrigVT.getVectorNumElements();
9999 int OrigSplitNumElements = OrigNumElements / 2;
10000 MVT OrigScalarVT = OrigVT.getVectorElementType();
10001 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
10005 auto *BV = dyn_cast<BuildVectorSDNode>(V);
10007 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10008 DAG.getIntPtrConstant(0, DL));
10009 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10010 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
10013 SmallVector<SDValue, 16> LoOps, HiOps;
10014 for (int i = 0; i < OrigSplitNumElements; ++i) {
10015 LoOps.push_back(BV->getOperand(i));
10016 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
10018 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
10019 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
10021 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
10022 DAG.getBitcast(SplitVT, HiV));
10025 SDValue LoV1, HiV1, LoV2, HiV2;
10026 std::tie(LoV1, HiV1) = SplitVector(V1);
10027 std::tie(LoV2, HiV2) = SplitVector(V2);
10029 // Now create two 4-way blends of these half-width vectors.
10030 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
10031 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
10032 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
10033 for (int i = 0; i < SplitNumElements; ++i) {
10034 int M = HalfMask[i];
10035 if (M >= NumElements) {
10036 if (M >= NumElements + SplitNumElements)
10040 V2BlendMask.push_back(M - NumElements);
10041 V1BlendMask.push_back(-1);
10042 BlendMask.push_back(SplitNumElements + i);
10043 } else if (M >= 0) {
10044 if (M >= SplitNumElements)
10048 V2BlendMask.push_back(-1);
10049 V1BlendMask.push_back(M);
10050 BlendMask.push_back(i);
10052 V2BlendMask.push_back(-1);
10053 V1BlendMask.push_back(-1);
10054 BlendMask.push_back(-1);
10058 // Because the lowering happens after all combining takes place, we need to
10059 // manually combine these blend masks as much as possible so that we create
10060 // a minimal number of high-level vector shuffle nodes.
10062 // First try just blending the halves of V1 or V2.
10063 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10064 return DAG.getUNDEF(SplitVT);
10065 if (!UseLoV2 && !UseHiV2)
10066 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10067 if (!UseLoV1 && !UseHiV1)
10068 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10070 SDValue V1Blend, V2Blend;
10071 if (UseLoV1 && UseHiV1) {
10073 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10075 // We only use half of V1 so map the usage down into the final blend mask.
10076 V1Blend = UseLoV1 ? LoV1 : HiV1;
10077 for (int i = 0; i < SplitNumElements; ++i)
10078 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10079 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10081 if (UseLoV2 && UseHiV2) {
10083 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10085 // We only use half of V2 so map the usage down into the final blend mask.
10086 V2Blend = UseLoV2 ? LoV2 : HiV2;
10087 for (int i = 0; i < SplitNumElements; ++i)
10088 if (BlendMask[i] >= SplitNumElements)
10089 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10091 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10093 SDValue Lo = HalfBlend(LoMask);
10094 SDValue Hi = HalfBlend(HiMask);
10095 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10098 /// \brief Either split a vector in halves or decompose the shuffles and the
10101 /// This is provided as a good fallback for many lowerings of non-single-input
10102 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10103 /// between splitting the shuffle into 128-bit components and stitching those
10104 /// back together vs. extracting the single-input shuffles and blending those
10106 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10107 SDValue V2, ArrayRef<int> Mask,
10108 SelectionDAG &DAG) {
10109 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10110 "lower single-input shuffles as it "
10111 "could then recurse on itself.");
10112 int Size = Mask.size();
10114 // If this can be modeled as a broadcast of two elements followed by a blend,
10115 // prefer that lowering. This is especially important because broadcasts can
10116 // often fold with memory operands.
10117 auto DoBothBroadcast = [&] {
10118 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10121 if (V2BroadcastIdx == -1)
10122 V2BroadcastIdx = M - Size;
10123 else if (M - Size != V2BroadcastIdx)
10125 } else if (M >= 0) {
10126 if (V1BroadcastIdx == -1)
10127 V1BroadcastIdx = M;
10128 else if (M != V1BroadcastIdx)
10133 if (DoBothBroadcast())
10134 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10137 // If the inputs all stem from a single 128-bit lane of each input, then we
10138 // split them rather than blending because the split will decompose to
10139 // unusually few instructions.
10140 int LaneCount = VT.getSizeInBits() / 128;
10141 int LaneSize = Size / LaneCount;
10142 SmallBitVector LaneInputs[2];
10143 LaneInputs[0].resize(LaneCount, false);
10144 LaneInputs[1].resize(LaneCount, false);
10145 for (int i = 0; i < Size; ++i)
10147 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10148 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10149 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10151 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10152 // that the decomposed single-input shuffles don't end up here.
10153 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10156 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10157 /// a permutation and blend of those lanes.
10159 /// This essentially blends the out-of-lane inputs to each lane into the lane
10160 /// from a permuted copy of the vector. This lowering strategy results in four
10161 /// instructions in the worst case for a single-input cross lane shuffle which
10162 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10163 /// of. Special cases for each particular shuffle pattern should be handled
10164 /// prior to trying this lowering.
10165 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10166 SDValue V1, SDValue V2,
10167 ArrayRef<int> Mask,
10168 SelectionDAG &DAG) {
10169 // FIXME: This should probably be generalized for 512-bit vectors as well.
10170 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
10171 int LaneSize = Mask.size() / 2;
10173 // If there are only inputs from one 128-bit lane, splitting will in fact be
10174 // less expensive. The flags track whether the given lane contains an element
10175 // that crosses to another lane.
10176 bool LaneCrossing[2] = {false, false};
10177 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10178 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10179 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10180 if (!LaneCrossing[0] || !LaneCrossing[1])
10181 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10183 if (isSingleInputShuffleMask(Mask)) {
10184 SmallVector<int, 32> FlippedBlendMask;
10185 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10186 FlippedBlendMask.push_back(
10187 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10189 : Mask[i] % LaneSize +
10190 (i / LaneSize) * LaneSize + Size));
10192 // Flip the vector, and blend the results which should now be in-lane. The
10193 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10194 // 5 for the high source. The value 3 selects the high half of source 2 and
10195 // the value 2 selects the low half of source 2. We only use source 2 to
10196 // allow folding it into a memory operand.
10197 unsigned PERMMask = 3 | 2 << 4;
10198 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10199 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10200 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10203 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10204 // will be handled by the above logic and a blend of the results, much like
10205 // other patterns in AVX.
10206 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10209 /// \brief Handle lowering 2-lane 128-bit shuffles.
10210 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10211 SDValue V2, ArrayRef<int> Mask,
10212 const X86Subtarget *Subtarget,
10213 SelectionDAG &DAG) {
10214 // TODO: If minimizing size and one of the inputs is a zero vector and the
10215 // the zero vector has only one use, we could use a VPERM2X128 to save the
10216 // instruction bytes needed to explicitly generate the zero vector.
10218 // Blends are faster and handle all the non-lane-crossing cases.
10219 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10223 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10224 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10226 // If either input operand is a zero vector, use VPERM2X128 because its mask
10227 // allows us to replace the zero input with an implicit zero.
10228 if (!IsV1Zero && !IsV2Zero) {
10229 // Check for patterns which can be matched with a single insert of a 128-bit
10231 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10232 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10233 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10234 VT.getVectorNumElements() / 2);
10235 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10236 DAG.getIntPtrConstant(0, DL));
10237 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10238 OnlyUsesV1 ? V1 : V2,
10239 DAG.getIntPtrConstant(0, DL));
10240 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10244 // Otherwise form a 128-bit permutation. After accounting for undefs,
10245 // convert the 64-bit shuffle mask selection values into 128-bit
10246 // selection bits by dividing the indexes by 2 and shifting into positions
10247 // defined by a vperm2*128 instruction's immediate control byte.
10249 // The immediate permute control byte looks like this:
10250 // [1:0] - select 128 bits from sources for low half of destination
10252 // [3] - zero low half of destination
10253 // [5:4] - select 128 bits from sources for high half of destination
10255 // [7] - zero high half of destination
10257 int MaskLO = Mask[0];
10258 if (MaskLO == SM_SentinelUndef)
10259 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10261 int MaskHI = Mask[2];
10262 if (MaskHI == SM_SentinelUndef)
10263 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10265 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10267 // If either input is a zero vector, replace it with an undef input.
10268 // Shuffle mask values < 4 are selecting elements of V1.
10269 // Shuffle mask values >= 4 are selecting elements of V2.
10270 // Adjust each half of the permute mask by clearing the half that was
10271 // selecting the zero vector and setting the zero mask bit.
10273 V1 = DAG.getUNDEF(VT);
10275 PermMask = (PermMask & 0xf0) | 0x08;
10277 PermMask = (PermMask & 0x0f) | 0x80;
10280 V2 = DAG.getUNDEF(VT);
10282 PermMask = (PermMask & 0xf0) | 0x08;
10284 PermMask = (PermMask & 0x0f) | 0x80;
10287 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10288 DAG.getConstant(PermMask, DL, MVT::i8));
10291 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10292 /// shuffling each lane.
10294 /// This will only succeed when the result of fixing the 128-bit lanes results
10295 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10296 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10297 /// the lane crosses early and then use simpler shuffles within each lane.
10299 /// FIXME: It might be worthwhile at some point to support this without
10300 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10301 /// in x86 only floating point has interesting non-repeating shuffles, and even
10302 /// those are still *marginally* more expensive.
10303 static SDValue lowerVectorShuffleByMerging128BitLanes(
10304 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10305 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10306 assert(!isSingleInputShuffleMask(Mask) &&
10307 "This is only useful with multiple inputs.");
10309 int Size = Mask.size();
10310 int LaneSize = 128 / VT.getScalarSizeInBits();
10311 int NumLanes = Size / LaneSize;
10312 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10314 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10315 // check whether the in-128-bit lane shuffles share a repeating pattern.
10316 SmallVector<int, 4> Lanes;
10317 Lanes.resize(NumLanes, -1);
10318 SmallVector<int, 4> InLaneMask;
10319 InLaneMask.resize(LaneSize, -1);
10320 for (int i = 0; i < Size; ++i) {
10324 int j = i / LaneSize;
10326 if (Lanes[j] < 0) {
10327 // First entry we've seen for this lane.
10328 Lanes[j] = Mask[i] / LaneSize;
10329 } else if (Lanes[j] != Mask[i] / LaneSize) {
10330 // This doesn't match the lane selected previously!
10334 // Check that within each lane we have a consistent shuffle mask.
10335 int k = i % LaneSize;
10336 if (InLaneMask[k] < 0) {
10337 InLaneMask[k] = Mask[i] % LaneSize;
10338 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10339 // This doesn't fit a repeating in-lane mask.
10344 // First shuffle the lanes into place.
10345 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10346 VT.getSizeInBits() / 64);
10347 SmallVector<int, 8> LaneMask;
10348 LaneMask.resize(NumLanes * 2, -1);
10349 for (int i = 0; i < NumLanes; ++i)
10350 if (Lanes[i] >= 0) {
10351 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10352 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10355 V1 = DAG.getBitcast(LaneVT, V1);
10356 V2 = DAG.getBitcast(LaneVT, V2);
10357 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10359 // Cast it back to the type we actually want.
10360 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10362 // Now do a simple shuffle that isn't lane crossing.
10363 SmallVector<int, 8> NewMask;
10364 NewMask.resize(Size, -1);
10365 for (int i = 0; i < Size; ++i)
10367 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10368 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10369 "Must not introduce lane crosses at this point!");
10371 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10374 /// Lower shuffles where an entire half of a 256-bit vector is UNDEF.
10375 /// This allows for fast cases such as subvector extraction/insertion
10376 /// or shuffling smaller vector types which can lower more efficiently.
10377 static SDValue lowerVectorShuffleWithUndefHalf(SDLoc DL, MVT VT, SDValue V1,
10378 SDValue V2, ArrayRef<int> Mask,
10379 const X86Subtarget *Subtarget,
10380 SelectionDAG &DAG) {
10381 assert(VT.getSizeInBits() == 256 && "Expected 256-bit vector");
10383 unsigned NumElts = VT.getVectorNumElements();
10384 unsigned HalfNumElts = NumElts / 2;
10385 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(), HalfNumElts);
10387 bool UndefLower = isUndefInRange(Mask, 0, HalfNumElts);
10388 bool UndefUpper = isUndefInRange(Mask, HalfNumElts, HalfNumElts);
10389 if (!UndefLower && !UndefUpper)
10392 // Upper half is undef and lower half is whole upper subvector.
10393 // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
10395 isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) {
10396 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10397 DAG.getIntPtrConstant(HalfNumElts, DL));
10398 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10399 DAG.getIntPtrConstant(0, DL));
10402 // Lower half is undef and upper half is whole lower subvector.
10403 // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
10405 isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) {
10406 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10407 DAG.getIntPtrConstant(0, DL));
10408 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10409 DAG.getIntPtrConstant(HalfNumElts, DL));
10412 // AVX2 supports efficient immediate 64-bit element cross-lane shuffles.
10413 if (UndefLower && Subtarget->hasAVX2() &&
10414 (VT == MVT::v4f64 || VT == MVT::v4i64))
10417 // If the shuffle only uses the lower halves of the input operands,
10418 // then extract them and perform the 'half' shuffle at half width.
10419 // e.g. vector_shuffle <X, X, X, X, u, u, u, u> or <X, X, u, u>
10420 int HalfIdx1 = -1, HalfIdx2 = -1;
10421 SmallVector<int, 8> HalfMask;
10422 unsigned Offset = UndefLower ? HalfNumElts : 0;
10423 for (unsigned i = 0; i != HalfNumElts; ++i) {
10424 int M = Mask[i + Offset];
10426 HalfMask.push_back(M);
10430 // Determine which of the 4 half vectors this element is from.
10431 // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2.
10432 int HalfIdx = M / HalfNumElts;
10434 // Only shuffle using the lower halves of the inputs.
10435 // TODO: Investigate usefulness of shuffling with upper halves.
10436 if (HalfIdx != 0 && HalfIdx != 2)
10439 // Determine the element index into its half vector source.
10440 int HalfElt = M % HalfNumElts;
10442 // We can shuffle with up to 2 half vectors, set the new 'half'
10443 // shuffle mask accordingly.
10444 if (-1 == HalfIdx1 || HalfIdx1 == HalfIdx) {
10445 HalfMask.push_back(HalfElt);
10446 HalfIdx1 = HalfIdx;
10449 if (-1 == HalfIdx2 || HalfIdx2 == HalfIdx) {
10450 HalfMask.push_back(HalfElt + HalfNumElts);
10451 HalfIdx2 = HalfIdx;
10455 // Too many half vectors referenced.
10458 assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length");
10460 auto GetHalfVector = [&](int HalfIdx) {
10462 return DAG.getUNDEF(HalfVT);
10463 SDValue V = (HalfIdx < 2 ? V1 : V2);
10464 HalfIdx = (HalfIdx % 2) * HalfNumElts;
10465 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V,
10466 DAG.getIntPtrConstant(HalfIdx, DL));
10469 SDValue Half1 = GetHalfVector(HalfIdx1);
10470 SDValue Half2 = GetHalfVector(HalfIdx2);
10471 SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask);
10472 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V,
10473 DAG.getIntPtrConstant(Offset, DL));
10476 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10479 /// This returns true if the elements from a particular input are already in the
10480 /// slot required by the given mask and require no permutation.
10481 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10482 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10483 int Size = Mask.size();
10484 for (int i = 0; i < Size; ++i)
10485 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10491 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10492 ArrayRef<int> Mask, SDValue V1,
10493 SDValue V2, SelectionDAG &DAG) {
10495 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10496 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10497 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10498 int NumElts = VT.getVectorNumElements();
10499 bool ShufpdMask = true;
10500 bool CommutableMask = true;
10501 unsigned Immediate = 0;
10502 for (int i = 0; i < NumElts; ++i) {
10505 int Val = (i & 6) + NumElts * (i & 1);
10506 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10507 if (Mask[i] < Val || Mask[i] > Val + 1)
10508 ShufpdMask = false;
10509 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10510 CommutableMask = false;
10511 Immediate |= (Mask[i] % 2) << i;
10514 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10515 DAG.getConstant(Immediate, DL, MVT::i8));
10516 if (CommutableMask)
10517 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10518 DAG.getConstant(Immediate, DL, MVT::i8));
10522 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10524 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10525 /// isn't available.
10526 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10527 const X86Subtarget *Subtarget,
10528 SelectionDAG &DAG) {
10530 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10531 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10532 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10533 ArrayRef<int> Mask = SVOp->getMask();
10534 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10536 SmallVector<int, 4> WidenedMask;
10537 if (canWidenShuffleElements(Mask, WidenedMask))
10538 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10541 if (isSingleInputShuffleMask(Mask)) {
10542 // Check for being able to broadcast a single element.
10543 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10544 Mask, Subtarget, DAG))
10547 // Use low duplicate instructions for masks that match their pattern.
10548 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10549 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10551 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10552 // Non-half-crossing single input shuffles can be lowerid with an
10553 // interleaved permutation.
10554 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10555 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10556 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10557 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10560 // With AVX2 we have direct support for this permutation.
10561 if (Subtarget->hasAVX2())
10562 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10563 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10565 // Otherwise, fall back.
10566 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10570 // Use dedicated unpack instructions for masks that match their pattern.
10572 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10575 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10579 // Check if the blend happens to exactly fit that of SHUFPD.
10581 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10584 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10585 // shuffle. However, if we have AVX2 and either inputs are already in place,
10586 // we will be able to shuffle even across lanes the other input in a single
10587 // instruction so skip this pattern.
10588 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10589 isShuffleMaskInputInPlace(1, Mask))))
10590 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10591 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10594 // If we have AVX2 then we always want to lower with a blend because an v4 we
10595 // can fully permute the elements.
10596 if (Subtarget->hasAVX2())
10597 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10600 // Otherwise fall back on generic lowering.
10601 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10604 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10606 /// This routine is only called when we have AVX2 and thus a reasonable
10607 /// instruction set for v4i64 shuffling..
10608 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10609 const X86Subtarget *Subtarget,
10610 SelectionDAG &DAG) {
10612 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10613 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10614 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10615 ArrayRef<int> Mask = SVOp->getMask();
10616 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10617 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10619 SmallVector<int, 4> WidenedMask;
10620 if (canWidenShuffleElements(Mask, WidenedMask))
10621 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10624 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10628 // Check for being able to broadcast a single element.
10629 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10630 Mask, Subtarget, DAG))
10633 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10634 // use lower latency instructions that will operate on both 128-bit lanes.
10635 SmallVector<int, 2> RepeatedMask;
10636 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10637 if (isSingleInputShuffleMask(Mask)) {
10638 int PSHUFDMask[] = {-1, -1, -1, -1};
10639 for (int i = 0; i < 2; ++i)
10640 if (RepeatedMask[i] >= 0) {
10641 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10642 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10644 return DAG.getBitcast(
10646 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10647 DAG.getBitcast(MVT::v8i32, V1),
10648 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10652 // AVX2 provides a direct instruction for permuting a single input across
10654 if (isSingleInputShuffleMask(Mask))
10655 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10656 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10658 // Try to use shift instructions.
10659 if (SDValue Shift =
10660 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10663 // Use dedicated unpack instructions for masks that match their pattern.
10665 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10668 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10669 // shuffle. However, if we have AVX2 and either inputs are already in place,
10670 // we will be able to shuffle even across lanes the other input in a single
10671 // instruction so skip this pattern.
10672 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10673 isShuffleMaskInputInPlace(1, Mask))))
10674 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10675 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10678 // Otherwise fall back on generic blend lowering.
10679 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10683 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10685 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10686 /// isn't available.
10687 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10688 const X86Subtarget *Subtarget,
10689 SelectionDAG &DAG) {
10691 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10692 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10693 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10694 ArrayRef<int> Mask = SVOp->getMask();
10695 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10697 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10701 // Check for being able to broadcast a single element.
10702 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10703 Mask, Subtarget, DAG))
10706 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10707 // options to efficiently lower the shuffle.
10708 SmallVector<int, 4> RepeatedMask;
10709 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10710 assert(RepeatedMask.size() == 4 &&
10711 "Repeated masks must be half the mask width!");
10713 // Use even/odd duplicate instructions for masks that match their pattern.
10714 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10715 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10716 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10717 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10719 if (isSingleInputShuffleMask(Mask))
10720 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10721 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10723 // Use dedicated unpack instructions for masks that match their pattern.
10725 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10728 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10729 // have already handled any direct blends. We also need to squash the
10730 // repeated mask into a simulated v4f32 mask.
10731 for (int i = 0; i < 4; ++i)
10732 if (RepeatedMask[i] >= 8)
10733 RepeatedMask[i] -= 4;
10734 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10737 // If we have a single input shuffle with different shuffle patterns in the
10738 // two 128-bit lanes use the variable mask to VPERMILPS.
10739 if (isSingleInputShuffleMask(Mask)) {
10740 SDValue VPermMask[8];
10741 for (int i = 0; i < 8; ++i)
10742 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10743 : DAG.getConstant(Mask[i], DL, MVT::i32);
10744 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10745 return DAG.getNode(
10746 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10747 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10749 if (Subtarget->hasAVX2())
10750 return DAG.getNode(
10751 X86ISD::VPERMV, DL, MVT::v8f32,
10752 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10754 // Otherwise, fall back.
10755 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10759 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10761 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10762 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10765 // If we have AVX2 then we always want to lower with a blend because at v8 we
10766 // can fully permute the elements.
10767 if (Subtarget->hasAVX2())
10768 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10771 // Otherwise fall back on generic lowering.
10772 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10775 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10777 /// This routine is only called when we have AVX2 and thus a reasonable
10778 /// instruction set for v8i32 shuffling..
10779 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10780 const X86Subtarget *Subtarget,
10781 SelectionDAG &DAG) {
10783 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10784 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10785 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10786 ArrayRef<int> Mask = SVOp->getMask();
10787 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10788 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10790 // Whenever we can lower this as a zext, that instruction is strictly faster
10791 // than any alternative. It also allows us to fold memory operands into the
10792 // shuffle in many cases.
10793 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10794 Mask, Subtarget, DAG))
10797 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10801 // Check for being able to broadcast a single element.
10802 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10803 Mask, Subtarget, DAG))
10806 // If the shuffle mask is repeated in each 128-bit lane we can use more
10807 // efficient instructions that mirror the shuffles across the two 128-bit
10809 SmallVector<int, 4> RepeatedMask;
10810 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10811 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10812 if (isSingleInputShuffleMask(Mask))
10813 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10814 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10816 // Use dedicated unpack instructions for masks that match their pattern.
10818 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10822 // Try to use shift instructions.
10823 if (SDValue Shift =
10824 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10827 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10828 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10831 // If the shuffle patterns aren't repeated but it is a single input, directly
10832 // generate a cross-lane VPERMD instruction.
10833 if (isSingleInputShuffleMask(Mask)) {
10834 SDValue VPermMask[8];
10835 for (int i = 0; i < 8; ++i)
10836 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10837 : DAG.getConstant(Mask[i], DL, MVT::i32);
10838 return DAG.getNode(
10839 X86ISD::VPERMV, DL, MVT::v8i32,
10840 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10843 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10845 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10846 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10849 // Otherwise fall back on generic blend lowering.
10850 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10854 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10856 /// This routine is only called when we have AVX2 and thus a reasonable
10857 /// instruction set for v16i16 shuffling..
10858 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10859 const X86Subtarget *Subtarget,
10860 SelectionDAG &DAG) {
10862 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10863 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10864 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10865 ArrayRef<int> Mask = SVOp->getMask();
10866 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10867 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10869 // Whenever we can lower this as a zext, that instruction is strictly faster
10870 // than any alternative. It also allows us to fold memory operands into the
10871 // shuffle in many cases.
10872 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10873 Mask, Subtarget, DAG))
10876 // Check for being able to broadcast a single element.
10877 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10878 Mask, Subtarget, DAG))
10881 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10885 // Use dedicated unpack instructions for masks that match their pattern.
10887 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10890 // Try to use shift instructions.
10891 if (SDValue Shift =
10892 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10895 // Try to use byte rotation instructions.
10896 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10897 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10900 if (isSingleInputShuffleMask(Mask)) {
10901 // There are no generalized cross-lane shuffle operations available on i16
10903 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10904 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10907 SmallVector<int, 8> RepeatedMask;
10908 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10909 // As this is a single-input shuffle, the repeated mask should be
10910 // a strictly valid v8i16 mask that we can pass through to the v8i16
10911 // lowering to handle even the v16 case.
10912 return lowerV8I16GeneralSingleInputVectorShuffle(
10913 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10916 SDValue PSHUFBMask[32];
10917 for (int i = 0; i < 16; ++i) {
10918 if (Mask[i] == -1) {
10919 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10923 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10924 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10925 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10926 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10928 return DAG.getBitcast(MVT::v16i16,
10929 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10930 DAG.getBitcast(MVT::v32i8, V1),
10931 DAG.getNode(ISD::BUILD_VECTOR, DL,
10932 MVT::v32i8, PSHUFBMask)));
10935 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10937 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10938 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10941 // Otherwise fall back on generic lowering.
10942 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10945 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10947 /// This routine is only called when we have AVX2 and thus a reasonable
10948 /// instruction set for v32i8 shuffling..
10949 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10950 const X86Subtarget *Subtarget,
10951 SelectionDAG &DAG) {
10953 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10954 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10955 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10956 ArrayRef<int> Mask = SVOp->getMask();
10957 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10958 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10960 // Whenever we can lower this as a zext, that instruction is strictly faster
10961 // than any alternative. It also allows us to fold memory operands into the
10962 // shuffle in many cases.
10963 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10964 Mask, Subtarget, DAG))
10967 // Check for being able to broadcast a single element.
10968 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10969 Mask, Subtarget, DAG))
10972 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10976 // Use dedicated unpack instructions for masks that match their pattern.
10978 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10981 // Try to use shift instructions.
10982 if (SDValue Shift =
10983 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10986 // Try to use byte rotation instructions.
10987 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10988 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10991 if (isSingleInputShuffleMask(Mask)) {
10992 // There are no generalized cross-lane shuffle operations available on i8
10994 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10995 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10998 SDValue PSHUFBMask[32];
10999 for (int i = 0; i < 32; ++i)
11002 ? DAG.getUNDEF(MVT::i8)
11003 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
11006 return DAG.getNode(
11007 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
11008 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
11011 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11013 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11014 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
11017 // Otherwise fall back on generic lowering.
11018 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
11021 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
11023 /// This routine either breaks down the specific type of a 256-bit x86 vector
11024 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
11025 /// together based on the available instructions.
11026 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11027 MVT VT, const X86Subtarget *Subtarget,
11028 SelectionDAG &DAG) {
11030 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11031 ArrayRef<int> Mask = SVOp->getMask();
11033 // If we have a single input to the zero element, insert that into V1 if we
11034 // can do so cheaply.
11035 int NumElts = VT.getVectorNumElements();
11036 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
11037 return M >= NumElts;
11040 if (NumV2Elements == 1 && Mask[0] >= NumElts)
11041 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
11042 DL, VT, V1, V2, Mask, Subtarget, DAG))
11045 // Handle special cases where the lower or upper half is UNDEF.
11047 lowerVectorShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG))
11050 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
11051 // can check for those subtargets here and avoid much of the subtarget
11052 // querying in the per-vector-type lowering routines. With AVX1 we have
11053 // essentially *zero* ability to manipulate a 256-bit vector with integer
11054 // types. Since we'll use floating point types there eventually, just
11055 // immediately cast everything to a float and operate entirely in that domain.
11056 if (VT.isInteger() && !Subtarget->hasAVX2()) {
11057 int ElementBits = VT.getScalarSizeInBits();
11058 if (ElementBits < 32)
11059 // No floating point type available, decompose into 128-bit vectors.
11060 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11062 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
11063 VT.getVectorNumElements());
11064 V1 = DAG.getBitcast(FpVT, V1);
11065 V2 = DAG.getBitcast(FpVT, V2);
11066 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
11069 switch (VT.SimpleTy) {
11071 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11073 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11075 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11077 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11079 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11081 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11084 llvm_unreachable("Not a valid 256-bit x86 vector type!");
11088 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
11089 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
11090 ArrayRef<int> Mask,
11091 SDValue V1, SDValue V2,
11092 SelectionDAG &DAG) {
11093 assert(VT.getScalarSizeInBits() == 64 &&
11094 "Unexpected element type size for 128bit shuffle.");
11096 // To handle 256 bit vector requires VLX and most probably
11097 // function lowerV2X128VectorShuffle() is better solution.
11098 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
11100 SmallVector<int, 4> WidenedMask;
11101 if (!canWidenShuffleElements(Mask, WidenedMask))
11104 // Form a 128-bit permutation.
11105 // Convert the 64-bit shuffle mask selection values into 128-bit selection
11106 // bits defined by a vshuf64x2 instruction's immediate control byte.
11107 unsigned PermMask = 0, Imm = 0;
11108 unsigned ControlBitsNum = WidenedMask.size() / 2;
11110 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
11111 if (WidenedMask[i] == SM_SentinelZero)
11114 // Use first element in place of undef mask.
11115 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
11116 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
11119 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
11120 DAG.getConstant(PermMask, DL, MVT::i8));
11123 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
11124 ArrayRef<int> Mask, SDValue V1,
11125 SDValue V2, SelectionDAG &DAG) {
11127 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
11129 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
11130 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
11132 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
11133 if (isSingleInputShuffleMask(Mask))
11134 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
11136 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
11139 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
11140 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11141 const X86Subtarget *Subtarget,
11142 SelectionDAG &DAG) {
11144 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11145 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11146 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11147 ArrayRef<int> Mask = SVOp->getMask();
11148 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11150 if (SDValue Shuf128 =
11151 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
11154 if (SDValue Unpck =
11155 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
11158 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
11161 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
11162 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11163 const X86Subtarget *Subtarget,
11164 SelectionDAG &DAG) {
11166 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11167 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11168 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11169 ArrayRef<int> Mask = SVOp->getMask();
11170 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11172 if (SDValue Unpck =
11173 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
11176 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
11179 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11180 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11181 const X86Subtarget *Subtarget,
11182 SelectionDAG &DAG) {
11184 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11185 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11186 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11187 ArrayRef<int> Mask = SVOp->getMask();
11188 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11190 if (SDValue Shuf128 =
11191 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
11194 if (SDValue Unpck =
11195 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
11198 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
11201 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11202 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11203 const X86Subtarget *Subtarget,
11204 SelectionDAG &DAG) {
11206 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11207 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11208 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11209 ArrayRef<int> Mask = SVOp->getMask();
11210 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11212 if (SDValue Unpck =
11213 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
11216 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
11219 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11220 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11221 const X86Subtarget *Subtarget,
11222 SelectionDAG &DAG) {
11224 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11225 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11226 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11227 ArrayRef<int> Mask = SVOp->getMask();
11228 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11229 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11231 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
11234 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11235 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11236 const X86Subtarget *Subtarget,
11237 SelectionDAG &DAG) {
11239 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11240 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11241 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11242 ArrayRef<int> Mask = SVOp->getMask();
11243 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11244 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11246 // FIXME: Implement direct support for this type!
11247 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11250 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11252 /// This routine either breaks down the specific type of a 512-bit x86 vector
11253 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11254 /// together based on the available instructions.
11255 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11256 MVT VT, const X86Subtarget *Subtarget,
11257 SelectionDAG &DAG) {
11259 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11260 ArrayRef<int> Mask = SVOp->getMask();
11261 assert(Subtarget->hasAVX512() &&
11262 "Cannot lower 512-bit vectors w/ basic ISA!");
11264 // Check for being able to broadcast a single element.
11265 if (SDValue Broadcast =
11266 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
11269 // Dispatch to each element type for lowering. If we don't have supprot for
11270 // specific element type shuffles at 512 bits, immediately split them and
11271 // lower them. Each lowering routine of a given type is allowed to assume that
11272 // the requisite ISA extensions for that element type are available.
11273 switch (VT.SimpleTy) {
11275 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11277 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11279 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11281 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11283 if (Subtarget->hasBWI())
11284 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11287 if (Subtarget->hasBWI())
11288 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11292 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11295 // Otherwise fall back on splitting.
11296 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11299 // Lower vXi1 vector shuffles.
11300 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11301 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11302 // vector, shuffle and then truncate it back.
11303 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11304 MVT VT, const X86Subtarget *Subtarget,
11305 SelectionDAG &DAG) {
11307 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11308 ArrayRef<int> Mask = SVOp->getMask();
11309 assert(Subtarget->hasAVX512() &&
11310 "Cannot lower 512-bit vectors w/o basic ISA!");
11312 switch (VT.SimpleTy) {
11314 llvm_unreachable("Expected a vector of i1 elements");
11316 ExtVT = MVT::v2i64;
11319 ExtVT = MVT::v4i32;
11322 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11325 ExtVT = MVT::v16i32;
11328 ExtVT = MVT::v32i16;
11331 ExtVT = MVT::v64i8;
11335 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11336 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11337 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11338 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11340 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11343 V2 = DAG.getUNDEF(ExtVT);
11344 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11345 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11346 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11347 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11349 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11350 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11351 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11353 /// \brief Top-level lowering for x86 vector shuffles.
11355 /// This handles decomposition, canonicalization, and lowering of all x86
11356 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11357 /// above in helper routines. The canonicalization attempts to widen shuffles
11358 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11359 /// s.t. only one of the two inputs needs to be tested, etc.
11360 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11361 SelectionDAG &DAG) {
11362 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11363 ArrayRef<int> Mask = SVOp->getMask();
11364 SDValue V1 = Op.getOperand(0);
11365 SDValue V2 = Op.getOperand(1);
11366 MVT VT = Op.getSimpleValueType();
11367 int NumElements = VT.getVectorNumElements();
11369 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11371 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11372 "Can't lower MMX shuffles");
11374 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11375 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11376 if (V1IsUndef && V2IsUndef)
11377 return DAG.getUNDEF(VT);
11379 // When we create a shuffle node we put the UNDEF node to second operand,
11380 // but in some cases the first operand may be transformed to UNDEF.
11381 // In this case we should just commute the node.
11383 return DAG.getCommutedVectorShuffle(*SVOp);
11385 // Check for non-undef masks pointing at an undef vector and make the masks
11386 // undef as well. This makes it easier to match the shuffle based solely on
11390 if (M >= NumElements) {
11391 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11392 for (int &M : NewMask)
11393 if (M >= NumElements)
11395 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11398 // We actually see shuffles that are entirely re-arrangements of a set of
11399 // zero inputs. This mostly happens while decomposing complex shuffles into
11400 // simple ones. Directly lower these as a buildvector of zeros.
11401 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11402 if (Zeroable.all())
11403 return getZeroVector(VT, Subtarget, DAG, dl);
11405 // Try to collapse shuffles into using a vector type with fewer elements but
11406 // wider element types. We cap this to not form integers or floating point
11407 // elements wider than 64 bits, but it might be interesting to form i128
11408 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11409 SmallVector<int, 16> WidenedMask;
11410 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11411 canWidenShuffleElements(Mask, WidenedMask)) {
11412 MVT NewEltVT = VT.isFloatingPoint()
11413 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11414 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11415 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11416 // Make sure that the new vector type is legal. For example, v2f64 isn't
11418 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11419 V1 = DAG.getBitcast(NewVT, V1);
11420 V2 = DAG.getBitcast(NewVT, V2);
11421 return DAG.getBitcast(
11422 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11426 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11427 for (int M : SVOp->getMask())
11429 ++NumUndefElements;
11430 else if (M < NumElements)
11435 // Commute the shuffle as needed such that more elements come from V1 than
11436 // V2. This allows us to match the shuffle pattern strictly on how many
11437 // elements come from V1 without handling the symmetric cases.
11438 if (NumV2Elements > NumV1Elements)
11439 return DAG.getCommutedVectorShuffle(*SVOp);
11441 // When the number of V1 and V2 elements are the same, try to minimize the
11442 // number of uses of V2 in the low half of the vector. When that is tied,
11443 // ensure that the sum of indices for V1 is equal to or lower than the sum
11444 // indices for V2. When those are equal, try to ensure that the number of odd
11445 // indices for V1 is lower than the number of odd indices for V2.
11446 if (NumV1Elements == NumV2Elements) {
11447 int LowV1Elements = 0, LowV2Elements = 0;
11448 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11449 if (M >= NumElements)
11453 if (LowV2Elements > LowV1Elements) {
11454 return DAG.getCommutedVectorShuffle(*SVOp);
11455 } else if (LowV2Elements == LowV1Elements) {
11456 int SumV1Indices = 0, SumV2Indices = 0;
11457 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11458 if (SVOp->getMask()[i] >= NumElements)
11460 else if (SVOp->getMask()[i] >= 0)
11462 if (SumV2Indices < SumV1Indices) {
11463 return DAG.getCommutedVectorShuffle(*SVOp);
11464 } else if (SumV2Indices == SumV1Indices) {
11465 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11466 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11467 if (SVOp->getMask()[i] >= NumElements)
11468 NumV2OddIndices += i % 2;
11469 else if (SVOp->getMask()[i] >= 0)
11470 NumV1OddIndices += i % 2;
11471 if (NumV2OddIndices < NumV1OddIndices)
11472 return DAG.getCommutedVectorShuffle(*SVOp);
11477 // For each vector width, delegate to a specialized lowering routine.
11478 if (VT.is128BitVector())
11479 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11481 if (VT.is256BitVector())
11482 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11484 if (VT.is512BitVector())
11485 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11488 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11489 llvm_unreachable("Unimplemented!");
11492 // This function assumes its argument is a BUILD_VECTOR of constants or
11493 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11495 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11496 unsigned &MaskValue) {
11498 unsigned NumElems = BuildVector->getNumOperands();
11500 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11501 // We don't handle the >2 lanes case right now.
11502 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11506 unsigned NumElemsInLane = NumElems / NumLanes;
11508 // Blend for v16i16 should be symmetric for the both lanes.
11509 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11510 SDValue EltCond = BuildVector->getOperand(i);
11511 SDValue SndLaneEltCond =
11512 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11514 int Lane1Cond = -1, Lane2Cond = -1;
11515 if (isa<ConstantSDNode>(EltCond))
11516 Lane1Cond = !isNullConstant(EltCond);
11517 if (isa<ConstantSDNode>(SndLaneEltCond))
11518 Lane2Cond = !isNullConstant(SndLaneEltCond);
11520 unsigned LaneMask = 0;
11521 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11522 // Lane1Cond != 0, means we want the first argument.
11523 // Lane1Cond == 0, means we want the second argument.
11524 // The encoding of this argument is 0 for the first argument, 1
11525 // for the second. Therefore, invert the condition.
11526 LaneMask = !Lane1Cond << i;
11527 else if (Lane1Cond < 0)
11528 LaneMask = !Lane2Cond << i;
11532 MaskValue |= LaneMask;
11534 MaskValue |= LaneMask << NumElemsInLane;
11539 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11540 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11541 const X86Subtarget *Subtarget,
11542 SelectionDAG &DAG) {
11543 SDValue Cond = Op.getOperand(0);
11544 SDValue LHS = Op.getOperand(1);
11545 SDValue RHS = Op.getOperand(2);
11547 MVT VT = Op.getSimpleValueType();
11549 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11551 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11553 // Only non-legal VSELECTs reach this lowering, convert those into generic
11554 // shuffles and re-use the shuffle lowering path for blends.
11555 SmallVector<int, 32> Mask;
11556 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11557 SDValue CondElt = CondBV->getOperand(i);
11559 isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
11562 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11565 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11566 // A vselect where all conditions and data are constants can be optimized into
11567 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11568 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11569 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11570 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11573 // Try to lower this to a blend-style vector shuffle. This can handle all
11574 // constant condition cases.
11575 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11578 // Variable blends are only legal from SSE4.1 onward.
11579 if (!Subtarget->hasSSE41())
11582 // Only some types will be legal on some subtargets. If we can emit a legal
11583 // VSELECT-matching blend, return Op, and but if we need to expand, return
11585 switch (Op.getSimpleValueType().SimpleTy) {
11587 // Most of the vector types have blends past SSE4.1.
11591 // The byte blends for AVX vectors were introduced only in AVX2.
11592 if (Subtarget->hasAVX2())
11599 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11600 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11603 // FIXME: We should custom lower this by fixing the condition and using i8
11609 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11610 MVT VT = Op.getSimpleValueType();
11613 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11616 if (VT.getSizeInBits() == 8) {
11617 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11618 Op.getOperand(0), Op.getOperand(1));
11619 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11620 DAG.getValueType(VT));
11621 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11624 if (VT.getSizeInBits() == 16) {
11625 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11626 if (isNullConstant(Op.getOperand(1)))
11627 return DAG.getNode(
11628 ISD::TRUNCATE, dl, MVT::i16,
11629 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11630 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11631 Op.getOperand(1)));
11632 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11633 Op.getOperand(0), Op.getOperand(1));
11634 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11635 DAG.getValueType(VT));
11636 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11639 if (VT == MVT::f32) {
11640 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11641 // the result back to FR32 register. It's only worth matching if the
11642 // result has a single use which is a store or a bitcast to i32. And in
11643 // the case of a store, it's not worth it if the index is a constant 0,
11644 // because a MOVSSmr can be used instead, which is smaller and faster.
11645 if (!Op.hasOneUse())
11647 SDNode *User = *Op.getNode()->use_begin();
11648 if ((User->getOpcode() != ISD::STORE ||
11649 isNullConstant(Op.getOperand(1))) &&
11650 (User->getOpcode() != ISD::BITCAST ||
11651 User->getValueType(0) != MVT::i32))
11653 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11654 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11656 return DAG.getBitcast(MVT::f32, Extract);
11659 if (VT == MVT::i32 || VT == MVT::i64) {
11660 // ExtractPS/pextrq works with constant index.
11661 if (isa<ConstantSDNode>(Op.getOperand(1)))
11667 /// Extract one bit from mask vector, like v16i1 or v8i1.
11668 /// AVX-512 feature.
11670 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11671 SDValue Vec = Op.getOperand(0);
11673 MVT VecVT = Vec.getSimpleValueType();
11674 SDValue Idx = Op.getOperand(1);
11675 MVT EltVT = Op.getSimpleValueType();
11677 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11678 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11679 "Unexpected vector type in ExtractBitFromMaskVector");
11681 // variable index can't be handled in mask registers,
11682 // extend vector to VR512
11683 if (!isa<ConstantSDNode>(Idx)) {
11684 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11685 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11686 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11687 ExtVT.getVectorElementType(), Ext, Idx);
11688 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11691 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11692 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11693 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11694 rc = getRegClassFor(MVT::v16i1);
11695 unsigned MaxSift = rc->getSize()*8 - 1;
11696 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11697 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11698 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11699 DAG.getConstant(MaxSift, dl, MVT::i8));
11700 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11701 DAG.getIntPtrConstant(0, dl));
11705 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11706 SelectionDAG &DAG) const {
11708 SDValue Vec = Op.getOperand(0);
11709 MVT VecVT = Vec.getSimpleValueType();
11710 SDValue Idx = Op.getOperand(1);
11712 if (Op.getSimpleValueType() == MVT::i1)
11713 return ExtractBitFromMaskVector(Op, DAG);
11715 if (!isa<ConstantSDNode>(Idx)) {
11716 if (VecVT.is512BitVector() ||
11717 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11718 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11721 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11722 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11723 MaskEltVT.getSizeInBits());
11725 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11726 auto PtrVT = getPointerTy(DAG.getDataLayout());
11727 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11728 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11729 DAG.getConstant(0, dl, PtrVT));
11730 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11731 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11732 DAG.getConstant(0, dl, PtrVT));
11737 // If this is a 256-bit vector result, first extract the 128-bit vector and
11738 // then extract the element from the 128-bit vector.
11739 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11741 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11742 // Get the 128-bit vector.
11743 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11744 MVT EltVT = VecVT.getVectorElementType();
11746 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11747 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11749 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11750 // this can be done with a mask.
11751 IdxVal &= ElemsPerChunk - 1;
11752 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11753 DAG.getConstant(IdxVal, dl, MVT::i32));
11756 assert(VecVT.is128BitVector() && "Unexpected vector length");
11758 if (Subtarget->hasSSE41())
11759 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11762 MVT VT = Op.getSimpleValueType();
11763 // TODO: handle v16i8.
11764 if (VT.getSizeInBits() == 16) {
11765 SDValue Vec = Op.getOperand(0);
11766 if (isNullConstant(Op.getOperand(1)))
11767 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11768 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11769 DAG.getBitcast(MVT::v4i32, Vec),
11770 Op.getOperand(1)));
11771 // Transform it so it match pextrw which produces a 32-bit result.
11772 MVT EltVT = MVT::i32;
11773 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11774 Op.getOperand(0), Op.getOperand(1));
11775 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11776 DAG.getValueType(VT));
11777 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11780 if (VT.getSizeInBits() == 32) {
11781 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11785 // SHUFPS the element to the lowest double word, then movss.
11786 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11787 MVT VVT = Op.getOperand(0).getSimpleValueType();
11788 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11789 DAG.getUNDEF(VVT), Mask);
11790 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11791 DAG.getIntPtrConstant(0, dl));
11794 if (VT.getSizeInBits() == 64) {
11795 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11796 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11797 // to match extract_elt for f64.
11798 if (isNullConstant(Op.getOperand(1)))
11801 // UNPCKHPD the element to the lowest double word, then movsd.
11802 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11803 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11804 int Mask[2] = { 1, -1 };
11805 MVT VVT = Op.getOperand(0).getSimpleValueType();
11806 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11807 DAG.getUNDEF(VVT), Mask);
11808 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11809 DAG.getIntPtrConstant(0, dl));
11815 /// Insert one bit to mask vector, like v16i1 or v8i1.
11816 /// AVX-512 feature.
11818 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11820 SDValue Vec = Op.getOperand(0);
11821 SDValue Elt = Op.getOperand(1);
11822 SDValue Idx = Op.getOperand(2);
11823 MVT VecVT = Vec.getSimpleValueType();
11825 if (!isa<ConstantSDNode>(Idx)) {
11826 // Non constant index. Extend source and destination,
11827 // insert element and then truncate the result.
11828 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11829 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11830 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11831 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11832 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11833 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11836 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11837 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11839 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11840 DAG.getConstant(IdxVal, dl, MVT::i8));
11841 if (Vec.getOpcode() == ISD::UNDEF)
11843 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11846 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11847 SelectionDAG &DAG) const {
11848 MVT VT = Op.getSimpleValueType();
11849 MVT EltVT = VT.getVectorElementType();
11851 if (EltVT == MVT::i1)
11852 return InsertBitToMaskVector(Op, DAG);
11855 SDValue N0 = Op.getOperand(0);
11856 SDValue N1 = Op.getOperand(1);
11857 SDValue N2 = Op.getOperand(2);
11858 if (!isa<ConstantSDNode>(N2))
11860 auto *N2C = cast<ConstantSDNode>(N2);
11861 unsigned IdxVal = N2C->getZExtValue();
11863 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11864 // into that, and then insert the subvector back into the result.
11865 if (VT.is256BitVector() || VT.is512BitVector()) {
11866 // With a 256-bit vector, we can insert into the zero element efficiently
11867 // using a blend if we have AVX or AVX2 and the right data type.
11868 if (VT.is256BitVector() && IdxVal == 0) {
11869 // TODO: It is worthwhile to cast integer to floating point and back
11870 // and incur a domain crossing penalty if that's what we'll end up
11871 // doing anyway after extracting to a 128-bit vector.
11872 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11873 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11874 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11875 N2 = DAG.getIntPtrConstant(1, dl);
11876 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11880 // Get the desired 128-bit vector chunk.
11881 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11883 // Insert the element into the desired chunk.
11884 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11885 assert(isPowerOf2_32(NumEltsIn128));
11886 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11887 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11889 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11890 DAG.getConstant(IdxIn128, dl, MVT::i32));
11892 // Insert the changed part back into the bigger vector
11893 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11895 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11897 if (Subtarget->hasSSE41()) {
11898 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11900 if (VT == MVT::v8i16) {
11901 Opc = X86ISD::PINSRW;
11903 assert(VT == MVT::v16i8);
11904 Opc = X86ISD::PINSRB;
11907 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11909 if (N1.getValueType() != MVT::i32)
11910 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11911 if (N2.getValueType() != MVT::i32)
11912 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11913 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11916 if (EltVT == MVT::f32) {
11917 // Bits [7:6] of the constant are the source select. This will always be
11918 // zero here. The DAG Combiner may combine an extract_elt index into
11919 // these bits. For example (insert (extract, 3), 2) could be matched by
11920 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11921 // Bits [5:4] of the constant are the destination select. This is the
11922 // value of the incoming immediate.
11923 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11924 // combine either bitwise AND or insert of float 0.0 to set these bits.
11926 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11927 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11928 // If this is an insertion of 32-bits into the low 32-bits of
11929 // a vector, we prefer to generate a blend with immediate rather
11930 // than an insertps. Blends are simpler operations in hardware and so
11931 // will always have equal or better performance than insertps.
11932 // But if optimizing for size and there's a load folding opportunity,
11933 // generate insertps because blendps does not have a 32-bit memory
11935 N2 = DAG.getIntPtrConstant(1, dl);
11936 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11937 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11939 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11940 // Create this as a scalar to vector..
11941 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11942 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11945 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11946 // PINSR* works with constant index.
11951 if (EltVT == MVT::i8)
11954 if (EltVT.getSizeInBits() == 16) {
11955 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11956 // as its second argument.
11957 if (N1.getValueType() != MVT::i32)
11958 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11959 if (N2.getValueType() != MVT::i32)
11960 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11961 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11966 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11968 MVT OpVT = Op.getSimpleValueType();
11970 // If this is a 256-bit vector result, first insert into a 128-bit
11971 // vector and then insert into the 256-bit vector.
11972 if (!OpVT.is128BitVector()) {
11973 // Insert into a 128-bit vector.
11974 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11975 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11976 OpVT.getVectorNumElements() / SizeFactor);
11978 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11980 // Insert the 128-bit vector.
11981 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11984 if (OpVT == MVT::v1i64 &&
11985 Op.getOperand(0).getValueType() == MVT::i64)
11986 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11988 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11989 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11990 return DAG.getBitcast(
11991 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11994 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11995 // a simple subregister reference or explicit instructions to grab
11996 // upper bits of a vector.
11997 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11998 SelectionDAG &DAG) {
12000 SDValue In = Op.getOperand(0);
12001 SDValue Idx = Op.getOperand(1);
12002 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12003 MVT ResVT = Op.getSimpleValueType();
12004 MVT InVT = In.getSimpleValueType();
12006 if (Subtarget->hasFp256()) {
12007 if (ResVT.is128BitVector() &&
12008 (InVT.is256BitVector() || InVT.is512BitVector()) &&
12009 isa<ConstantSDNode>(Idx)) {
12010 return Extract128BitVector(In, IdxVal, DAG, dl);
12012 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
12013 isa<ConstantSDNode>(Idx)) {
12014 return Extract256BitVector(In, IdxVal, DAG, dl);
12020 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
12021 // simple superregister reference or explicit instructions to insert
12022 // the upper bits of a vector.
12023 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12024 SelectionDAG &DAG) {
12025 if (!Subtarget->hasAVX())
12029 SDValue Vec = Op.getOperand(0);
12030 SDValue SubVec = Op.getOperand(1);
12031 SDValue Idx = Op.getOperand(2);
12033 if (!isa<ConstantSDNode>(Idx))
12036 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12037 MVT OpVT = Op.getSimpleValueType();
12038 MVT SubVecVT = SubVec.getSimpleValueType();
12040 // Fold two 16-byte subvector loads into one 32-byte load:
12041 // (insert_subvector (insert_subvector undef, (load addr), 0),
12042 // (load addr + 16), Elts/2)
12044 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
12045 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
12046 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
12047 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
12048 if (Idx2 && Idx2->getZExtValue() == 0) {
12049 SDValue SubVec2 = Vec.getOperand(1);
12050 // If needed, look through a bitcast to get to the load.
12051 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
12052 SubVec2 = SubVec2.getOperand(0);
12054 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
12056 unsigned Alignment = FirstLd->getAlignment();
12057 unsigned AS = FirstLd->getAddressSpace();
12058 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
12059 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
12060 OpVT, AS, Alignment, &Fast) && Fast) {
12061 SDValue Ops[] = { SubVec2, SubVec };
12062 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
12069 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
12070 SubVecVT.is128BitVector())
12071 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
12073 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
12074 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
12076 if (OpVT.getVectorElementType() == MVT::i1)
12077 return Insert1BitVector(Op, DAG);
12082 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
12083 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
12084 // one of the above mentioned nodes. It has to be wrapped because otherwise
12085 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
12086 // be used to form addressing mode. These wrapped nodes will be selected
12089 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
12090 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
12092 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12093 // global base reg.
12094 unsigned char OpFlag = 0;
12095 unsigned WrapperKind = X86ISD::Wrapper;
12096 CodeModel::Model M = DAG.getTarget().getCodeModel();
12098 if (Subtarget->isPICStyleRIPRel() &&
12099 (M == CodeModel::Small || M == CodeModel::Kernel))
12100 WrapperKind = X86ISD::WrapperRIP;
12101 else if (Subtarget->isPICStyleGOT())
12102 OpFlag = X86II::MO_GOTOFF;
12103 else if (Subtarget->isPICStyleStubPIC())
12104 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12106 auto PtrVT = getPointerTy(DAG.getDataLayout());
12107 SDValue Result = DAG.getTargetConstantPool(
12108 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
12110 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12111 // With PIC, the address is actually $g + Offset.
12114 DAG.getNode(ISD::ADD, DL, PtrVT,
12115 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12121 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
12122 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
12124 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12125 // global base reg.
12126 unsigned char OpFlag = 0;
12127 unsigned WrapperKind = X86ISD::Wrapper;
12128 CodeModel::Model M = DAG.getTarget().getCodeModel();
12130 if (Subtarget->isPICStyleRIPRel() &&
12131 (M == CodeModel::Small || M == CodeModel::Kernel))
12132 WrapperKind = X86ISD::WrapperRIP;
12133 else if (Subtarget->isPICStyleGOT())
12134 OpFlag = X86II::MO_GOTOFF;
12135 else if (Subtarget->isPICStyleStubPIC())
12136 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12138 auto PtrVT = getPointerTy(DAG.getDataLayout());
12139 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
12141 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12143 // With PIC, the address is actually $g + Offset.
12146 DAG.getNode(ISD::ADD, DL, PtrVT,
12147 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12153 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
12154 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
12156 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12157 // global base reg.
12158 unsigned char OpFlag = 0;
12159 unsigned WrapperKind = X86ISD::Wrapper;
12160 CodeModel::Model M = DAG.getTarget().getCodeModel();
12162 if (Subtarget->isPICStyleRIPRel() &&
12163 (M == CodeModel::Small || M == CodeModel::Kernel)) {
12164 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
12165 OpFlag = X86II::MO_GOTPCREL;
12166 WrapperKind = X86ISD::WrapperRIP;
12167 } else if (Subtarget->isPICStyleGOT()) {
12168 OpFlag = X86II::MO_GOT;
12169 } else if (Subtarget->isPICStyleStubPIC()) {
12170 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
12171 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12172 OpFlag = X86II::MO_DARWIN_NONLAZY;
12175 auto PtrVT = getPointerTy(DAG.getDataLayout());
12176 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
12179 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12181 // With PIC, the address is actually $g + Offset.
12182 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12183 !Subtarget->is64Bit()) {
12185 DAG.getNode(ISD::ADD, DL, PtrVT,
12186 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12189 // For symbols that require a load from a stub to get the address, emit the
12191 if (isGlobalStubReference(OpFlag))
12192 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
12193 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12194 false, false, false, 0);
12200 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12201 // Create the TargetBlockAddressAddress node.
12202 unsigned char OpFlags =
12203 Subtarget->ClassifyBlockAddressReference();
12204 CodeModel::Model M = DAG.getTarget().getCodeModel();
12205 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12206 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12208 auto PtrVT = getPointerTy(DAG.getDataLayout());
12209 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
12211 if (Subtarget->isPICStyleRIPRel() &&
12212 (M == CodeModel::Small || M == CodeModel::Kernel))
12213 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12215 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12217 // With PIC, the address is actually $g + Offset.
12218 if (isGlobalRelativeToPICBase(OpFlags)) {
12219 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12220 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12227 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12228 int64_t Offset, SelectionDAG &DAG) const {
12229 // Create the TargetGlobalAddress node, folding in the constant
12230 // offset if it is legal.
12231 unsigned char OpFlags =
12232 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12233 CodeModel::Model M = DAG.getTarget().getCodeModel();
12234 auto PtrVT = getPointerTy(DAG.getDataLayout());
12236 if (OpFlags == X86II::MO_NO_FLAG &&
12237 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12238 // A direct static reference to a global.
12239 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
12242 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
12245 if (Subtarget->isPICStyleRIPRel() &&
12246 (M == CodeModel::Small || M == CodeModel::Kernel))
12247 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12249 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12251 // With PIC, the address is actually $g + Offset.
12252 if (isGlobalRelativeToPICBase(OpFlags)) {
12253 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12254 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12257 // For globals that require a load from a stub to get the address, emit the
12259 if (isGlobalStubReference(OpFlags))
12260 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12261 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12262 false, false, false, 0);
12264 // If there was a non-zero offset that we didn't fold, create an explicit
12265 // addition for it.
12267 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12268 DAG.getConstant(Offset, dl, PtrVT));
12274 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12275 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12276 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12277 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12281 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12282 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12283 unsigned char OperandFlags, bool LocalDynamic = false) {
12284 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12285 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12287 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12288 GA->getValueType(0),
12292 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12296 SDValue Ops[] = { Chain, TGA, *InFlag };
12297 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12299 SDValue Ops[] = { Chain, TGA };
12300 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12303 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12304 MFI->setAdjustsStack(true);
12305 MFI->setHasCalls(true);
12307 SDValue Flag = Chain.getValue(1);
12308 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12311 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12313 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12316 SDLoc dl(GA); // ? function entry point might be better
12317 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12318 DAG.getNode(X86ISD::GlobalBaseReg,
12319 SDLoc(), PtrVT), InFlag);
12320 InFlag = Chain.getValue(1);
12322 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12325 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12327 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12329 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12330 X86::RAX, X86II::MO_TLSGD);
12333 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12339 // Get the start address of the TLS block for this module.
12340 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12341 .getInfo<X86MachineFunctionInfo>();
12342 MFI->incNumLocalDynamicTLSAccesses();
12346 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12347 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12350 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12351 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12352 InFlag = Chain.getValue(1);
12353 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12354 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12357 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12361 unsigned char OperandFlags = X86II::MO_DTPOFF;
12362 unsigned WrapperKind = X86ISD::Wrapper;
12363 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12364 GA->getValueType(0),
12365 GA->getOffset(), OperandFlags);
12366 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12368 // Add x@dtpoff with the base.
12369 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12372 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12373 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12374 const EVT PtrVT, TLSModel::Model model,
12375 bool is64Bit, bool isPIC) {
12378 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12379 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12380 is64Bit ? 257 : 256));
12382 SDValue ThreadPointer =
12383 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12384 MachinePointerInfo(Ptr), false, false, false, 0);
12386 unsigned char OperandFlags = 0;
12387 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12389 unsigned WrapperKind = X86ISD::Wrapper;
12390 if (model == TLSModel::LocalExec) {
12391 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12392 } else if (model == TLSModel::InitialExec) {
12394 OperandFlags = X86II::MO_GOTTPOFF;
12395 WrapperKind = X86ISD::WrapperRIP;
12397 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12400 llvm_unreachable("Unexpected model");
12403 // emit "addl x@ntpoff,%eax" (local exec)
12404 // or "addl x@indntpoff,%eax" (initial exec)
12405 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12407 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12408 GA->getOffset(), OperandFlags);
12409 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12411 if (model == TLSModel::InitialExec) {
12412 if (isPIC && !is64Bit) {
12413 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12414 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12418 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12419 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12420 false, false, false, 0);
12423 // The address of the thread local variable is the add of the thread
12424 // pointer with the offset of the variable.
12425 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12429 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12431 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12433 // Cygwin uses emutls.
12434 // FIXME: It may be EmulatedTLS-generic also for X86-Android.
12435 if (Subtarget->isTargetWindowsCygwin())
12436 return LowerToTLSEmulatedModel(GA, DAG);
12438 const GlobalValue *GV = GA->getGlobal();
12439 auto PtrVT = getPointerTy(DAG.getDataLayout());
12441 if (Subtarget->isTargetELF()) {
12442 if (DAG.getTarget().Options.EmulatedTLS)
12443 return LowerToTLSEmulatedModel(GA, DAG);
12444 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12446 case TLSModel::GeneralDynamic:
12447 if (Subtarget->is64Bit())
12448 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12449 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12450 case TLSModel::LocalDynamic:
12451 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12452 Subtarget->is64Bit());
12453 case TLSModel::InitialExec:
12454 case TLSModel::LocalExec:
12455 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12456 DAG.getTarget().getRelocationModel() ==
12459 llvm_unreachable("Unknown TLS model.");
12462 if (Subtarget->isTargetDarwin()) {
12463 // Darwin only has one model of TLS. Lower to that.
12464 unsigned char OpFlag = 0;
12465 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12466 X86ISD::WrapperRIP : X86ISD::Wrapper;
12468 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12469 // global base reg.
12470 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12471 !Subtarget->is64Bit();
12473 OpFlag = X86II::MO_TLVP_PIC_BASE;
12475 OpFlag = X86II::MO_TLVP;
12477 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12478 GA->getValueType(0),
12479 GA->getOffset(), OpFlag);
12480 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12482 // With PIC32, the address is actually $g + Offset.
12484 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12485 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12488 // Lowering the machine isd will make sure everything is in the right
12490 SDValue Chain = DAG.getEntryNode();
12491 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12492 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, DL, true), DL);
12493 SDValue Args[] = { Chain, Offset };
12494 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12496 DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true),
12497 DAG.getIntPtrConstant(0, DL, true), SDValue(), DL);
12499 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12500 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12501 MFI->setAdjustsStack(true);
12503 // And our return value (tls address) is in the standard call return value
12505 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12506 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12509 if (Subtarget->isTargetKnownWindowsMSVC() ||
12510 Subtarget->isTargetWindowsGNU()) {
12511 // Just use the implicit TLS architecture
12512 // Need to generate someting similar to:
12513 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12515 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12516 // mov rcx, qword [rdx+rcx*8]
12517 // mov eax, .tls$:tlsvar
12518 // [rax+rcx] contains the address
12519 // Windows 64bit: gs:0x58
12520 // Windows 32bit: fs:__tls_array
12523 SDValue Chain = DAG.getEntryNode();
12525 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12526 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12527 // use its literal value of 0x2C.
12528 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12529 ? Type::getInt8PtrTy(*DAG.getContext(),
12531 : Type::getInt32PtrTy(*DAG.getContext(),
12534 SDValue TlsArray = Subtarget->is64Bit()
12535 ? DAG.getIntPtrConstant(0x58, dl)
12536 : (Subtarget->isTargetWindowsGNU()
12537 ? DAG.getIntPtrConstant(0x2C, dl)
12538 : DAG.getExternalSymbol("_tls_array", PtrVT));
12540 SDValue ThreadPointer =
12541 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12545 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12546 res = ThreadPointer;
12548 // Load the _tls_index variable
12549 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12550 if (Subtarget->is64Bit())
12551 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12552 MachinePointerInfo(), MVT::i32, false, false,
12555 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12558 auto &DL = DAG.getDataLayout();
12560 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12561 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12563 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12566 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12569 // Get the offset of start of .tls section
12570 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12571 GA->getValueType(0),
12572 GA->getOffset(), X86II::MO_SECREL);
12573 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12575 // The address of the thread local variable is the add of the thread
12576 // pointer with the offset of the variable.
12577 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12580 llvm_unreachable("TLS not implemented for this target.");
12583 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12584 /// and take a 2 x i32 value to shift plus a shift amount.
12585 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12586 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12587 MVT VT = Op.getSimpleValueType();
12588 unsigned VTBits = VT.getSizeInBits();
12590 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12591 SDValue ShOpLo = Op.getOperand(0);
12592 SDValue ShOpHi = Op.getOperand(1);
12593 SDValue ShAmt = Op.getOperand(2);
12594 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12595 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12597 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12598 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12599 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12600 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12601 : DAG.getConstant(0, dl, VT);
12603 SDValue Tmp2, Tmp3;
12604 if (Op.getOpcode() == ISD::SHL_PARTS) {
12605 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12606 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12608 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12609 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12612 // If the shift amount is larger or equal than the width of a part we can't
12613 // rely on the results of shld/shrd. Insert a test and select the appropriate
12614 // values for large shift amounts.
12615 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12616 DAG.getConstant(VTBits, dl, MVT::i8));
12617 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12618 AndNode, DAG.getConstant(0, dl, MVT::i8));
12621 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12622 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12623 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12625 if (Op.getOpcode() == ISD::SHL_PARTS) {
12626 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12627 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12629 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12630 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12633 SDValue Ops[2] = { Lo, Hi };
12634 return DAG.getMergeValues(Ops, dl);
12637 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12638 SelectionDAG &DAG) const {
12639 SDValue Src = Op.getOperand(0);
12640 MVT SrcVT = Src.getSimpleValueType();
12641 MVT VT = Op.getSimpleValueType();
12644 if (SrcVT.isVector()) {
12645 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12646 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12647 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12648 DAG.getUNDEF(SrcVT)));
12650 if (SrcVT.getVectorElementType() == MVT::i1) {
12651 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12652 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12653 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12658 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12659 "Unknown SINT_TO_FP to lower!");
12661 // These are really Legal; return the operand so the caller accepts it as
12663 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12665 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12666 Subtarget->is64Bit()) {
12670 unsigned Size = SrcVT.getSizeInBits()/8;
12671 MachineFunction &MF = DAG.getMachineFunction();
12672 auto PtrVT = getPointerTy(MF.getDataLayout());
12673 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12674 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12675 SDValue Chain = DAG.getStore(
12676 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12677 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12679 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12682 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12684 SelectionDAG &DAG) const {
12688 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12690 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12692 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12694 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12696 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12697 MachineMemOperand *MMO;
12699 int SSFI = FI->getIndex();
12700 MMO = DAG.getMachineFunction().getMachineMemOperand(
12701 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12702 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12704 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12705 StackSlot = StackSlot.getOperand(1);
12707 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12708 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12710 Tys, Ops, SrcVT, MMO);
12713 Chain = Result.getValue(1);
12714 SDValue InFlag = Result.getValue(2);
12716 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12717 // shouldn't be necessary except that RFP cannot be live across
12718 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12719 MachineFunction &MF = DAG.getMachineFunction();
12720 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12721 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12722 auto PtrVT = getPointerTy(MF.getDataLayout());
12723 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12724 Tys = DAG.getVTList(MVT::Other);
12726 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12728 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12729 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12730 MachineMemOperand::MOStore, SSFISize, SSFISize);
12732 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12733 Ops, Op.getValueType(), MMO);
12734 Result = DAG.getLoad(
12735 Op.getValueType(), DL, Chain, StackSlot,
12736 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12737 false, false, false, 0);
12743 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12744 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12745 SelectionDAG &DAG) const {
12746 // This algorithm is not obvious. Here it is what we're trying to output:
12749 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12750 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12752 haddpd %xmm0, %xmm0
12754 pshufd $0x4e, %xmm0, %xmm1
12760 LLVMContext *Context = DAG.getContext();
12762 // Build some magic constants.
12763 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12764 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12765 auto PtrVT = getPointerTy(DAG.getDataLayout());
12766 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12768 SmallVector<Constant*,2> CV1;
12770 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12771 APInt(64, 0x4330000000000000ULL))));
12773 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12774 APInt(64, 0x4530000000000000ULL))));
12775 Constant *C1 = ConstantVector::get(CV1);
12776 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12778 // Load the 64-bit value into an XMM register.
12779 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12782 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12783 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12784 false, false, false, 16);
12786 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12789 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12790 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12791 false, false, false, 16);
12792 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12793 // TODO: Are there any fast-math-flags to propagate here?
12794 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12797 if (Subtarget->hasSSE3()) {
12798 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12799 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12801 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12802 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12804 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12805 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12808 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12809 DAG.getIntPtrConstant(0, dl));
12812 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12813 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12814 SelectionDAG &DAG) const {
12816 // FP constant to bias correct the final result.
12817 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12820 // Load the 32-bit value into an XMM register.
12821 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12824 // Zero out the upper parts of the register.
12825 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12827 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12828 DAG.getBitcast(MVT::v2f64, Load),
12829 DAG.getIntPtrConstant(0, dl));
12831 // Or the load with the bias.
12832 SDValue Or = DAG.getNode(
12833 ISD::OR, dl, MVT::v2i64,
12834 DAG.getBitcast(MVT::v2i64,
12835 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12836 DAG.getBitcast(MVT::v2i64,
12837 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12839 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12840 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12842 // Subtract the bias.
12843 // TODO: Are there any fast-math-flags to propagate here?
12844 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12846 // Handle final rounding.
12847 MVT DestVT = Op.getSimpleValueType();
12849 if (DestVT.bitsLT(MVT::f64))
12850 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12851 DAG.getIntPtrConstant(0, dl));
12852 if (DestVT.bitsGT(MVT::f64))
12853 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12855 // Handle final rounding.
12859 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12860 const X86Subtarget &Subtarget) {
12861 // The algorithm is the following:
12862 // #ifdef __SSE4_1__
12863 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12864 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12865 // (uint4) 0x53000000, 0xaa);
12867 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12868 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12870 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12871 // return (float4) lo + fhi;
12873 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12874 // reassociate the two FADDs, and if we do that, the algorithm fails
12875 // spectacularly (PR24512).
12876 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12877 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12878 // there's also the MachineCombiner reassociations happening on Machine IR.
12879 if (DAG.getTarget().Options.UnsafeFPMath)
12883 SDValue V = Op->getOperand(0);
12884 MVT VecIntVT = V.getSimpleValueType();
12885 bool Is128 = VecIntVT == MVT::v4i32;
12886 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12887 // If we convert to something else than the supported type, e.g., to v4f64,
12889 if (VecFloatVT != Op->getSimpleValueType(0))
12892 unsigned NumElts = VecIntVT.getVectorNumElements();
12893 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12894 "Unsupported custom type");
12895 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12897 // In the #idef/#else code, we have in common:
12898 // - The vector of constants:
12904 // Create the splat vector for 0x4b000000.
12905 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12906 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12907 CstLow, CstLow, CstLow, CstLow};
12908 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12909 makeArrayRef(&CstLowArray[0], NumElts));
12910 // Create the splat vector for 0x53000000.
12911 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12912 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12913 CstHigh, CstHigh, CstHigh, CstHigh};
12914 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12915 makeArrayRef(&CstHighArray[0], NumElts));
12917 // Create the right shift.
12918 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12919 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12920 CstShift, CstShift, CstShift, CstShift};
12921 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12922 makeArrayRef(&CstShiftArray[0], NumElts));
12923 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12926 if (Subtarget.hasSSE41()) {
12927 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12928 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12929 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12930 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12931 // Low will be bitcasted right away, so do not bother bitcasting back to its
12933 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12934 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12935 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12936 // (uint4) 0x53000000, 0xaa);
12937 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12938 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12939 // High will be bitcasted right away, so do not bother bitcasting back to
12940 // its original type.
12941 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12942 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12944 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12945 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12946 CstMask, CstMask, CstMask);
12947 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12948 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12949 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12951 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12952 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12955 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12956 SDValue CstFAdd = DAG.getConstantFP(
12957 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12958 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12959 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12960 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12961 makeArrayRef(&CstFAddArray[0], NumElts));
12963 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12964 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12965 // TODO: Are there any fast-math-flags to propagate here?
12967 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12968 // return (float4) lo + fhi;
12969 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12970 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12973 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12974 SelectionDAG &DAG) const {
12975 SDValue N0 = Op.getOperand(0);
12976 MVT SVT = N0.getSimpleValueType();
12979 switch (SVT.SimpleTy) {
12981 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12986 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12987 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12988 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12992 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12995 assert(Subtarget->hasAVX512());
12996 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12997 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
13001 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
13002 SelectionDAG &DAG) const {
13003 SDValue N0 = Op.getOperand(0);
13005 auto PtrVT = getPointerTy(DAG.getDataLayout());
13007 if (Op.getSimpleValueType().isVector())
13008 return lowerUINT_TO_FP_vec(Op, DAG);
13010 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
13011 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
13012 // the optimization here.
13013 if (DAG.SignBitIsZero(N0))
13014 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
13016 MVT SrcVT = N0.getSimpleValueType();
13017 MVT DstVT = Op.getSimpleValueType();
13019 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
13020 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
13021 // Conversions from unsigned i32 to f32/f64 are legal,
13022 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
13026 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
13027 return LowerUINT_TO_FP_i64(Op, DAG);
13028 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
13029 return LowerUINT_TO_FP_i32(Op, DAG);
13030 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
13033 // Make a 64-bit buffer, and use it to build an FILD.
13034 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
13035 if (SrcVT == MVT::i32) {
13036 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
13037 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
13038 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13039 StackSlot, MachinePointerInfo(),
13041 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
13042 OffsetSlot, MachinePointerInfo(),
13044 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
13048 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
13049 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13050 StackSlot, MachinePointerInfo(),
13052 // For i64 source, we need to add the appropriate power of 2 if the input
13053 // was negative. This is the same as the optimization in
13054 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
13055 // we must be careful to do the computation in x87 extended precision, not
13056 // in SSE. (The generic code can't know it's OK to do this, or how to.)
13057 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
13058 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
13059 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
13060 MachineMemOperand::MOLoad, 8, 8);
13062 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
13063 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
13064 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
13067 APInt FF(32, 0x5F800000ULL);
13069 // Check whether the sign bit is set.
13070 SDValue SignSet = DAG.getSetCC(
13071 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
13072 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
13074 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
13075 SDValue FudgePtr = DAG.getConstantPool(
13076 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
13078 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
13079 SDValue Zero = DAG.getIntPtrConstant(0, dl);
13080 SDValue Four = DAG.getIntPtrConstant(4, dl);
13081 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
13083 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
13085 // Load the value out, extending it from f32 to f80.
13086 // FIXME: Avoid the extend by constructing the right constant pool?
13087 SDValue Fudge = DAG.getExtLoad(
13088 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
13089 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
13090 false, false, false, 4);
13091 // Extend everything to 80 bits to force it to be done on x87.
13092 // TODO: Are there any fast-math-flags to propagate here?
13093 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
13094 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
13095 DAG.getIntPtrConstant(0, dl));
13098 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
13099 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
13100 // just return an <SDValue(), SDValue()> pair.
13101 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
13102 // to i16, i32 or i64, and we lower it to a legal sequence.
13103 // If lowered to the final integer result we return a <result, SDValue()> pair.
13104 // Otherwise we lower it to a sequence ending with a FIST, return a
13105 // <FIST, StackSlot> pair, and the caller is responsible for loading
13106 // the final integer result from StackSlot.
13107 std::pair<SDValue,SDValue>
13108 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
13109 bool IsSigned, bool IsReplace) const {
13112 EVT DstTy = Op.getValueType();
13113 EVT TheVT = Op.getOperand(0).getValueType();
13114 auto PtrVT = getPointerTy(DAG.getDataLayout());
13116 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
13117 // f16 must be promoted before using the lowering in this routine.
13118 // fp128 does not use this lowering.
13119 return std::make_pair(SDValue(), SDValue());
13122 // If using FIST to compute an unsigned i64, we'll need some fixup
13123 // to handle values above the maximum signed i64. A FIST is always
13124 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
13125 bool UnsignedFixup = !IsSigned &&
13126 DstTy == MVT::i64 &&
13127 (!Subtarget->is64Bit() ||
13128 !isScalarFPTypeInSSEReg(TheVT));
13130 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
13131 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
13132 // The low 32 bits of the fist result will have the correct uint32 result.
13133 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
13137 assert(DstTy.getSimpleVT() <= MVT::i64 &&
13138 DstTy.getSimpleVT() >= MVT::i16 &&
13139 "Unknown FP_TO_INT to lower!");
13141 // These are really Legal.
13142 if (DstTy == MVT::i32 &&
13143 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13144 return std::make_pair(SDValue(), SDValue());
13145 if (Subtarget->is64Bit() &&
13146 DstTy == MVT::i64 &&
13147 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13148 return std::make_pair(SDValue(), SDValue());
13150 // We lower FP->int64 into FISTP64 followed by a load from a temporary
13152 MachineFunction &MF = DAG.getMachineFunction();
13153 unsigned MemSize = DstTy.getSizeInBits()/8;
13154 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13155 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13158 switch (DstTy.getSimpleVT().SimpleTy) {
13159 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
13160 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
13161 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
13162 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
13165 SDValue Chain = DAG.getEntryNode();
13166 SDValue Value = Op.getOperand(0);
13167 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
13169 if (UnsignedFixup) {
13171 // Conversion to unsigned i64 is implemented with a select,
13172 // depending on whether the source value fits in the range
13173 // of a signed i64. Let Thresh be the FP equivalent of
13174 // 0x8000000000000000ULL.
13176 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
13177 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
13178 // Fist-to-mem64 FistSrc
13179 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
13180 // to XOR'ing the high 32 bits with Adjust.
13182 // Being a power of 2, Thresh is exactly representable in all FP formats.
13183 // For X87 we'd like to use the smallest FP type for this constant, but
13184 // for DAG type consistency we have to match the FP operand type.
13186 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
13187 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
13188 bool LosesInfo = false;
13189 if (TheVT == MVT::f64)
13190 // The rounding mode is irrelevant as the conversion should be exact.
13191 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
13193 else if (TheVT == MVT::f80)
13194 Status = Thresh.convert(APFloat::x87DoubleExtended,
13195 APFloat::rmNearestTiesToEven, &LosesInfo);
13197 assert(Status == APFloat::opOK && !LosesInfo &&
13198 "FP conversion should have been exact");
13200 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
13202 SDValue Cmp = DAG.getSetCC(DL,
13203 getSetCCResultType(DAG.getDataLayout(),
13204 *DAG.getContext(), TheVT),
13205 Value, ThreshVal, ISD::SETLT);
13206 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
13207 DAG.getConstant(0, DL, MVT::i32),
13208 DAG.getConstant(0x80000000, DL, MVT::i32));
13209 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
13210 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
13211 *DAG.getContext(), TheVT),
13212 Value, ThreshVal, ISD::SETLT);
13213 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
13216 // FIXME This causes a redundant load/store if the SSE-class value is already
13217 // in memory, such as if it is on the callstack.
13218 if (isScalarFPTypeInSSEReg(TheVT)) {
13219 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13220 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13221 MachinePointerInfo::getFixedStack(MF, SSFI), false,
13223 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13225 Chain, StackSlot, DAG.getValueType(TheVT)
13228 MachineMemOperand *MMO =
13229 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13230 MachineMemOperand::MOLoad, MemSize, MemSize);
13231 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13232 Chain = Value.getValue(1);
13233 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13234 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13237 MachineMemOperand *MMO =
13238 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13239 MachineMemOperand::MOStore, MemSize, MemSize);
13241 if (UnsignedFixup) {
13243 // Insert the FIST, load its result as two i32's,
13244 // and XOR the high i32 with Adjust.
13246 SDValue FistOps[] = { Chain, Value, StackSlot };
13247 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13248 FistOps, DstTy, MMO);
13250 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
13251 MachinePointerInfo(),
13252 false, false, false, 0);
13253 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
13254 DAG.getConstant(4, DL, PtrVT));
13256 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
13257 MachinePointerInfo(),
13258 false, false, false, 0);
13259 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13261 if (Subtarget->is64Bit()) {
13262 // Join High32 and Low32 into a 64-bit result.
13263 // (High32 << 32) | Low32
13264 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13265 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13266 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13267 DAG.getConstant(32, DL, MVT::i8));
13268 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13269 return std::make_pair(Result, SDValue());
13272 SDValue ResultOps[] = { Low32, High32 };
13274 SDValue pair = IsReplace
13275 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13276 : DAG.getMergeValues(ResultOps, DL);
13277 return std::make_pair(pair, SDValue());
13279 // Build the FP_TO_INT*_IN_MEM
13280 SDValue Ops[] = { Chain, Value, StackSlot };
13281 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13283 return std::make_pair(FIST, StackSlot);
13287 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13288 const X86Subtarget *Subtarget) {
13289 MVT VT = Op->getSimpleValueType(0);
13290 SDValue In = Op->getOperand(0);
13291 MVT InVT = In.getSimpleValueType();
13294 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13295 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13297 // Optimize vectors in AVX mode:
13300 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13301 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13302 // Concat upper and lower parts.
13305 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13306 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13307 // Concat upper and lower parts.
13310 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13311 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13312 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13315 if (Subtarget->hasInt256())
13316 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13318 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13319 SDValue Undef = DAG.getUNDEF(InVT);
13320 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13321 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13322 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13324 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13325 VT.getVectorNumElements()/2);
13327 OpLo = DAG.getBitcast(HVT, OpLo);
13328 OpHi = DAG.getBitcast(HVT, OpHi);
13330 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13333 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13334 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13335 MVT VT = Op->getSimpleValueType(0);
13336 SDValue In = Op->getOperand(0);
13337 MVT InVT = In.getSimpleValueType();
13339 unsigned int NumElts = VT.getVectorNumElements();
13340 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13343 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13344 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13346 assert(InVT.getVectorElementType() == MVT::i1);
13347 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13349 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13351 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13353 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13354 if (VT.is512BitVector())
13356 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13359 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13360 SelectionDAG &DAG) {
13361 if (Subtarget->hasFp256())
13362 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13368 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13369 SelectionDAG &DAG) {
13371 MVT VT = Op.getSimpleValueType();
13372 SDValue In = Op.getOperand(0);
13373 MVT SVT = In.getSimpleValueType();
13375 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13376 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13378 if (Subtarget->hasFp256())
13379 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13382 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13383 VT.getVectorNumElements() != SVT.getVectorNumElements());
13387 static SDValue LowerTruncateVecI1(SDValue Op, SelectionDAG &DAG,
13388 const X86Subtarget *Subtarget) {
13391 MVT VT = Op.getSimpleValueType();
13392 SDValue In = Op.getOperand(0);
13393 MVT InVT = In.getSimpleValueType();
13395 assert(VT.getVectorElementType() == MVT::i1 && "Unexected vector type.");
13397 // Shift LSB to MSB and use VPMOVB2M - SKX.
13398 unsigned ShiftInx = InVT.getScalarSizeInBits() - 1;
13399 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13400 Subtarget->hasBWI()) || // legal, will go to VPMOVB2M, VPMOVW2M
13401 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13402 InVT.getScalarSizeInBits() <= 16 && Subtarget->hasBWI() &&
13403 Subtarget->hasVLX())) { // legal, will go to VPMOVB2M, VPMOVW2M
13404 // Shift packed bytes not supported natively, bitcast to dword
13405 MVT ExtVT = MVT::getVectorVT(MVT::i16, InVT.getSizeInBits()/16);
13406 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, ExtVT,
13407 DAG.getBitcast(ExtVT, In),
13408 DAG.getConstant(ShiftInx, DL, ExtVT));
13409 ShiftNode = DAG.getBitcast(InVT, ShiftNode);
13410 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13412 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13413 Subtarget->hasDQI()) || // legal, will go to VPMOVD2M, VPMOVQ2M
13414 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13415 InVT.getScalarSizeInBits() >= 32 && Subtarget->hasDQI() &&
13416 Subtarget->hasVLX())) { // legal, will go to VPMOVD2M, VPMOVQ2M
13418 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13419 DAG.getConstant(ShiftInx, DL, InVT));
13420 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13423 // Shift LSB to MSB, extend if necessary and use TESTM.
13424 unsigned NumElts = InVT.getVectorNumElements();
13425 if (InVT.getSizeInBits() < 512 &&
13426 (InVT.getScalarType() == MVT::i8 || InVT.getScalarType() == MVT::i16 ||
13427 !Subtarget->hasVLX())) {
13428 assert((NumElts == 8 || NumElts == 16) && "Unexected vector type.");
13430 // TESTD/Q should be used (if BW supported we use CVT2MASK above),
13431 // so vector should be extended to packed dword/qword.
13432 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts);
13433 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13435 ShiftInx = InVT.getScalarSizeInBits() - 1;
13438 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13439 DAG.getConstant(ShiftInx, DL, InVT));
13440 return DAG.getNode(X86ISD::TESTM, DL, VT, ShiftNode, ShiftNode);
13443 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13445 MVT VT = Op.getSimpleValueType();
13446 SDValue In = Op.getOperand(0);
13447 MVT InVT = In.getSimpleValueType();
13449 if (VT == MVT::i1) {
13450 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13451 "Invalid scalar TRUNCATE operation");
13452 if (InVT.getSizeInBits() >= 32)
13454 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13455 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13457 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13458 "Invalid TRUNCATE operation");
13460 if (VT.getVectorElementType() == MVT::i1)
13461 return LowerTruncateVecI1(Op, DAG, Subtarget);
13463 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13464 if (Subtarget->hasAVX512()) {
13465 // word to byte only under BWI
13466 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13467 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13468 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13469 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13471 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13472 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13473 if (Subtarget->hasInt256()) {
13474 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13475 In = DAG.getBitcast(MVT::v8i32, In);
13476 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13478 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13479 DAG.getIntPtrConstant(0, DL));
13482 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13483 DAG.getIntPtrConstant(0, DL));
13484 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13485 DAG.getIntPtrConstant(2, DL));
13486 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13487 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13488 static const int ShufMask[] = {0, 2, 4, 6};
13489 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13492 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13493 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13494 if (Subtarget->hasInt256()) {
13495 In = DAG.getBitcast(MVT::v32i8, In);
13497 SmallVector<SDValue,32> pshufbMask;
13498 for (unsigned i = 0; i < 2; ++i) {
13499 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13500 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13501 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13502 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13503 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13504 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13505 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13506 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13507 for (unsigned j = 0; j < 8; ++j)
13508 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13510 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13511 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13512 In = DAG.getBitcast(MVT::v4i64, In);
13514 static const int ShufMask[] = {0, 2, -1, -1};
13515 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13517 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13518 DAG.getIntPtrConstant(0, DL));
13519 return DAG.getBitcast(VT, In);
13522 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13523 DAG.getIntPtrConstant(0, DL));
13525 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13526 DAG.getIntPtrConstant(4, DL));
13528 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13529 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13531 // The PSHUFB mask:
13532 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13533 -1, -1, -1, -1, -1, -1, -1, -1};
13535 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13536 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13537 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13539 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13540 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13542 // The MOVLHPS Mask:
13543 static const int ShufMask2[] = {0, 1, 4, 5};
13544 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13545 return DAG.getBitcast(MVT::v8i16, res);
13548 // Handle truncation of V256 to V128 using shuffles.
13549 if (!VT.is128BitVector() || !InVT.is256BitVector())
13552 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13554 unsigned NumElems = VT.getVectorNumElements();
13555 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13557 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13558 // Prepare truncation shuffle mask
13559 for (unsigned i = 0; i != NumElems; ++i)
13560 MaskVec[i] = i * 2;
13561 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13562 DAG.getUNDEF(NVT), &MaskVec[0]);
13563 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13564 DAG.getIntPtrConstant(0, DL));
13567 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13568 SelectionDAG &DAG) const {
13569 assert(!Op.getSimpleValueType().isVector());
13571 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13572 /*IsSigned=*/ true, /*IsReplace=*/ false);
13573 SDValue FIST = Vals.first, StackSlot = Vals.second;
13574 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13575 if (!FIST.getNode())
13578 if (StackSlot.getNode())
13579 // Load the result.
13580 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13581 FIST, StackSlot, MachinePointerInfo(),
13582 false, false, false, 0);
13584 // The node is the result.
13588 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13589 SelectionDAG &DAG) const {
13590 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13591 /*IsSigned=*/ false, /*IsReplace=*/ false);
13592 SDValue FIST = Vals.first, StackSlot = Vals.second;
13593 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13594 if (!FIST.getNode())
13597 if (StackSlot.getNode())
13598 // Load the result.
13599 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13600 FIST, StackSlot, MachinePointerInfo(),
13601 false, false, false, 0);
13603 // The node is the result.
13607 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13609 MVT VT = Op.getSimpleValueType();
13610 SDValue In = Op.getOperand(0);
13611 MVT SVT = In.getSimpleValueType();
13613 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13615 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13616 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13617 In, DAG.getUNDEF(SVT)));
13620 /// The only differences between FABS and FNEG are the mask and the logic op.
13621 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13622 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13623 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13624 "Wrong opcode for lowering FABS or FNEG.");
13626 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13628 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13629 // into an FNABS. We'll lower the FABS after that if it is still in use.
13631 for (SDNode *User : Op->uses())
13632 if (User->getOpcode() == ISD::FNEG)
13636 MVT VT = Op.getSimpleValueType();
13638 bool IsF128 = (VT == MVT::f128);
13640 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13641 // decide if we should generate a 16-byte constant mask when we only need 4 or
13642 // 8 bytes for the scalar case.
13648 if (VT.isVector()) {
13650 EltVT = VT.getVectorElementType();
13651 NumElts = VT.getVectorNumElements();
13652 } else if (IsF128) {
13653 // SSE instructions are used for optimized f128 logical operations.
13654 LogicVT = MVT::f128;
13658 // There are no scalar bitwise logical SSE/AVX instructions, so we
13659 // generate a 16-byte vector constant and logic op even for the scalar case.
13660 // Using a 16-byte mask allows folding the load of the mask with
13661 // the logic op, so it can save (~4 bytes) on code size.
13662 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13664 NumElts = (VT == MVT::f64) ? 2 : 4;
13667 unsigned EltBits = EltVT.getSizeInBits();
13668 LLVMContext *Context = DAG.getContext();
13669 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13671 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13672 Constant *C = ConstantInt::get(*Context, MaskElt);
13673 C = ConstantVector::getSplat(NumElts, C);
13674 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13675 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13676 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13678 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13679 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13680 false, false, false, Alignment);
13682 SDValue Op0 = Op.getOperand(0);
13683 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13685 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13686 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13688 if (VT.isVector() || IsF128)
13689 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13691 // For the scalar case extend to a 128-bit vector, perform the logic op,
13692 // and extract the scalar result back out.
13693 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13694 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13695 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13696 DAG.getIntPtrConstant(0, dl));
13699 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13700 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13701 LLVMContext *Context = DAG.getContext();
13702 SDValue Op0 = Op.getOperand(0);
13703 SDValue Op1 = Op.getOperand(1);
13705 MVT VT = Op.getSimpleValueType();
13706 MVT SrcVT = Op1.getSimpleValueType();
13707 bool IsF128 = (VT == MVT::f128);
13709 // If second operand is smaller, extend it first.
13710 if (SrcVT.bitsLT(VT)) {
13711 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13714 // And if it is bigger, shrink it first.
13715 if (SrcVT.bitsGT(VT)) {
13716 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13720 // At this point the operands and the result should have the same
13721 // type, and that won't be f80 since that is not custom lowered.
13722 assert((VT == MVT::f64 || VT == MVT::f32 || IsF128) &&
13723 "Unexpected type in LowerFCOPYSIGN");
13725 const fltSemantics &Sem =
13726 VT == MVT::f64 ? APFloat::IEEEdouble :
13727 (IsF128 ? APFloat::IEEEquad : APFloat::IEEEsingle);
13728 const unsigned SizeInBits = VT.getSizeInBits();
13730 SmallVector<Constant *, 4> CV(
13731 VT == MVT::f64 ? 2 : (IsF128 ? 1 : 4),
13732 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13734 // First, clear all bits but the sign bit from the second operand (sign).
13735 CV[0] = ConstantFP::get(*Context,
13736 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13737 Constant *C = ConstantVector::get(CV);
13738 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13739 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13741 // Perform all logic operations as 16-byte vectors because there are no
13742 // scalar FP logic instructions in SSE. This allows load folding of the
13743 // constants into the logic instructions.
13744 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : (IsF128 ? MVT::f128 : MVT::v4f32);
13746 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13747 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13748 false, false, false, 16);
13750 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13751 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13753 // Next, clear the sign bit from the first operand (magnitude).
13754 // If it's a constant, we can clear it here.
13755 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13756 APFloat APF = Op0CN->getValueAPF();
13757 // If the magnitude is a positive zero, the sign bit alone is enough.
13758 if (APF.isPosZero())
13759 return IsF128 ? SignBit :
13760 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13761 DAG.getIntPtrConstant(0, dl));
13763 CV[0] = ConstantFP::get(*Context, APF);
13765 CV[0] = ConstantFP::get(
13767 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13769 C = ConstantVector::get(CV);
13770 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13772 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13773 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13774 false, false, false, 16);
13775 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13776 if (!isa<ConstantFPSDNode>(Op0)) {
13778 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13779 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13781 // OR the magnitude value with the sign bit.
13782 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13783 return IsF128 ? Val :
13784 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13785 DAG.getIntPtrConstant(0, dl));
13788 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13789 SDValue N0 = Op.getOperand(0);
13791 MVT VT = Op.getSimpleValueType();
13793 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13794 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13795 DAG.getConstant(1, dl, VT));
13796 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13799 // Check whether an OR'd tree is PTEST-able.
13800 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13801 SelectionDAG &DAG) {
13802 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13804 if (!Subtarget->hasSSE41())
13807 if (!Op->hasOneUse())
13810 SDNode *N = Op.getNode();
13813 SmallVector<SDValue, 8> Opnds;
13814 DenseMap<SDValue, unsigned> VecInMap;
13815 SmallVector<SDValue, 8> VecIns;
13816 EVT VT = MVT::Other;
13818 // Recognize a special case where a vector is casted into wide integer to
13820 Opnds.push_back(N->getOperand(0));
13821 Opnds.push_back(N->getOperand(1));
13823 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13824 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13825 // BFS traverse all OR'd operands.
13826 if (I->getOpcode() == ISD::OR) {
13827 Opnds.push_back(I->getOperand(0));
13828 Opnds.push_back(I->getOperand(1));
13829 // Re-evaluate the number of nodes to be traversed.
13830 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13834 // Quit if a non-EXTRACT_VECTOR_ELT
13835 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13838 // Quit if without a constant index.
13839 SDValue Idx = I->getOperand(1);
13840 if (!isa<ConstantSDNode>(Idx))
13843 SDValue ExtractedFromVec = I->getOperand(0);
13844 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13845 if (M == VecInMap.end()) {
13846 VT = ExtractedFromVec.getValueType();
13847 // Quit if not 128/256-bit vector.
13848 if (!VT.is128BitVector() && !VT.is256BitVector())
13850 // Quit if not the same type.
13851 if (VecInMap.begin() != VecInMap.end() &&
13852 VT != VecInMap.begin()->first.getValueType())
13854 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13855 VecIns.push_back(ExtractedFromVec);
13857 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13860 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13861 "Not extracted from 128-/256-bit vector.");
13863 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13865 for (DenseMap<SDValue, unsigned>::const_iterator
13866 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13867 // Quit if not all elements are used.
13868 if (I->second != FullMask)
13872 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13874 // Cast all vectors into TestVT for PTEST.
13875 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13876 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13878 // If more than one full vectors are evaluated, OR them first before PTEST.
13879 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13880 // Each iteration will OR 2 nodes and append the result until there is only
13881 // 1 node left, i.e. the final OR'd value of all vectors.
13882 SDValue LHS = VecIns[Slot];
13883 SDValue RHS = VecIns[Slot + 1];
13884 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13887 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13888 VecIns.back(), VecIns.back());
13891 /// \brief return true if \c Op has a use that doesn't just read flags.
13892 static bool hasNonFlagsUse(SDValue Op) {
13893 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13895 SDNode *User = *UI;
13896 unsigned UOpNo = UI.getOperandNo();
13897 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13898 // Look pass truncate.
13899 UOpNo = User->use_begin().getOperandNo();
13900 User = *User->use_begin();
13903 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13904 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13910 /// Emit nodes that will be selected as "test Op0,Op0", or something
13912 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13913 SelectionDAG &DAG) const {
13914 if (Op.getValueType() == MVT::i1) {
13915 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13916 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13917 DAG.getConstant(0, dl, MVT::i8));
13919 // CF and OF aren't always set the way we want. Determine which
13920 // of these we need.
13921 bool NeedCF = false;
13922 bool NeedOF = false;
13925 case X86::COND_A: case X86::COND_AE:
13926 case X86::COND_B: case X86::COND_BE:
13929 case X86::COND_G: case X86::COND_GE:
13930 case X86::COND_L: case X86::COND_LE:
13931 case X86::COND_O: case X86::COND_NO: {
13932 // Check if we really need to set the
13933 // Overflow flag. If NoSignedWrap is present
13934 // that is not actually needed.
13935 switch (Op->getOpcode()) {
13940 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13941 if (BinNode->Flags.hasNoSignedWrap())
13951 // See if we can use the EFLAGS value from the operand instead of
13952 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13953 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13954 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13955 // Emit a CMP with 0, which is the TEST pattern.
13956 //if (Op.getValueType() == MVT::i1)
13957 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13958 // DAG.getConstant(0, MVT::i1));
13959 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13960 DAG.getConstant(0, dl, Op.getValueType()));
13962 unsigned Opcode = 0;
13963 unsigned NumOperands = 0;
13965 // Truncate operations may prevent the merge of the SETCC instruction
13966 // and the arithmetic instruction before it. Attempt to truncate the operands
13967 // of the arithmetic instruction and use a reduced bit-width instruction.
13968 bool NeedTruncation = false;
13969 SDValue ArithOp = Op;
13970 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13971 SDValue Arith = Op->getOperand(0);
13972 // Both the trunc and the arithmetic op need to have one user each.
13973 if (Arith->hasOneUse())
13974 switch (Arith.getOpcode()) {
13981 NeedTruncation = true;
13987 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13988 // which may be the result of a CAST. We use the variable 'Op', which is the
13989 // non-casted variable when we check for possible users.
13990 switch (ArithOp.getOpcode()) {
13992 // Due to an isel shortcoming, be conservative if this add is likely to be
13993 // selected as part of a load-modify-store instruction. When the root node
13994 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13995 // uses of other nodes in the match, such as the ADD in this case. This
13996 // leads to the ADD being left around and reselected, with the result being
13997 // two adds in the output. Alas, even if none our users are stores, that
13998 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13999 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
14000 // climbing the DAG back to the root, and it doesn't seem to be worth the
14002 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14003 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14004 if (UI->getOpcode() != ISD::CopyToReg &&
14005 UI->getOpcode() != ISD::SETCC &&
14006 UI->getOpcode() != ISD::STORE)
14009 if (ConstantSDNode *C =
14010 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
14011 // An add of one will be selected as an INC.
14012 if (C->isOne() && !Subtarget->slowIncDec()) {
14013 Opcode = X86ISD::INC;
14018 // An add of negative one (subtract of one) will be selected as a DEC.
14019 if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
14020 Opcode = X86ISD::DEC;
14026 // Otherwise use a regular EFLAGS-setting add.
14027 Opcode = X86ISD::ADD;
14032 // If we have a constant logical shift that's only used in a comparison
14033 // against zero turn it into an equivalent AND. This allows turning it into
14034 // a TEST instruction later.
14035 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
14036 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
14037 EVT VT = Op.getValueType();
14038 unsigned BitWidth = VT.getSizeInBits();
14039 unsigned ShAmt = Op->getConstantOperandVal(1);
14040 if (ShAmt >= BitWidth) // Avoid undefined shifts.
14042 APInt Mask = ArithOp.getOpcode() == ISD::SRL
14043 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
14044 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
14045 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
14047 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
14048 DAG.getConstant(Mask, dl, VT));
14049 DAG.ReplaceAllUsesWith(Op, New);
14055 // If the primary and result isn't used, don't bother using X86ISD::AND,
14056 // because a TEST instruction will be better.
14057 if (!hasNonFlagsUse(Op))
14063 // Due to the ISEL shortcoming noted above, be conservative if this op is
14064 // likely to be selected as part of a load-modify-store instruction.
14065 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14066 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14067 if (UI->getOpcode() == ISD::STORE)
14070 // Otherwise use a regular EFLAGS-setting instruction.
14071 switch (ArithOp.getOpcode()) {
14072 default: llvm_unreachable("unexpected operator!");
14073 case ISD::SUB: Opcode = X86ISD::SUB; break;
14074 case ISD::XOR: Opcode = X86ISD::XOR; break;
14075 case ISD::AND: Opcode = X86ISD::AND; break;
14077 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
14078 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
14079 if (EFLAGS.getNode())
14082 Opcode = X86ISD::OR;
14096 return SDValue(Op.getNode(), 1);
14102 // If we found that truncation is beneficial, perform the truncation and
14104 if (NeedTruncation) {
14105 EVT VT = Op.getValueType();
14106 SDValue WideVal = Op->getOperand(0);
14107 EVT WideVT = WideVal.getValueType();
14108 unsigned ConvertedOp = 0;
14109 // Use a target machine opcode to prevent further DAGCombine
14110 // optimizations that may separate the arithmetic operations
14111 // from the setcc node.
14112 switch (WideVal.getOpcode()) {
14114 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
14115 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
14116 case ISD::AND: ConvertedOp = X86ISD::AND; break;
14117 case ISD::OR: ConvertedOp = X86ISD::OR; break;
14118 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
14122 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14123 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
14124 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
14125 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
14126 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
14132 // Emit a CMP with 0, which is the TEST pattern.
14133 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14134 DAG.getConstant(0, dl, Op.getValueType()));
14136 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14137 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
14139 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
14140 DAG.ReplaceAllUsesWith(Op, New);
14141 return SDValue(New.getNode(), 1);
14144 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
14146 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
14147 SDLoc dl, SelectionDAG &DAG) const {
14148 if (isNullConstant(Op1))
14149 return EmitTest(Op0, X86CC, dl, DAG);
14151 assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
14152 "Unexpected comparison operation for MVT::i1 operands");
14154 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
14155 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
14156 // Do the comparison at i32 if it's smaller, besides the Atom case.
14157 // This avoids subregister aliasing issues. Keep the smaller reference
14158 // if we're optimizing for size, however, as that'll allow better folding
14159 // of memory operations.
14160 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
14161 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
14162 !Subtarget->isAtom()) {
14163 unsigned ExtendOp =
14164 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
14165 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
14166 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
14168 // Use SUB instead of CMP to enable CSE between SUB and CMP.
14169 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
14170 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
14172 return SDValue(Sub.getNode(), 1);
14174 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
14177 /// Convert a comparison if required by the subtarget.
14178 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
14179 SelectionDAG &DAG) const {
14180 // If the subtarget does not support the FUCOMI instruction, floating-point
14181 // comparisons have to be converted.
14182 if (Subtarget->hasCMov() ||
14183 Cmp.getOpcode() != X86ISD::CMP ||
14184 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
14185 !Cmp.getOperand(1).getValueType().isFloatingPoint())
14188 // The instruction selector will select an FUCOM instruction instead of
14189 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
14190 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
14191 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
14193 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
14194 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
14195 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
14196 DAG.getConstant(8, dl, MVT::i8));
14197 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
14199 // Some 64-bit targets lack SAHF support, but they do support FCOMI.
14200 assert(Subtarget->hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?");
14201 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
14204 /// The minimum architected relative accuracy is 2^-12. We need one
14205 /// Newton-Raphson step to have a good float result (24 bits of precision).
14206 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
14207 DAGCombinerInfo &DCI,
14208 unsigned &RefinementSteps,
14209 bool &UseOneConstNR) const {
14210 EVT VT = Op.getValueType();
14211 const char *RecipOp;
14213 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
14214 // TODO: Add support for AVX512 (v16f32).
14215 // It is likely not profitable to do this for f64 because a double-precision
14216 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
14217 // instructions: convert to single, rsqrtss, convert back to double, refine
14218 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
14219 // along with FMA, this could be a throughput win.
14220 if (VT == MVT::f32 && Subtarget->hasSSE1())
14222 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14223 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14224 RecipOp = "vec-sqrtf";
14228 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14229 if (!Recips.isEnabled(RecipOp))
14232 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14233 UseOneConstNR = false;
14234 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
14237 /// The minimum architected relative accuracy is 2^-12. We need one
14238 /// Newton-Raphson step to have a good float result (24 bits of precision).
14239 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
14240 DAGCombinerInfo &DCI,
14241 unsigned &RefinementSteps) const {
14242 EVT VT = Op.getValueType();
14243 const char *RecipOp;
14245 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
14246 // TODO: Add support for AVX512 (v16f32).
14247 // It is likely not profitable to do this for f64 because a double-precision
14248 // reciprocal estimate with refinement on x86 prior to FMA requires
14249 // 15 instructions: convert to single, rcpss, convert back to double, refine
14250 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
14251 // along with FMA, this could be a throughput win.
14252 if (VT == MVT::f32 && Subtarget->hasSSE1())
14254 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14255 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14256 RecipOp = "vec-divf";
14260 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14261 if (!Recips.isEnabled(RecipOp))
14264 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14265 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
14268 /// If we have at least two divisions that use the same divisor, convert to
14269 /// multplication by a reciprocal. This may need to be adjusted for a given
14270 /// CPU if a division's cost is not at least twice the cost of a multiplication.
14271 /// This is because we still need one division to calculate the reciprocal and
14272 /// then we need two multiplies by that reciprocal as replacements for the
14273 /// original divisions.
14274 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
14278 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
14279 /// if it's possible.
14280 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
14281 SDLoc dl, SelectionDAG &DAG) const {
14282 SDValue Op0 = And.getOperand(0);
14283 SDValue Op1 = And.getOperand(1);
14284 if (Op0.getOpcode() == ISD::TRUNCATE)
14285 Op0 = Op0.getOperand(0);
14286 if (Op1.getOpcode() == ISD::TRUNCATE)
14287 Op1 = Op1.getOperand(0);
14290 if (Op1.getOpcode() == ISD::SHL)
14291 std::swap(Op0, Op1);
14292 if (Op0.getOpcode() == ISD::SHL) {
14293 if (isOneConstant(Op0.getOperand(0))) {
14294 // If we looked past a truncate, check that it's only truncating away
14296 unsigned BitWidth = Op0.getValueSizeInBits();
14297 unsigned AndBitWidth = And.getValueSizeInBits();
14298 if (BitWidth > AndBitWidth) {
14300 DAG.computeKnownBits(Op0, Zeros, Ones);
14301 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14305 RHS = Op0.getOperand(1);
14307 } else if (Op1.getOpcode() == ISD::Constant) {
14308 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14309 uint64_t AndRHSVal = AndRHS->getZExtValue();
14310 SDValue AndLHS = Op0;
14312 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14313 LHS = AndLHS.getOperand(0);
14314 RHS = AndLHS.getOperand(1);
14317 // Use BT if the immediate can't be encoded in a TEST instruction.
14318 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14320 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14324 if (LHS.getNode()) {
14325 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14326 // instruction. Since the shift amount is in-range-or-undefined, we know
14327 // that doing a bittest on the i32 value is ok. We extend to i32 because
14328 // the encoding for the i16 version is larger than the i32 version.
14329 // Also promote i16 to i32 for performance / code size reason.
14330 if (LHS.getValueType() == MVT::i8 ||
14331 LHS.getValueType() == MVT::i16)
14332 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14334 // If the operand types disagree, extend the shift amount to match. Since
14335 // BT ignores high bits (like shifts) we can use anyextend.
14336 if (LHS.getValueType() != RHS.getValueType())
14337 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14339 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14340 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14341 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14342 DAG.getConstant(Cond, dl, MVT::i8), BT);
14348 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14350 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14355 // SSE Condition code mapping:
14364 switch (SetCCOpcode) {
14365 default: llvm_unreachable("Unexpected SETCC condition");
14367 case ISD::SETEQ: SSECC = 0; break;
14369 case ISD::SETGT: Swap = true; // Fallthrough
14371 case ISD::SETOLT: SSECC = 1; break;
14373 case ISD::SETGE: Swap = true; // Fallthrough
14375 case ISD::SETOLE: SSECC = 2; break;
14376 case ISD::SETUO: SSECC = 3; break;
14378 case ISD::SETNE: SSECC = 4; break;
14379 case ISD::SETULE: Swap = true; // Fallthrough
14380 case ISD::SETUGE: SSECC = 5; break;
14381 case ISD::SETULT: Swap = true; // Fallthrough
14382 case ISD::SETUGT: SSECC = 6; break;
14383 case ISD::SETO: SSECC = 7; break;
14385 case ISD::SETONE: SSECC = 8; break;
14388 std::swap(Op0, Op1);
14393 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14394 // ones, and then concatenate the result back.
14395 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14396 MVT VT = Op.getSimpleValueType();
14398 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14399 "Unsupported value type for operation");
14401 unsigned NumElems = VT.getVectorNumElements();
14403 SDValue CC = Op.getOperand(2);
14405 // Extract the LHS vectors
14406 SDValue LHS = Op.getOperand(0);
14407 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14408 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14410 // Extract the RHS vectors
14411 SDValue RHS = Op.getOperand(1);
14412 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14413 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14415 // Issue the operation on the smaller types and concatenate the result back
14416 MVT EltVT = VT.getVectorElementType();
14417 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14418 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14419 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14420 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14423 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14424 SDValue Op0 = Op.getOperand(0);
14425 SDValue Op1 = Op.getOperand(1);
14426 SDValue CC = Op.getOperand(2);
14427 MVT VT = Op.getSimpleValueType();
14430 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14431 "Unexpected type for boolean compare operation");
14432 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14433 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14434 DAG.getConstant(-1, dl, VT));
14435 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14436 DAG.getConstant(-1, dl, VT));
14437 switch (SetCCOpcode) {
14438 default: llvm_unreachable("Unexpected SETCC condition");
14440 // (x == y) -> ~(x ^ y)
14441 return DAG.getNode(ISD::XOR, dl, VT,
14442 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14443 DAG.getConstant(-1, dl, VT));
14445 // (x != y) -> (x ^ y)
14446 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14449 // (x > y) -> (x & ~y)
14450 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14453 // (x < y) -> (~x & y)
14454 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14457 // (x <= y) -> (~x | y)
14458 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14461 // (x >=y) -> (x | ~y)
14462 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14466 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14467 const X86Subtarget *Subtarget) {
14468 SDValue Op0 = Op.getOperand(0);
14469 SDValue Op1 = Op.getOperand(1);
14470 SDValue CC = Op.getOperand(2);
14471 MVT VT = Op.getSimpleValueType();
14474 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14475 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14476 "Cannot set masked compare for this operation");
14478 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14480 bool Unsigned = false;
14483 switch (SetCCOpcode) {
14484 default: llvm_unreachable("Unexpected SETCC condition");
14485 case ISD::SETNE: SSECC = 4; break;
14486 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14487 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14488 case ISD::SETLT: Swap = true; //fall-through
14489 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14490 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14491 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14492 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14493 case ISD::SETULE: Unsigned = true; //fall-through
14494 case ISD::SETLE: SSECC = 2; break;
14498 std::swap(Op0, Op1);
14500 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14501 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14502 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14503 DAG.getConstant(SSECC, dl, MVT::i8));
14506 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14507 /// operand \p Op1. If non-trivial (for example because it's not constant)
14508 /// return an empty value.
14509 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14511 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14515 MVT VT = Op1.getSimpleValueType();
14516 MVT EVT = VT.getVectorElementType();
14517 unsigned n = VT.getVectorNumElements();
14518 SmallVector<SDValue, 8> ULTOp1;
14520 for (unsigned i = 0; i < n; ++i) {
14521 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14522 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14525 // Avoid underflow.
14526 APInt Val = Elt->getAPIntValue();
14530 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14533 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14536 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14537 SelectionDAG &DAG) {
14538 SDValue Op0 = Op.getOperand(0);
14539 SDValue Op1 = Op.getOperand(1);
14540 SDValue CC = Op.getOperand(2);
14541 MVT VT = Op.getSimpleValueType();
14542 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14543 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14548 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14549 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14552 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14553 unsigned Opc = X86ISD::CMPP;
14554 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14555 assert(VT.getVectorNumElements() <= 16);
14556 Opc = X86ISD::CMPM;
14558 // In the two special cases we can't handle, emit two comparisons.
14561 unsigned CombineOpc;
14562 if (SetCCOpcode == ISD::SETUEQ) {
14563 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14565 assert(SetCCOpcode == ISD::SETONE);
14566 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14569 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14570 DAG.getConstant(CC0, dl, MVT::i8));
14571 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14572 DAG.getConstant(CC1, dl, MVT::i8));
14573 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14575 // Handle all other FP comparisons here.
14576 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14577 DAG.getConstant(SSECC, dl, MVT::i8));
14580 MVT VTOp0 = Op0.getSimpleValueType();
14581 assert(VTOp0 == Op1.getSimpleValueType() &&
14582 "Expected operands with same type!");
14583 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14584 "Invalid number of packed elements for source and destination!");
14586 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14587 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14588 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14589 // legalizer firstly checks if the first operand in input to the setcc has
14590 // a legal type. If so, then it promotes the return type to that same type.
14591 // Otherwise, the return type is promoted to the 'next legal type' which,
14592 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14594 // We reach this code only if the following two conditions are met:
14595 // 1. Both return type and operand type have been promoted to wider types
14596 // by the type legalizer.
14597 // 2. The original operand type has been promoted to a 256-bit vector.
14599 // Note that condition 2. only applies for AVX targets.
14600 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14601 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14604 // The non-AVX512 code below works under the assumption that source and
14605 // destination types are the same.
14606 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14607 "Value types for source and destination must be the same!");
14609 // Break 256-bit integer vector compare into smaller ones.
14610 if (VT.is256BitVector() && !Subtarget->hasInt256())
14611 return Lower256IntVSETCC(Op, DAG);
14613 MVT OpVT = Op1.getSimpleValueType();
14614 if (OpVT.getVectorElementType() == MVT::i1)
14615 return LowerBoolVSETCC_AVX512(Op, DAG);
14617 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14618 if (Subtarget->hasAVX512()) {
14619 if (Op1.getSimpleValueType().is512BitVector() ||
14620 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14621 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14622 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14624 // In AVX-512 architecture setcc returns mask with i1 elements,
14625 // But there is no compare instruction for i8 and i16 elements in KNL.
14626 // We are not talking about 512-bit operands in this case, these
14627 // types are illegal.
14629 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14630 OpVT.getVectorElementType().getSizeInBits() >= 8))
14631 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14632 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14635 // Lower using XOP integer comparisons.
14636 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14637 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14638 // Translate compare code to XOP PCOM compare mode.
14639 unsigned CmpMode = 0;
14640 switch (SetCCOpcode) {
14641 default: llvm_unreachable("Unexpected SETCC condition");
14643 case ISD::SETLT: CmpMode = 0x00; break;
14645 case ISD::SETLE: CmpMode = 0x01; break;
14647 case ISD::SETGT: CmpMode = 0x02; break;
14649 case ISD::SETGE: CmpMode = 0x03; break;
14650 case ISD::SETEQ: CmpMode = 0x04; break;
14651 case ISD::SETNE: CmpMode = 0x05; break;
14654 // Are we comparing unsigned or signed integers?
14655 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14656 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14658 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14659 DAG.getConstant(CmpMode, dl, MVT::i8));
14662 // We are handling one of the integer comparisons here. Since SSE only has
14663 // GT and EQ comparisons for integer, swapping operands and multiple
14664 // operations may be required for some comparisons.
14666 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14667 bool Subus = false;
14669 switch (SetCCOpcode) {
14670 default: llvm_unreachable("Unexpected SETCC condition");
14671 case ISD::SETNE: Invert = true;
14672 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14673 case ISD::SETLT: Swap = true;
14674 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14675 case ISD::SETGE: Swap = true;
14676 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14677 Invert = true; break;
14678 case ISD::SETULT: Swap = true;
14679 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14680 FlipSigns = true; break;
14681 case ISD::SETUGE: Swap = true;
14682 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14683 FlipSigns = true; Invert = true; break;
14686 // Special case: Use min/max operations for SETULE/SETUGE
14687 MVT VET = VT.getVectorElementType();
14689 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14690 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14693 switch (SetCCOpcode) {
14695 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14696 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14699 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14702 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14703 if (!MinMax && hasSubus) {
14704 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14706 // t = psubus Op0, Op1
14707 // pcmpeq t, <0..0>
14708 switch (SetCCOpcode) {
14710 case ISD::SETULT: {
14711 // If the comparison is against a constant we can turn this into a
14712 // setule. With psubus, setule does not require a swap. This is
14713 // beneficial because the constant in the register is no longer
14714 // destructed as the destination so it can be hoisted out of a loop.
14715 // Only do this pre-AVX since vpcmp* is no longer destructive.
14716 if (Subtarget->hasAVX())
14718 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14719 if (ULEOp1.getNode()) {
14721 Subus = true; Invert = false; Swap = false;
14725 // Psubus is better than flip-sign because it requires no inversion.
14726 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14727 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14731 Opc = X86ISD::SUBUS;
14737 std::swap(Op0, Op1);
14739 // Check that the operation in question is available (most are plain SSE2,
14740 // but PCMPGTQ and PCMPEQQ have different requirements).
14741 if (VT == MVT::v2i64) {
14742 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14743 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14745 // First cast everything to the right type.
14746 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14747 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14749 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14750 // bits of the inputs before performing those operations. The lower
14751 // compare is always unsigned.
14754 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14756 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14757 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14758 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14759 Sign, Zero, Sign, Zero);
14761 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14762 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14764 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14765 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14766 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14768 // Create masks for only the low parts/high parts of the 64 bit integers.
14769 static const int MaskHi[] = { 1, 1, 3, 3 };
14770 static const int MaskLo[] = { 0, 0, 2, 2 };
14771 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14772 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14773 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14775 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14776 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14779 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14781 return DAG.getBitcast(VT, Result);
14784 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14785 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14786 // pcmpeqd + pshufd + pand.
14787 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14789 // First cast everything to the right type.
14790 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14791 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14794 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14796 // Make sure the lower and upper halves are both all-ones.
14797 static const int Mask[] = { 1, 0, 3, 2 };
14798 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14799 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14802 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14804 return DAG.getBitcast(VT, Result);
14808 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14809 // bits of the inputs before performing those operations.
14811 MVT EltVT = VT.getVectorElementType();
14812 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14814 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14815 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14818 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14820 // If the logical-not of the result is required, perform that now.
14822 Result = DAG.getNOT(dl, Result, VT);
14825 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14828 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14829 getZeroVector(VT, Subtarget, DAG, dl));
14834 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14836 MVT VT = Op.getSimpleValueType();
14838 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14840 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14841 && "SetCC type must be 8-bit or 1-bit integer");
14842 SDValue Op0 = Op.getOperand(0);
14843 SDValue Op1 = Op.getOperand(1);
14845 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14847 // Optimize to BT if possible.
14848 // Lower (X & (1 << N)) == 0 to BT(X, N).
14849 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14850 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14851 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14852 isNullConstant(Op1) &&
14853 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14854 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14856 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14861 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14863 if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
14864 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14866 // If the input is a setcc, then reuse the input setcc or use a new one with
14867 // the inverted condition.
14868 if (Op0.getOpcode() == X86ISD::SETCC) {
14869 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14870 bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
14874 CCode = X86::GetOppositeBranchCondition(CCode);
14875 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14876 DAG.getConstant(CCode, dl, MVT::i8),
14877 Op0.getOperand(1));
14879 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14883 if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
14884 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14886 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14887 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14890 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14891 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14892 if (X86CC == X86::COND_INVALID)
14895 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14896 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14897 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14898 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14900 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14904 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14905 SDValue LHS = Op.getOperand(0);
14906 SDValue RHS = Op.getOperand(1);
14907 SDValue Carry = Op.getOperand(2);
14908 SDValue Cond = Op.getOperand(3);
14911 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14912 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14914 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14915 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14916 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14917 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14918 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14921 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14922 static bool isX86LogicalCmp(SDValue Op) {
14923 unsigned Opc = Op.getNode()->getOpcode();
14924 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14925 Opc == X86ISD::SAHF)
14927 if (Op.getResNo() == 1 &&
14928 (Opc == X86ISD::ADD ||
14929 Opc == X86ISD::SUB ||
14930 Opc == X86ISD::ADC ||
14931 Opc == X86ISD::SBB ||
14932 Opc == X86ISD::SMUL ||
14933 Opc == X86ISD::UMUL ||
14934 Opc == X86ISD::INC ||
14935 Opc == X86ISD::DEC ||
14936 Opc == X86ISD::OR ||
14937 Opc == X86ISD::XOR ||
14938 Opc == X86ISD::AND))
14941 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14947 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14948 if (V.getOpcode() != ISD::TRUNCATE)
14951 SDValue VOp0 = V.getOperand(0);
14952 unsigned InBits = VOp0.getValueSizeInBits();
14953 unsigned Bits = V.getValueSizeInBits();
14954 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14957 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14958 bool addTest = true;
14959 SDValue Cond = Op.getOperand(0);
14960 SDValue Op1 = Op.getOperand(1);
14961 SDValue Op2 = Op.getOperand(2);
14963 MVT VT = Op1.getSimpleValueType();
14966 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14967 // are available or VBLENDV if AVX is available.
14968 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14969 if (Cond.getOpcode() == ISD::SETCC &&
14970 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14971 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14972 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14973 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14974 int SSECC = translateX86FSETCC(
14975 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14978 if (Subtarget->hasAVX512()) {
14979 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14980 DAG.getConstant(SSECC, DL, MVT::i8));
14981 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14984 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14985 DAG.getConstant(SSECC, DL, MVT::i8));
14987 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14988 // of 3 logic instructions for size savings and potentially speed.
14989 // Unfortunately, there is no scalar form of VBLENDV.
14991 // If either operand is a constant, don't try this. We can expect to
14992 // optimize away at least one of the logic instructions later in that
14993 // case, so that sequence would be faster than a variable blend.
14995 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14996 // uses XMM0 as the selection register. That may need just as many
14997 // instructions as the AND/ANDN/OR sequence due to register moves, so
15000 if (Subtarget->hasAVX() &&
15001 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
15003 // Convert to vectors, do a VSELECT, and convert back to scalar.
15004 // All of the conversions should be optimized away.
15006 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
15007 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
15008 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
15009 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
15011 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
15012 VCmp = DAG.getBitcast(VCmpVT, VCmp);
15014 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
15016 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
15017 VSel, DAG.getIntPtrConstant(0, DL));
15019 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
15020 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
15021 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
15025 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
15027 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
15028 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
15029 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
15030 Op1Scalar = Op1.getOperand(0);
15032 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
15033 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
15034 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
15035 Op2Scalar = Op2.getOperand(0);
15036 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
15037 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
15038 Op1Scalar.getValueType(),
15039 Cond, Op1Scalar, Op2Scalar);
15040 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
15041 return DAG.getBitcast(VT, newSelect);
15042 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
15043 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
15044 DAG.getIntPtrConstant(0, DL));
15048 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
15049 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
15050 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15051 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
15052 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15053 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
15054 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
15056 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
15059 if (Cond.getOpcode() == ISD::SETCC) {
15060 SDValue NewCond = LowerSETCC(Cond, DAG);
15061 if (NewCond.getNode())
15065 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
15066 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
15067 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
15068 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
15069 if (Cond.getOpcode() == X86ISD::SETCC &&
15070 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
15071 isNullConstant(Cond.getOperand(1).getOperand(1))) {
15072 SDValue Cmp = Cond.getOperand(1);
15074 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
15076 if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15077 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
15078 SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
15080 SDValue CmpOp0 = Cmp.getOperand(0);
15081 // Apply further optimizations for special cases
15082 // (select (x != 0), -1, 0) -> neg & sbb
15083 // (select (x == 0), 0, -1) -> neg & sbb
15084 if (isNullConstant(Y) &&
15085 (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
15086 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
15087 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
15088 DAG.getConstant(0, DL,
15089 CmpOp0.getValueType()),
15091 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15092 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15093 SDValue(Neg.getNode(), 1));
15097 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
15098 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
15099 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15101 SDValue Res = // Res = 0 or -1.
15102 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15103 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
15105 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
15106 Res = DAG.getNOT(DL, Res, Res.getValueType());
15108 if (!isNullConstant(Op2))
15109 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
15114 // Look past (and (setcc_carry (cmp ...)), 1).
15115 if (Cond.getOpcode() == ISD::AND &&
15116 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15117 isOneConstant(Cond.getOperand(1)))
15118 Cond = Cond.getOperand(0);
15120 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15121 // setting operand in place of the X86ISD::SETCC.
15122 unsigned CondOpcode = Cond.getOpcode();
15123 if (CondOpcode == X86ISD::SETCC ||
15124 CondOpcode == X86ISD::SETCC_CARRY) {
15125 CC = Cond.getOperand(0);
15127 SDValue Cmp = Cond.getOperand(1);
15128 unsigned Opc = Cmp.getOpcode();
15129 MVT VT = Op.getSimpleValueType();
15131 bool IllegalFPCMov = false;
15132 if (VT.isFloatingPoint() && !VT.isVector() &&
15133 !isScalarFPTypeInSSEReg(VT)) // FPStack?
15134 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
15136 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
15137 Opc == X86ISD::BT) { // FIXME
15141 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15142 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15143 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15144 Cond.getOperand(0).getValueType() != MVT::i8)) {
15145 SDValue LHS = Cond.getOperand(0);
15146 SDValue RHS = Cond.getOperand(1);
15147 unsigned X86Opcode;
15150 switch (CondOpcode) {
15151 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15152 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15153 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15154 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15155 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15156 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15157 default: llvm_unreachable("unexpected overflowing operator");
15159 if (CondOpcode == ISD::UMULO)
15160 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15163 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15165 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
15167 if (CondOpcode == ISD::UMULO)
15168 Cond = X86Op.getValue(2);
15170 Cond = X86Op.getValue(1);
15172 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
15177 // Look past the truncate if the high bits are known zero.
15178 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15179 Cond = Cond.getOperand(0);
15181 // We know the result of AND is compared against zero. Try to match
15183 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15184 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
15185 CC = NewSetCC.getOperand(0);
15186 Cond = NewSetCC.getOperand(1);
15193 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
15194 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
15197 // a < b ? -1 : 0 -> RES = ~setcc_carry
15198 // a < b ? 0 : -1 -> RES = setcc_carry
15199 // a >= b ? -1 : 0 -> RES = setcc_carry
15200 // a >= b ? 0 : -1 -> RES = ~setcc_carry
15201 if (Cond.getOpcode() == X86ISD::SUB) {
15202 Cond = ConvertCmpIfNecessary(Cond, DAG);
15203 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
15205 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
15206 (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15207 (isNullConstant(Op1) || isNullConstant(Op2))) {
15208 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15209 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15211 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
15212 return DAG.getNOT(DL, Res, Res.getValueType());
15217 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15218 // widen the cmov and push the truncate through. This avoids introducing a new
15219 // branch during isel and doesn't add any extensions.
15220 if (Op.getValueType() == MVT::i8 &&
15221 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15222 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15223 if (T1.getValueType() == T2.getValueType() &&
15224 // Blacklist CopyFromReg to avoid partial register stalls.
15225 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15226 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15227 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15228 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15232 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15233 // condition is true.
15234 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15235 SDValue Ops[] = { Op2, Op1, CC, Cond };
15236 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15239 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
15240 const X86Subtarget *Subtarget,
15241 SelectionDAG &DAG) {
15242 MVT VT = Op->getSimpleValueType(0);
15243 SDValue In = Op->getOperand(0);
15244 MVT InVT = In.getSimpleValueType();
15245 MVT VTElt = VT.getVectorElementType();
15246 MVT InVTElt = InVT.getVectorElementType();
15250 if ((InVTElt == MVT::i1) &&
15251 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15252 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15254 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15255 VTElt.getSizeInBits() <= 16)) ||
15257 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15258 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15260 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15261 VTElt.getSizeInBits() >= 32))))
15262 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15264 unsigned int NumElts = VT.getVectorNumElements();
15266 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
15269 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15270 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15271 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15272 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15275 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15276 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
15278 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
15281 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
15283 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
15284 if (VT.is512BitVector())
15286 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
15289 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
15290 const X86Subtarget *Subtarget,
15291 SelectionDAG &DAG) {
15292 SDValue In = Op->getOperand(0);
15293 MVT VT = Op->getSimpleValueType(0);
15294 MVT InVT = In.getSimpleValueType();
15295 assert(VT.getSizeInBits() == InVT.getSizeInBits());
15297 MVT InSVT = InVT.getVectorElementType();
15298 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
15300 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
15302 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15307 // SSE41 targets can use the pmovsx* instructions directly.
15308 if (Subtarget->hasSSE41())
15309 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15311 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15315 // As SRAI is only available on i16/i32 types, we expand only up to i32
15316 // and handle i64 separately.
15317 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15318 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15319 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15320 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15321 Curr = DAG.getBitcast(CurrVT, Curr);
15324 SDValue SignExt = Curr;
15325 if (CurrVT != InVT) {
15326 unsigned SignExtShift =
15327 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15328 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15329 DAG.getConstant(SignExtShift, dl, MVT::i8));
15335 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15336 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15337 DAG.getConstant(31, dl, MVT::i8));
15338 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15339 return DAG.getBitcast(VT, Ext);
15345 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15346 SelectionDAG &DAG) {
15347 MVT VT = Op->getSimpleValueType(0);
15348 SDValue In = Op->getOperand(0);
15349 MVT InVT = In.getSimpleValueType();
15352 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15353 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15355 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15356 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15357 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15360 if (Subtarget->hasInt256())
15361 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15363 // Optimize vectors in AVX mode
15364 // Sign extend v8i16 to v8i32 and
15367 // Divide input vector into two parts
15368 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15369 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15370 // concat the vectors to original VT
15372 unsigned NumElems = InVT.getVectorNumElements();
15373 SDValue Undef = DAG.getUNDEF(InVT);
15375 SmallVector<int,8> ShufMask1(NumElems, -1);
15376 for (unsigned i = 0; i != NumElems/2; ++i)
15379 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15381 SmallVector<int,8> ShufMask2(NumElems, -1);
15382 for (unsigned i = 0; i != NumElems/2; ++i)
15383 ShufMask2[i] = i + NumElems/2;
15385 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15387 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15388 VT.getVectorNumElements()/2);
15390 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15391 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15393 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15396 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15397 // may emit an illegal shuffle but the expansion is still better than scalar
15398 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15399 // we'll emit a shuffle and a arithmetic shift.
15400 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15401 // TODO: It is possible to support ZExt by zeroing the undef values during
15402 // the shuffle phase or after the shuffle.
15403 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15404 SelectionDAG &DAG) {
15405 MVT RegVT = Op.getSimpleValueType();
15406 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15407 assert(RegVT.isInteger() &&
15408 "We only custom lower integer vector sext loads.");
15410 // Nothing useful we can do without SSE2 shuffles.
15411 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15413 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15415 EVT MemVT = Ld->getMemoryVT();
15416 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15417 unsigned RegSz = RegVT.getSizeInBits();
15419 ISD::LoadExtType Ext = Ld->getExtensionType();
15421 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15422 && "Only anyext and sext are currently implemented.");
15423 assert(MemVT != RegVT && "Cannot extend to the same type");
15424 assert(MemVT.isVector() && "Must load a vector from memory");
15426 unsigned NumElems = RegVT.getVectorNumElements();
15427 unsigned MemSz = MemVT.getSizeInBits();
15428 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15430 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15431 // The only way in which we have a legal 256-bit vector result but not the
15432 // integer 256-bit operations needed to directly lower a sextload is if we
15433 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15434 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15435 // correctly legalized. We do this late to allow the canonical form of
15436 // sextload to persist throughout the rest of the DAG combiner -- it wants
15437 // to fold together any extensions it can, and so will fuse a sign_extend
15438 // of an sextload into a sextload targeting a wider value.
15440 if (MemSz == 128) {
15441 // Just switch this to a normal load.
15442 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15443 "it must be a legal 128-bit vector "
15445 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15446 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15447 Ld->isInvariant(), Ld->getAlignment());
15449 assert(MemSz < 128 &&
15450 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15451 // Do an sext load to a 128-bit vector type. We want to use the same
15452 // number of elements, but elements half as wide. This will end up being
15453 // recursively lowered by this routine, but will succeed as we definitely
15454 // have all the necessary features if we're using AVX1.
15456 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15457 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15459 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15460 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15461 Ld->isNonTemporal(), Ld->isInvariant(),
15462 Ld->getAlignment());
15465 // Replace chain users with the new chain.
15466 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15467 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15469 // Finally, do a normal sign-extend to the desired register.
15470 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15473 // All sizes must be a power of two.
15474 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15475 "Non-power-of-two elements are not custom lowered!");
15477 // Attempt to load the original value using scalar loads.
15478 // Find the largest scalar type that divides the total loaded size.
15479 MVT SclrLoadTy = MVT::i8;
15480 for (MVT Tp : MVT::integer_valuetypes()) {
15481 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15486 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15487 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15489 SclrLoadTy = MVT::f64;
15491 // Calculate the number of scalar loads that we need to perform
15492 // in order to load our vector from memory.
15493 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15495 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15496 "Can only lower sext loads with a single scalar load!");
15498 unsigned loadRegZize = RegSz;
15499 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15502 // Represent our vector as a sequence of elements which are the
15503 // largest scalar that we can load.
15504 EVT LoadUnitVecVT = EVT::getVectorVT(
15505 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15507 // Represent the data using the same element type that is stored in
15508 // memory. In practice, we ''widen'' MemVT.
15510 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15511 loadRegZize / MemVT.getScalarSizeInBits());
15513 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15514 "Invalid vector type");
15516 // We can't shuffle using an illegal type.
15517 assert(TLI.isTypeLegal(WideVecVT) &&
15518 "We only lower types that form legal widened vector types");
15520 SmallVector<SDValue, 8> Chains;
15521 SDValue Ptr = Ld->getBasePtr();
15522 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15523 TLI.getPointerTy(DAG.getDataLayout()));
15524 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15526 for (unsigned i = 0; i < NumLoads; ++i) {
15527 // Perform a single load.
15528 SDValue ScalarLoad =
15529 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15530 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15531 Ld->getAlignment());
15532 Chains.push_back(ScalarLoad.getValue(1));
15533 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15534 // another round of DAGCombining.
15536 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15538 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15539 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15541 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15544 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15546 // Bitcast the loaded value to a vector of the original element type, in
15547 // the size of the target vector type.
15548 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15549 unsigned SizeRatio = RegSz / MemSz;
15551 if (Ext == ISD::SEXTLOAD) {
15552 // If we have SSE4.1, we can directly emit a VSEXT node.
15553 if (Subtarget->hasSSE41()) {
15554 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15555 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15559 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15561 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15562 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15564 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15565 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15569 // Redistribute the loaded elements into the different locations.
15570 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15571 for (unsigned i = 0; i != NumElems; ++i)
15572 ShuffleVec[i * SizeRatio] = i;
15574 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15575 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15577 // Bitcast to the requested type.
15578 Shuff = DAG.getBitcast(RegVT, Shuff);
15579 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15583 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15584 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15585 // from the AND / OR.
15586 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15587 Opc = Op.getOpcode();
15588 if (Opc != ISD::OR && Opc != ISD::AND)
15590 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15591 Op.getOperand(0).hasOneUse() &&
15592 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15593 Op.getOperand(1).hasOneUse());
15596 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15597 // 1 and that the SETCC node has a single use.
15598 static bool isXor1OfSetCC(SDValue Op) {
15599 if (Op.getOpcode() != ISD::XOR)
15601 if (isOneConstant(Op.getOperand(1)))
15602 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15603 Op.getOperand(0).hasOneUse();
15607 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15608 bool addTest = true;
15609 SDValue Chain = Op.getOperand(0);
15610 SDValue Cond = Op.getOperand(1);
15611 SDValue Dest = Op.getOperand(2);
15614 bool Inverted = false;
15616 if (Cond.getOpcode() == ISD::SETCC) {
15617 // Check for setcc([su]{add,sub,mul}o == 0).
15618 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15619 isNullConstant(Cond.getOperand(1)) &&
15620 Cond.getOperand(0).getResNo() == 1 &&
15621 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15622 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15623 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15624 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15625 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15626 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15628 Cond = Cond.getOperand(0);
15630 SDValue NewCond = LowerSETCC(Cond, DAG);
15631 if (NewCond.getNode())
15636 // FIXME: LowerXALUO doesn't handle these!!
15637 else if (Cond.getOpcode() == X86ISD::ADD ||
15638 Cond.getOpcode() == X86ISD::SUB ||
15639 Cond.getOpcode() == X86ISD::SMUL ||
15640 Cond.getOpcode() == X86ISD::UMUL)
15641 Cond = LowerXALUO(Cond, DAG);
15644 // Look pass (and (setcc_carry (cmp ...)), 1).
15645 if (Cond.getOpcode() == ISD::AND &&
15646 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15647 isOneConstant(Cond.getOperand(1)))
15648 Cond = Cond.getOperand(0);
15650 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15651 // setting operand in place of the X86ISD::SETCC.
15652 unsigned CondOpcode = Cond.getOpcode();
15653 if (CondOpcode == X86ISD::SETCC ||
15654 CondOpcode == X86ISD::SETCC_CARRY) {
15655 CC = Cond.getOperand(0);
15657 SDValue Cmp = Cond.getOperand(1);
15658 unsigned Opc = Cmp.getOpcode();
15659 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15660 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15664 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15668 // These can only come from an arithmetic instruction with overflow,
15669 // e.g. SADDO, UADDO.
15670 Cond = Cond.getNode()->getOperand(1);
15676 CondOpcode = Cond.getOpcode();
15677 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15678 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15679 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15680 Cond.getOperand(0).getValueType() != MVT::i8)) {
15681 SDValue LHS = Cond.getOperand(0);
15682 SDValue RHS = Cond.getOperand(1);
15683 unsigned X86Opcode;
15686 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15687 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15689 switch (CondOpcode) {
15690 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15692 if (isOneConstant(RHS)) {
15693 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15696 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15697 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15699 if (isOneConstant(RHS)) {
15700 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15703 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15704 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15705 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15706 default: llvm_unreachable("unexpected overflowing operator");
15709 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15710 if (CondOpcode == ISD::UMULO)
15711 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15714 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15716 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15718 if (CondOpcode == ISD::UMULO)
15719 Cond = X86Op.getValue(2);
15721 Cond = X86Op.getValue(1);
15723 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15727 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15728 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15729 if (CondOpc == ISD::OR) {
15730 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15731 // two branches instead of an explicit OR instruction with a
15733 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15734 isX86LogicalCmp(Cmp)) {
15735 CC = Cond.getOperand(0).getOperand(0);
15736 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15737 Chain, Dest, CC, Cmp);
15738 CC = Cond.getOperand(1).getOperand(0);
15742 } else { // ISD::AND
15743 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15744 // two branches instead of an explicit AND instruction with a
15745 // separate test. However, we only do this if this block doesn't
15746 // have a fall-through edge, because this requires an explicit
15747 // jmp when the condition is false.
15748 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15749 isX86LogicalCmp(Cmp) &&
15750 Op.getNode()->hasOneUse()) {
15751 X86::CondCode CCode =
15752 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15753 CCode = X86::GetOppositeBranchCondition(CCode);
15754 CC = DAG.getConstant(CCode, dl, MVT::i8);
15755 SDNode *User = *Op.getNode()->use_begin();
15756 // Look for an unconditional branch following this conditional branch.
15757 // We need this because we need to reverse the successors in order
15758 // to implement FCMP_OEQ.
15759 if (User->getOpcode() == ISD::BR) {
15760 SDValue FalseBB = User->getOperand(1);
15762 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15763 assert(NewBR == User);
15767 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15768 Chain, Dest, CC, Cmp);
15769 X86::CondCode CCode =
15770 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15771 CCode = X86::GetOppositeBranchCondition(CCode);
15772 CC = DAG.getConstant(CCode, dl, MVT::i8);
15778 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15779 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15780 // It should be transformed during dag combiner except when the condition
15781 // is set by a arithmetics with overflow node.
15782 X86::CondCode CCode =
15783 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15784 CCode = X86::GetOppositeBranchCondition(CCode);
15785 CC = DAG.getConstant(CCode, dl, MVT::i8);
15786 Cond = Cond.getOperand(0).getOperand(1);
15788 } else if (Cond.getOpcode() == ISD::SETCC &&
15789 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15790 // For FCMP_OEQ, we can emit
15791 // two branches instead of an explicit AND instruction with a
15792 // separate test. However, we only do this if this block doesn't
15793 // have a fall-through edge, because this requires an explicit
15794 // jmp when the condition is false.
15795 if (Op.getNode()->hasOneUse()) {
15796 SDNode *User = *Op.getNode()->use_begin();
15797 // Look for an unconditional branch following this conditional branch.
15798 // We need this because we need to reverse the successors in order
15799 // to implement FCMP_OEQ.
15800 if (User->getOpcode() == ISD::BR) {
15801 SDValue FalseBB = User->getOperand(1);
15803 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15804 assert(NewBR == User);
15808 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15809 Cond.getOperand(0), Cond.getOperand(1));
15810 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15811 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15812 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15813 Chain, Dest, CC, Cmp);
15814 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15819 } else if (Cond.getOpcode() == ISD::SETCC &&
15820 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15821 // For FCMP_UNE, we can emit
15822 // two branches instead of an explicit AND instruction with a
15823 // separate test. However, we only do this if this block doesn't
15824 // have a fall-through edge, because this requires an explicit
15825 // jmp when the condition is false.
15826 if (Op.getNode()->hasOneUse()) {
15827 SDNode *User = *Op.getNode()->use_begin();
15828 // Look for an unconditional branch following this conditional branch.
15829 // We need this because we need to reverse the successors in order
15830 // to implement FCMP_UNE.
15831 if (User->getOpcode() == ISD::BR) {
15832 SDValue FalseBB = User->getOperand(1);
15834 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15835 assert(NewBR == User);
15838 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15839 Cond.getOperand(0), Cond.getOperand(1));
15840 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15841 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15842 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15843 Chain, Dest, CC, Cmp);
15844 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15854 // Look pass the truncate if the high bits are known zero.
15855 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15856 Cond = Cond.getOperand(0);
15858 // We know the result of AND is compared against zero. Try to match
15860 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15861 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15862 CC = NewSetCC.getOperand(0);
15863 Cond = NewSetCC.getOperand(1);
15870 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15871 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15872 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15874 Cond = ConvertCmpIfNecessary(Cond, DAG);
15875 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15876 Chain, Dest, CC, Cond);
15879 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15880 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15881 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15882 // that the guard pages used by the OS virtual memory manager are allocated in
15883 // correct sequence.
15885 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15886 SelectionDAG &DAG) const {
15887 MachineFunction &MF = DAG.getMachineFunction();
15888 bool SplitStack = MF.shouldSplitStack();
15889 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15894 SDNode *Node = Op.getNode();
15895 SDValue Chain = Op.getOperand(0);
15896 SDValue Size = Op.getOperand(1);
15897 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15898 EVT VT = Node->getValueType(0);
15900 // Chain the dynamic stack allocation so that it doesn't modify the stack
15901 // pointer when other instructions are using the stack.
15902 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
15904 bool Is64Bit = Subtarget->is64Bit();
15905 MVT SPTy = getPointerTy(DAG.getDataLayout());
15909 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15910 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15911 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15912 " not tell us which reg is the stack pointer!");
15913 EVT VT = Node->getValueType(0);
15914 SDValue Tmp3 = Node->getOperand(2);
15916 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15917 Chain = SP.getValue(1);
15918 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15919 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15920 unsigned StackAlign = TFI.getStackAlignment();
15921 Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15922 if (Align > StackAlign)
15923 Result = DAG.getNode(ISD::AND, dl, VT, Result,
15924 DAG.getConstant(-(uint64_t)Align, dl, VT));
15925 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain
15926 } else if (SplitStack) {
15927 MachineRegisterInfo &MRI = MF.getRegInfo();
15930 // The 64 bit implementation of segmented stacks needs to clobber both r10
15931 // r11. This makes it impossible to use it along with nested parameters.
15932 const Function *F = MF.getFunction();
15934 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15936 if (I->hasNestAttr())
15937 report_fatal_error("Cannot use segmented stacks with functions that "
15938 "have nested arguments.");
15941 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15942 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15943 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15944 Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15945 DAG.getRegister(Vreg, SPTy));
15948 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15950 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15951 Flag = Chain.getValue(1);
15952 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15954 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15956 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15957 unsigned SPReg = RegInfo->getStackRegister();
15958 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15959 Chain = SP.getValue(1);
15962 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15963 DAG.getConstant(-(uint64_t)Align, dl, VT));
15964 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15970 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15971 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
15973 SDValue Ops[2] = {Result, Chain};
15974 return DAG.getMergeValues(Ops, dl);
15977 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15978 MachineFunction &MF = DAG.getMachineFunction();
15979 auto PtrVT = getPointerTy(MF.getDataLayout());
15980 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15982 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15985 if (!Subtarget->is64Bit() ||
15986 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15987 // vastart just stores the address of the VarArgsFrameIndex slot into the
15988 // memory location argument.
15989 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15990 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15991 MachinePointerInfo(SV), false, false, 0);
15995 // gp_offset (0 - 6 * 8)
15996 // fp_offset (48 - 48 + 8 * 16)
15997 // overflow_arg_area (point to parameters coming in memory).
15999 SmallVector<SDValue, 8> MemOps;
16000 SDValue FIN = Op.getOperand(1);
16002 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
16003 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
16005 FIN, MachinePointerInfo(SV), false, false, 0);
16006 MemOps.push_back(Store);
16009 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
16010 Store = DAG.getStore(Op.getOperand(0), DL,
16011 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
16013 FIN, MachinePointerInfo(SV, 4), false, false, 0);
16014 MemOps.push_back(Store);
16016 // Store ptr to overflow_arg_area
16017 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
16018 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
16019 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
16020 MachinePointerInfo(SV, 8),
16022 MemOps.push_back(Store);
16024 // Store ptr to reg_save_area.
16025 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
16026 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
16027 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
16028 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
16029 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
16030 MemOps.push_back(Store);
16031 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
16034 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
16035 assert(Subtarget->is64Bit() &&
16036 "LowerVAARG only handles 64-bit va_arg!");
16037 assert(Op.getNode()->getNumOperands() == 4);
16039 MachineFunction &MF = DAG.getMachineFunction();
16040 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
16041 // The Win64 ABI uses char* instead of a structure.
16042 return DAG.expandVAArg(Op.getNode());
16044 SDValue Chain = Op.getOperand(0);
16045 SDValue SrcPtr = Op.getOperand(1);
16046 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16047 unsigned Align = Op.getConstantOperandVal(3);
16050 EVT ArgVT = Op.getNode()->getValueType(0);
16051 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16052 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
16055 // Decide which area this value should be read from.
16056 // TODO: Implement the AMD64 ABI in its entirety. This simple
16057 // selection mechanism works only for the basic types.
16058 if (ArgVT == MVT::f80) {
16059 llvm_unreachable("va_arg for f80 not yet implemented");
16060 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
16061 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
16062 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
16063 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
16065 llvm_unreachable("Unhandled argument type in LowerVAARG");
16068 if (ArgMode == 2) {
16069 // Sanity Check: Make sure using fp_offset makes sense.
16070 assert(!Subtarget->useSoftFloat() &&
16071 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
16072 Subtarget->hasSSE1());
16075 // Insert VAARG_64 node into the DAG
16076 // VAARG_64 returns two values: Variable Argument Address, Chain
16077 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
16078 DAG.getConstant(ArgMode, dl, MVT::i8),
16079 DAG.getConstant(Align, dl, MVT::i32)};
16080 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
16081 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
16082 VTs, InstOps, MVT::i64,
16083 MachinePointerInfo(SV),
16085 /*Volatile=*/false,
16087 /*WriteMem=*/true);
16088 Chain = VAARG.getValue(1);
16090 // Load the next argument and return it
16091 return DAG.getLoad(ArgVT, dl,
16094 MachinePointerInfo(),
16095 false, false, false, 0);
16098 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
16099 SelectionDAG &DAG) {
16100 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
16101 // where a va_list is still an i8*.
16102 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
16103 if (Subtarget->isCallingConvWin64(
16104 DAG.getMachineFunction().getFunction()->getCallingConv()))
16105 // Probably a Win64 va_copy.
16106 return DAG.expandVACopy(Op.getNode());
16108 SDValue Chain = Op.getOperand(0);
16109 SDValue DstPtr = Op.getOperand(1);
16110 SDValue SrcPtr = Op.getOperand(2);
16111 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
16112 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16115 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
16116 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
16118 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
16121 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
16122 // amount is a constant. Takes immediate version of shift as input.
16123 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
16124 SDValue SrcOp, uint64_t ShiftAmt,
16125 SelectionDAG &DAG) {
16126 MVT ElementType = VT.getVectorElementType();
16128 // Fold this packed shift into its first operand if ShiftAmt is 0.
16132 // Check for ShiftAmt >= element width
16133 if (ShiftAmt >= ElementType.getSizeInBits()) {
16134 if (Opc == X86ISD::VSRAI)
16135 ShiftAmt = ElementType.getSizeInBits() - 1;
16137 return DAG.getConstant(0, dl, VT);
16140 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
16141 && "Unknown target vector shift-by-constant node");
16143 // Fold this packed vector shift into a build vector if SrcOp is a
16144 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
16145 if (VT == SrcOp.getSimpleValueType() &&
16146 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
16147 SmallVector<SDValue, 8> Elts;
16148 unsigned NumElts = SrcOp->getNumOperands();
16149 ConstantSDNode *ND;
16152 default: llvm_unreachable(nullptr);
16153 case X86ISD::VSHLI:
16154 for (unsigned i=0; i!=NumElts; ++i) {
16155 SDValue CurrentOp = SrcOp->getOperand(i);
16156 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16157 Elts.push_back(CurrentOp);
16160 ND = cast<ConstantSDNode>(CurrentOp);
16161 const APInt &C = ND->getAPIntValue();
16162 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
16165 case X86ISD::VSRLI:
16166 for (unsigned i=0; i!=NumElts; ++i) {
16167 SDValue CurrentOp = SrcOp->getOperand(i);
16168 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16169 Elts.push_back(CurrentOp);
16172 ND = cast<ConstantSDNode>(CurrentOp);
16173 const APInt &C = ND->getAPIntValue();
16174 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
16177 case X86ISD::VSRAI:
16178 for (unsigned i=0; i!=NumElts; ++i) {
16179 SDValue CurrentOp = SrcOp->getOperand(i);
16180 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16181 Elts.push_back(CurrentOp);
16184 ND = cast<ConstantSDNode>(CurrentOp);
16185 const APInt &C = ND->getAPIntValue();
16186 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
16191 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16194 return DAG.getNode(Opc, dl, VT, SrcOp,
16195 DAG.getConstant(ShiftAmt, dl, MVT::i8));
16198 // getTargetVShiftNode - Handle vector element shifts where the shift amount
16199 // may or may not be a constant. Takes immediate version of shift as input.
16200 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
16201 SDValue SrcOp, SDValue ShAmt,
16202 SelectionDAG &DAG) {
16203 MVT SVT = ShAmt.getSimpleValueType();
16204 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
16206 // Catch shift-by-constant.
16207 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16208 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16209 CShAmt->getZExtValue(), DAG);
16211 // Change opcode to non-immediate version
16213 default: llvm_unreachable("Unknown target vector shift node");
16214 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16215 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16216 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16219 const X86Subtarget &Subtarget =
16220 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16221 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16222 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16223 // Let the shuffle legalizer expand this shift amount node.
16224 SDValue Op0 = ShAmt.getOperand(0);
16225 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16226 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16228 // Need to build a vector containing shift amount.
16229 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16230 SmallVector<SDValue, 4> ShOps;
16231 ShOps.push_back(ShAmt);
16232 if (SVT == MVT::i32) {
16233 ShOps.push_back(DAG.getConstant(0, dl, SVT));
16234 ShOps.push_back(DAG.getUNDEF(SVT));
16236 ShOps.push_back(DAG.getUNDEF(SVT));
16238 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16239 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16242 // The return type has to be a 128-bit type with the same element
16243 // type as the input type.
16244 MVT EltVT = VT.getVectorElementType();
16245 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16247 ShAmt = DAG.getBitcast(ShVT, ShAmt);
16248 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16251 /// \brief Return Mask with the necessary casting or extending
16252 /// for \p Mask according to \p MaskVT when lowering masking intrinsics
16253 static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
16254 const X86Subtarget *Subtarget,
16255 SelectionDAG &DAG, SDLoc dl) {
16257 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
16258 // Mask should be extended
16259 Mask = DAG.getNode(ISD::ANY_EXTEND, dl,
16260 MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask);
16263 if (Mask.getSimpleValueType() == MVT::i64 && Subtarget->is32Bit()) {
16264 if (MaskVT == MVT::v64i1) {
16265 assert(Subtarget->hasBWI() && "Expected AVX512BW target!");
16266 // In case 32bit mode, bitcast i64 is illegal, extend/split it.
16268 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16269 DAG.getConstant(0, dl, MVT::i32));
16270 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16271 DAG.getConstant(1, dl, MVT::i32));
16273 Lo = DAG.getBitcast(MVT::v32i1, Lo);
16274 Hi = DAG.getBitcast(MVT::v32i1, Hi);
16276 return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
16278 // MaskVT require < 64bit. Truncate mask (should succeed in any case),
16280 MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
16281 return DAG.getBitcast(MaskVT,
16282 DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask));
16286 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16287 Mask.getSimpleValueType().getSizeInBits());
16288 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16289 // are extracted by EXTRACT_SUBVECTOR.
16290 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16291 DAG.getBitcast(BitcastVT, Mask),
16292 DAG.getIntPtrConstant(0, dl));
16296 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16297 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16298 /// necessary casting or extending for \p Mask when lowering masking intrinsics
16299 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16300 SDValue PreservedSrc,
16301 const X86Subtarget *Subtarget,
16302 SelectionDAG &DAG) {
16303 MVT VT = Op.getSimpleValueType();
16304 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16305 unsigned OpcodeSelect = ISD::VSELECT;
16308 if (isAllOnesConstant(Mask))
16311 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16313 switch (Op.getOpcode()) {
16315 case X86ISD::PCMPEQM:
16316 case X86ISD::PCMPGTM:
16318 case X86ISD::CMPMU:
16319 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16320 case X86ISD::VFPCLASS:
16321 case X86ISD::VFPCLASSS:
16322 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16323 case X86ISD::VTRUNC:
16324 case X86ISD::VTRUNCS:
16325 case X86ISD::VTRUNCUS:
16326 // We can't use ISD::VSELECT here because it is not always "Legal"
16327 // for the destination type. For example vpmovqb require only AVX512
16328 // and vselect that can operate on byte element type require BWI
16329 OpcodeSelect = X86ISD::SELECT;
16332 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16333 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16334 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16337 /// \brief Creates an SDNode for a predicated scalar operation.
16338 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16339 /// The mask is coming as MVT::i8 and it should be truncated
16340 /// to MVT::i1 while lowering masking intrinsics.
16341 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16342 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16343 /// for a scalar instruction.
16344 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16345 SDValue PreservedSrc,
16346 const X86Subtarget *Subtarget,
16347 SelectionDAG &DAG) {
16348 if (isAllOnesConstant(Mask))
16351 MVT VT = Op.getSimpleValueType();
16353 // The mask should be of type MVT::i1
16354 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16356 if (Op.getOpcode() == X86ISD::FSETCC)
16357 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16358 if (Op.getOpcode() == X86ISD::VFPCLASS ||
16359 Op.getOpcode() == X86ISD::VFPCLASSS)
16360 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16362 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16363 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16364 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16367 static int getSEHRegistrationNodeSize(const Function *Fn) {
16368 if (!Fn->hasPersonalityFn())
16369 report_fatal_error(
16370 "querying registration node size for function without personality");
16371 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16372 // WinEHStatePass for the full struct definition.
16373 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16374 case EHPersonality::MSVC_X86SEH: return 24;
16375 case EHPersonality::MSVC_CXX: return 16;
16378 report_fatal_error(
16379 "can only recover FP for 32-bit MSVC EH personality functions");
16382 /// When the MSVC runtime transfers control to us, either to an outlined
16383 /// function or when returning to a parent frame after catching an exception, we
16384 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16385 /// Here's the math:
16386 /// RegNodeBase = EntryEBP - RegNodeSize
16387 /// ParentFP = RegNodeBase - ParentFrameOffset
16388 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16389 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16390 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16391 SDValue EntryEBP) {
16392 MachineFunction &MF = DAG.getMachineFunction();
16395 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16396 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16398 // It's possible that the parent function no longer has a personality function
16399 // if the exceptional code was optimized away, in which case we just return
16400 // the incoming EBP.
16401 if (!Fn->hasPersonalityFn())
16404 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16405 // registration, or the .set_setframe offset.
16406 MCSymbol *OffsetSym =
16407 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16408 GlobalValue::getRealLinkageName(Fn->getName()));
16409 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16410 SDValue ParentFrameOffset =
16411 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16413 // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after
16414 // prologue to RBP in the parent function.
16415 const X86Subtarget &Subtarget =
16416 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16417 if (Subtarget.is64Bit())
16418 return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset);
16420 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16421 // RegNodeBase = EntryEBP - RegNodeSize
16422 // ParentFP = RegNodeBase - ParentFrameOffset
16423 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16424 DAG.getConstant(RegNodeSize, dl, PtrVT));
16425 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset);
16428 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16429 SelectionDAG &DAG) {
16431 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16432 MVT VT = Op.getSimpleValueType();
16433 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16435 switch(IntrData->Type) {
16436 case INTR_TYPE_1OP:
16437 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16438 case INTR_TYPE_2OP:
16439 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16441 case INTR_TYPE_2OP_IMM8:
16442 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16443 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16444 case INTR_TYPE_3OP:
16445 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16446 Op.getOperand(2), Op.getOperand(3));
16447 case INTR_TYPE_4OP:
16448 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16449 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16450 case INTR_TYPE_1OP_MASK_RM: {
16451 SDValue Src = Op.getOperand(1);
16452 SDValue PassThru = Op.getOperand(2);
16453 SDValue Mask = Op.getOperand(3);
16454 SDValue RoundingMode;
16455 // We allways add rounding mode to the Node.
16456 // If the rounding mode is not specified, we add the
16457 // "current direction" mode.
16458 if (Op.getNumOperands() == 4)
16460 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16462 RoundingMode = Op.getOperand(4);
16463 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16464 if (IntrWithRoundingModeOpcode != 0)
16465 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16466 X86::STATIC_ROUNDING::CUR_DIRECTION)
16467 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16468 dl, Op.getValueType(), Src, RoundingMode),
16469 Mask, PassThru, Subtarget, DAG);
16470 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16472 Mask, PassThru, Subtarget, DAG);
16474 case INTR_TYPE_1OP_MASK: {
16475 SDValue Src = Op.getOperand(1);
16476 SDValue PassThru = Op.getOperand(2);
16477 SDValue Mask = Op.getOperand(3);
16478 // We add rounding mode to the Node when
16479 // - RM Opcode is specified and
16480 // - RM is not "current direction".
16481 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16482 if (IntrWithRoundingModeOpcode != 0) {
16483 SDValue Rnd = Op.getOperand(4);
16484 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16485 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16486 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16487 dl, Op.getValueType(),
16489 Mask, PassThru, Subtarget, DAG);
16492 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16493 Mask, PassThru, Subtarget, DAG);
16495 case INTR_TYPE_SCALAR_MASK: {
16496 SDValue Src1 = Op.getOperand(1);
16497 SDValue Src2 = Op.getOperand(2);
16498 SDValue passThru = Op.getOperand(3);
16499 SDValue Mask = Op.getOperand(4);
16500 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16501 Mask, passThru, Subtarget, DAG);
16503 case INTR_TYPE_SCALAR_MASK_RM: {
16504 SDValue Src1 = Op.getOperand(1);
16505 SDValue Src2 = Op.getOperand(2);
16506 SDValue Src0 = Op.getOperand(3);
16507 SDValue Mask = Op.getOperand(4);
16508 // There are 2 kinds of intrinsics in this group:
16509 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16510 // (2) With rounding mode and sae - 7 operands.
16511 if (Op.getNumOperands() == 6) {
16512 SDValue Sae = Op.getOperand(5);
16513 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16514 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16516 Mask, Src0, Subtarget, DAG);
16518 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16519 SDValue RoundingMode = Op.getOperand(5);
16520 SDValue Sae = Op.getOperand(6);
16521 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16522 RoundingMode, Sae),
16523 Mask, Src0, Subtarget, DAG);
16525 case INTR_TYPE_2OP_MASK:
16526 case INTR_TYPE_2OP_IMM8_MASK: {
16527 SDValue Src1 = Op.getOperand(1);
16528 SDValue Src2 = Op.getOperand(2);
16529 SDValue PassThru = Op.getOperand(3);
16530 SDValue Mask = Op.getOperand(4);
16532 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16533 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16535 // We specify 2 possible opcodes for intrinsics with rounding modes.
16536 // First, we check if the intrinsic may have non-default rounding mode,
16537 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16538 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16539 if (IntrWithRoundingModeOpcode != 0) {
16540 SDValue Rnd = Op.getOperand(5);
16541 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16542 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16543 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16544 dl, Op.getValueType(),
16546 Mask, PassThru, Subtarget, DAG);
16549 // TODO: Intrinsics should have fast-math-flags to propagate.
16550 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16551 Mask, PassThru, Subtarget, DAG);
16553 case INTR_TYPE_2OP_MASK_RM: {
16554 SDValue Src1 = Op.getOperand(1);
16555 SDValue Src2 = Op.getOperand(2);
16556 SDValue PassThru = Op.getOperand(3);
16557 SDValue Mask = Op.getOperand(4);
16558 // We specify 2 possible modes for intrinsics, with/without rounding
16560 // First, we check if the intrinsic have rounding mode (6 operands),
16561 // if not, we set rounding mode to "current".
16563 if (Op.getNumOperands() == 6)
16564 Rnd = Op.getOperand(5);
16566 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16567 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16569 Mask, PassThru, Subtarget, DAG);
16571 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16572 SDValue Src1 = Op.getOperand(1);
16573 SDValue Src2 = Op.getOperand(2);
16574 SDValue Src3 = Op.getOperand(3);
16575 SDValue PassThru = Op.getOperand(4);
16576 SDValue Mask = Op.getOperand(5);
16577 SDValue Sae = Op.getOperand(6);
16579 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16581 Mask, PassThru, Subtarget, DAG);
16583 case INTR_TYPE_3OP_MASK_RM: {
16584 SDValue Src1 = Op.getOperand(1);
16585 SDValue Src2 = Op.getOperand(2);
16586 SDValue Imm = Op.getOperand(3);
16587 SDValue PassThru = Op.getOperand(4);
16588 SDValue Mask = Op.getOperand(5);
16589 // We specify 2 possible modes for intrinsics, with/without rounding
16591 // First, we check if the intrinsic have rounding mode (7 operands),
16592 // if not, we set rounding mode to "current".
16594 if (Op.getNumOperands() == 7)
16595 Rnd = Op.getOperand(6);
16597 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16598 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16599 Src1, Src2, Imm, Rnd),
16600 Mask, PassThru, Subtarget, DAG);
16602 case INTR_TYPE_3OP_IMM8_MASK:
16603 case INTR_TYPE_3OP_MASK:
16604 case INSERT_SUBVEC: {
16605 SDValue Src1 = Op.getOperand(1);
16606 SDValue Src2 = Op.getOperand(2);
16607 SDValue Src3 = Op.getOperand(3);
16608 SDValue PassThru = Op.getOperand(4);
16609 SDValue Mask = Op.getOperand(5);
16611 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16612 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16613 else if (IntrData->Type == INSERT_SUBVEC) {
16614 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16615 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16616 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16617 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16618 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16621 // We specify 2 possible opcodes for intrinsics with rounding modes.
16622 // First, we check if the intrinsic may have non-default rounding mode,
16623 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16624 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16625 if (IntrWithRoundingModeOpcode != 0) {
16626 SDValue Rnd = Op.getOperand(6);
16627 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16628 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16629 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16630 dl, Op.getValueType(),
16631 Src1, Src2, Src3, Rnd),
16632 Mask, PassThru, Subtarget, DAG);
16635 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16637 Mask, PassThru, Subtarget, DAG);
16639 case VPERM_3OP_MASKZ:
16640 case VPERM_3OP_MASK:{
16641 // Src2 is the PassThru
16642 SDValue Src1 = Op.getOperand(1);
16643 SDValue Src2 = Op.getOperand(2);
16644 SDValue Src3 = Op.getOperand(3);
16645 SDValue Mask = Op.getOperand(4);
16646 MVT VT = Op.getSimpleValueType();
16647 SDValue PassThru = SDValue();
16649 // set PassThru element
16650 if (IntrData->Type == VPERM_3OP_MASKZ)
16651 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16653 PassThru = DAG.getBitcast(VT, Src2);
16655 // Swap Src1 and Src2 in the node creation
16656 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16657 dl, Op.getValueType(),
16659 Mask, PassThru, Subtarget, DAG);
16663 case FMA_OP_MASK: {
16664 SDValue Src1 = Op.getOperand(1);
16665 SDValue Src2 = Op.getOperand(2);
16666 SDValue Src3 = Op.getOperand(3);
16667 SDValue Mask = Op.getOperand(4);
16668 MVT VT = Op.getSimpleValueType();
16669 SDValue PassThru = SDValue();
16671 // set PassThru element
16672 if (IntrData->Type == FMA_OP_MASKZ)
16673 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16674 else if (IntrData->Type == FMA_OP_MASK3)
16679 // We specify 2 possible opcodes for intrinsics with rounding modes.
16680 // First, we check if the intrinsic may have non-default rounding mode,
16681 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16682 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16683 if (IntrWithRoundingModeOpcode != 0) {
16684 SDValue Rnd = Op.getOperand(5);
16685 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16686 X86::STATIC_ROUNDING::CUR_DIRECTION)
16687 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16688 dl, Op.getValueType(),
16689 Src1, Src2, Src3, Rnd),
16690 Mask, PassThru, Subtarget, DAG);
16692 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16693 dl, Op.getValueType(),
16695 Mask, PassThru, Subtarget, DAG);
16697 case TERLOG_OP_MASK:
16698 case TERLOG_OP_MASKZ: {
16699 SDValue Src1 = Op.getOperand(1);
16700 SDValue Src2 = Op.getOperand(2);
16701 SDValue Src3 = Op.getOperand(3);
16702 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16703 SDValue Mask = Op.getOperand(5);
16704 MVT VT = Op.getSimpleValueType();
16705 SDValue PassThru = Src1;
16706 // Set PassThru element.
16707 if (IntrData->Type == TERLOG_OP_MASKZ)
16708 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16710 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16711 Src1, Src2, Src3, Src4),
16712 Mask, PassThru, Subtarget, DAG);
16715 // FPclass intrinsics with mask
16716 SDValue Src1 = Op.getOperand(1);
16717 MVT VT = Src1.getSimpleValueType();
16718 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16719 SDValue Imm = Op.getOperand(2);
16720 SDValue Mask = Op.getOperand(3);
16721 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16722 Mask.getSimpleValueType().getSizeInBits());
16723 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16724 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16725 DAG.getTargetConstant(0, dl, MaskVT),
16727 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16728 DAG.getUNDEF(BitcastVT), FPclassMask,
16729 DAG.getIntPtrConstant(0, dl));
16730 return DAG.getBitcast(Op.getValueType(), Res);
16733 SDValue Src1 = Op.getOperand(1);
16734 SDValue Imm = Op.getOperand(2);
16735 SDValue Mask = Op.getOperand(3);
16736 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16737 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16738 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16739 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16742 case CMP_MASK_CC: {
16743 // Comparison intrinsics with masks.
16744 // Example of transformation:
16745 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16746 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16748 // (v8i1 (insert_subvector undef,
16749 // (v2i1 (and (PCMPEQM %a, %b),
16750 // (extract_subvector
16751 // (v8i1 (bitcast %mask)), 0))), 0))))
16752 MVT VT = Op.getOperand(1).getSimpleValueType();
16753 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16754 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16755 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16756 Mask.getSimpleValueType().getSizeInBits());
16758 if (IntrData->Type == CMP_MASK_CC) {
16759 SDValue CC = Op.getOperand(3);
16760 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16761 // We specify 2 possible opcodes for intrinsics with rounding modes.
16762 // First, we check if the intrinsic may have non-default rounding mode,
16763 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16764 if (IntrData->Opc1 != 0) {
16765 SDValue Rnd = Op.getOperand(5);
16766 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16767 X86::STATIC_ROUNDING::CUR_DIRECTION)
16768 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16769 Op.getOperand(2), CC, Rnd);
16771 //default rounding mode
16773 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16774 Op.getOperand(2), CC);
16777 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16778 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16781 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16782 DAG.getTargetConstant(0, dl,
16785 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16786 DAG.getUNDEF(BitcastVT), CmpMask,
16787 DAG.getIntPtrConstant(0, dl));
16788 return DAG.getBitcast(Op.getValueType(), Res);
16790 case CMP_MASK_SCALAR_CC: {
16791 SDValue Src1 = Op.getOperand(1);
16792 SDValue Src2 = Op.getOperand(2);
16793 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16794 SDValue Mask = Op.getOperand(4);
16797 if (IntrData->Opc1 != 0) {
16798 SDValue Rnd = Op.getOperand(5);
16799 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16800 X86::STATIC_ROUNDING::CUR_DIRECTION)
16801 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16803 //default rounding mode
16805 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16807 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16808 DAG.getTargetConstant(0, dl,
16812 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16813 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16814 DAG.getValueType(MVT::i1));
16816 case COMI: { // Comparison intrinsics
16817 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16818 SDValue LHS = Op.getOperand(1);
16819 SDValue RHS = Op.getOperand(2);
16820 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16821 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16822 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16823 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16824 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16825 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16827 case COMI_RM: { // Comparison intrinsics with Sae
16828 SDValue LHS = Op.getOperand(1);
16829 SDValue RHS = Op.getOperand(2);
16830 SDValue CC = Op.getOperand(3);
16831 SDValue Sae = Op.getOperand(4);
16832 auto ComiType = TranslateX86ConstCondToX86CC(CC);
16833 // choose between ordered and unordered (comi/ucomi)
16834 unsigned comiOp = std::get<0>(ComiType) ? IntrData->Opc0 : IntrData->Opc1;
16836 if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
16837 X86::STATIC_ROUNDING::CUR_DIRECTION)
16838 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
16840 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
16841 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16842 DAG.getConstant(std::get<1>(ComiType), dl, MVT::i8), Cond);
16843 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16846 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16847 Op.getOperand(1), Op.getOperand(2), DAG);
16849 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16850 Op.getSimpleValueType(),
16852 Op.getOperand(2), DAG),
16853 Op.getOperand(4), Op.getOperand(3), Subtarget,
16855 case COMPRESS_EXPAND_IN_REG: {
16856 SDValue Mask = Op.getOperand(3);
16857 SDValue DataToCompress = Op.getOperand(1);
16858 SDValue PassThru = Op.getOperand(2);
16859 if (isAllOnesConstant(Mask)) // return data as is
16860 return Op.getOperand(1);
16862 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16864 Mask, PassThru, Subtarget, DAG);
16867 SDValue Mask = Op.getOperand(1);
16868 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16869 Mask.getSimpleValueType().getSizeInBits());
16870 Mask = DAG.getBitcast(MaskVT, Mask);
16871 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16874 SDValue Mask = Op.getOperand(3);
16875 MVT VT = Op.getSimpleValueType();
16876 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16877 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16878 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16882 MVT VT = Op.getSimpleValueType();
16883 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2);
16885 SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl);
16886 SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl);
16887 // Arguments should be swapped.
16888 SDValue Res = DAG.getNode(IntrData->Opc0, dl,
16889 MVT::getVectorVT(MVT::i1, VT.getSizeInBits()),
16891 return DAG.getBitcast(VT, Res);
16893 case CONVERT_TO_MASK: {
16894 MVT SrcVT = Op.getOperand(1).getSimpleValueType();
16895 MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements());
16896 MVT BitcastVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits());
16898 SDValue CvtMask = DAG.getNode(IntrData->Opc0, dl, MaskVT,
16900 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16901 DAG.getUNDEF(BitcastVT), CvtMask,
16902 DAG.getIntPtrConstant(0, dl));
16903 return DAG.getBitcast(Op.getValueType(), Res);
16905 case CONVERT_MASK_TO_VEC: {
16906 SDValue Mask = Op.getOperand(1);
16907 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16908 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16909 return DAG.getNode(IntrData->Opc0, dl, VT, VMask);
16911 case BRCST_SUBVEC_TO_VEC: {
16912 SDValue Src = Op.getOperand(1);
16913 SDValue Passthru = Op.getOperand(2);
16914 SDValue Mask = Op.getOperand(3);
16915 EVT resVT = Passthru.getValueType();
16916 SDValue subVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, resVT,
16917 DAG.getUNDEF(resVT), Src,
16918 DAG.getIntPtrConstant(0, dl));
16920 if (Src.getSimpleValueType().is256BitVector() && resVT.is512BitVector())
16921 immVal = DAG.getConstant(0x44, dl, MVT::i8);
16923 immVal = DAG.getConstant(0, dl, MVT::i8);
16924 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16925 subVec, subVec, immVal),
16926 Mask, Passthru, Subtarget, DAG);
16934 default: return SDValue(); // Don't custom lower most intrinsics.
16936 case Intrinsic::x86_avx2_permd:
16937 case Intrinsic::x86_avx2_permps:
16938 // Operands intentionally swapped. Mask is last operand to intrinsic,
16939 // but second operand for node/instruction.
16940 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16941 Op.getOperand(2), Op.getOperand(1));
16943 // ptest and testp intrinsics. The intrinsic these come from are designed to
16944 // return an integer value, not just an instruction so lower it to the ptest
16945 // or testp pattern and a setcc for the result.
16946 case Intrinsic::x86_sse41_ptestz:
16947 case Intrinsic::x86_sse41_ptestc:
16948 case Intrinsic::x86_sse41_ptestnzc:
16949 case Intrinsic::x86_avx_ptestz_256:
16950 case Intrinsic::x86_avx_ptestc_256:
16951 case Intrinsic::x86_avx_ptestnzc_256:
16952 case Intrinsic::x86_avx_vtestz_ps:
16953 case Intrinsic::x86_avx_vtestc_ps:
16954 case Intrinsic::x86_avx_vtestnzc_ps:
16955 case Intrinsic::x86_avx_vtestz_pd:
16956 case Intrinsic::x86_avx_vtestc_pd:
16957 case Intrinsic::x86_avx_vtestnzc_pd:
16958 case Intrinsic::x86_avx_vtestz_ps_256:
16959 case Intrinsic::x86_avx_vtestc_ps_256:
16960 case Intrinsic::x86_avx_vtestnzc_ps_256:
16961 case Intrinsic::x86_avx_vtestz_pd_256:
16962 case Intrinsic::x86_avx_vtestc_pd_256:
16963 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16964 bool IsTestPacked = false;
16967 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16968 case Intrinsic::x86_avx_vtestz_ps:
16969 case Intrinsic::x86_avx_vtestz_pd:
16970 case Intrinsic::x86_avx_vtestz_ps_256:
16971 case Intrinsic::x86_avx_vtestz_pd_256:
16972 IsTestPacked = true; // Fallthrough
16973 case Intrinsic::x86_sse41_ptestz:
16974 case Intrinsic::x86_avx_ptestz_256:
16976 X86CC = X86::COND_E;
16978 case Intrinsic::x86_avx_vtestc_ps:
16979 case Intrinsic::x86_avx_vtestc_pd:
16980 case Intrinsic::x86_avx_vtestc_ps_256:
16981 case Intrinsic::x86_avx_vtestc_pd_256:
16982 IsTestPacked = true; // Fallthrough
16983 case Intrinsic::x86_sse41_ptestc:
16984 case Intrinsic::x86_avx_ptestc_256:
16986 X86CC = X86::COND_B;
16988 case Intrinsic::x86_avx_vtestnzc_ps:
16989 case Intrinsic::x86_avx_vtestnzc_pd:
16990 case Intrinsic::x86_avx_vtestnzc_ps_256:
16991 case Intrinsic::x86_avx_vtestnzc_pd_256:
16992 IsTestPacked = true; // Fallthrough
16993 case Intrinsic::x86_sse41_ptestnzc:
16994 case Intrinsic::x86_avx_ptestnzc_256:
16996 X86CC = X86::COND_A;
17000 SDValue LHS = Op.getOperand(1);
17001 SDValue RHS = Op.getOperand(2);
17002 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
17003 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
17004 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
17005 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
17006 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17008 case Intrinsic::x86_avx512_kortestz_w:
17009 case Intrinsic::x86_avx512_kortestc_w: {
17010 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
17011 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
17012 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
17013 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
17014 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
17015 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17016 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17019 case Intrinsic::x86_sse42_pcmpistria128:
17020 case Intrinsic::x86_sse42_pcmpestria128:
17021 case Intrinsic::x86_sse42_pcmpistric128:
17022 case Intrinsic::x86_sse42_pcmpestric128:
17023 case Intrinsic::x86_sse42_pcmpistrio128:
17024 case Intrinsic::x86_sse42_pcmpestrio128:
17025 case Intrinsic::x86_sse42_pcmpistris128:
17026 case Intrinsic::x86_sse42_pcmpestris128:
17027 case Intrinsic::x86_sse42_pcmpistriz128:
17028 case Intrinsic::x86_sse42_pcmpestriz128: {
17032 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17033 case Intrinsic::x86_sse42_pcmpistria128:
17034 Opcode = X86ISD::PCMPISTRI;
17035 X86CC = X86::COND_A;
17037 case Intrinsic::x86_sse42_pcmpestria128:
17038 Opcode = X86ISD::PCMPESTRI;
17039 X86CC = X86::COND_A;
17041 case Intrinsic::x86_sse42_pcmpistric128:
17042 Opcode = X86ISD::PCMPISTRI;
17043 X86CC = X86::COND_B;
17045 case Intrinsic::x86_sse42_pcmpestric128:
17046 Opcode = X86ISD::PCMPESTRI;
17047 X86CC = X86::COND_B;
17049 case Intrinsic::x86_sse42_pcmpistrio128:
17050 Opcode = X86ISD::PCMPISTRI;
17051 X86CC = X86::COND_O;
17053 case Intrinsic::x86_sse42_pcmpestrio128:
17054 Opcode = X86ISD::PCMPESTRI;
17055 X86CC = X86::COND_O;
17057 case Intrinsic::x86_sse42_pcmpistris128:
17058 Opcode = X86ISD::PCMPISTRI;
17059 X86CC = X86::COND_S;
17061 case Intrinsic::x86_sse42_pcmpestris128:
17062 Opcode = X86ISD::PCMPESTRI;
17063 X86CC = X86::COND_S;
17065 case Intrinsic::x86_sse42_pcmpistriz128:
17066 Opcode = X86ISD::PCMPISTRI;
17067 X86CC = X86::COND_E;
17069 case Intrinsic::x86_sse42_pcmpestriz128:
17070 Opcode = X86ISD::PCMPESTRI;
17071 X86CC = X86::COND_E;
17074 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17075 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17076 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17077 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17078 DAG.getConstant(X86CC, dl, MVT::i8),
17079 SDValue(PCMP.getNode(), 1));
17080 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17083 case Intrinsic::x86_sse42_pcmpistri128:
17084 case Intrinsic::x86_sse42_pcmpestri128: {
17086 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17087 Opcode = X86ISD::PCMPISTRI;
17089 Opcode = X86ISD::PCMPESTRI;
17091 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17092 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17093 return DAG.getNode(Opcode, dl, VTs, NewOps);
17096 case Intrinsic::x86_seh_lsda: {
17097 // Compute the symbol for the LSDA. We know it'll get emitted later.
17098 MachineFunction &MF = DAG.getMachineFunction();
17099 SDValue Op1 = Op.getOperand(1);
17100 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
17101 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
17102 GlobalValue::getRealLinkageName(Fn->getName()));
17104 // Generate a simple absolute symbol reference. This intrinsic is only
17105 // supported on 32-bit Windows, which isn't PIC.
17106 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
17107 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
17110 case Intrinsic::x86_seh_recoverfp: {
17111 SDValue FnOp = Op.getOperand(1);
17112 SDValue IncomingFPOp = Op.getOperand(2);
17113 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
17114 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
17116 report_fatal_error(
17117 "llvm.x86.seh.recoverfp must take a function as the first argument");
17118 return recoverFramePointer(DAG, Fn, IncomingFPOp);
17121 case Intrinsic::localaddress: {
17122 // Returns one of the stack, base, or frame pointer registers, depending on
17123 // which is used to reference local variables.
17124 MachineFunction &MF = DAG.getMachineFunction();
17125 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17127 if (RegInfo->hasBasePointer(MF))
17128 Reg = RegInfo->getBaseRegister();
17129 else // This function handles the SP or FP case.
17130 Reg = RegInfo->getPtrSizedFrameRegister(MF);
17131 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
17136 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17137 SDValue Src, SDValue Mask, SDValue Base,
17138 SDValue Index, SDValue ScaleOp, SDValue Chain,
17139 const X86Subtarget * Subtarget) {
17141 auto *C = cast<ConstantSDNode>(ScaleOp);
17142 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17143 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17144 Index.getSimpleValueType().getVectorNumElements());
17146 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17148 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17150 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17151 Mask.getSimpleValueType().getSizeInBits());
17153 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17154 // are extracted by EXTRACT_SUBVECTOR.
17155 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17156 DAG.getBitcast(BitcastVT, Mask),
17157 DAG.getIntPtrConstant(0, dl));
17159 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17160 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17161 SDValue Segment = DAG.getRegister(0, MVT::i32);
17162 if (Src.getOpcode() == ISD::UNDEF)
17163 Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl);
17164 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17165 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17166 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17167 return DAG.getMergeValues(RetOps, dl);
17170 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17171 SDValue Src, SDValue Mask, SDValue Base,
17172 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17174 auto *C = cast<ConstantSDNode>(ScaleOp);
17175 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17176 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17177 SDValue Segment = DAG.getRegister(0, MVT::i32);
17178 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17179 Index.getSimpleValueType().getVectorNumElements());
17181 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17183 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17185 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17186 Mask.getSimpleValueType().getSizeInBits());
17188 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17189 // are extracted by EXTRACT_SUBVECTOR.
17190 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17191 DAG.getBitcast(BitcastVT, Mask),
17192 DAG.getIntPtrConstant(0, dl));
17194 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17195 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17196 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17197 return SDValue(Res, 1);
17200 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17201 SDValue Mask, SDValue Base, SDValue Index,
17202 SDValue ScaleOp, SDValue Chain) {
17204 auto *C = cast<ConstantSDNode>(ScaleOp);
17205 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17206 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17207 SDValue Segment = DAG.getRegister(0, MVT::i32);
17209 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17211 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17213 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17215 MaskInReg = DAG.getBitcast(MaskVT, Mask);
17216 //SDVTList VTs = DAG.getVTList(MVT::Other);
17217 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17218 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17219 return SDValue(Res, 0);
17222 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17223 // read performance monitor counters (x86_rdpmc).
17224 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17225 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17226 SmallVectorImpl<SDValue> &Results) {
17227 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17228 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17231 // The ECX register is used to select the index of the performance counter
17233 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17235 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17237 // Reads the content of a 64-bit performance counter and returns it in the
17238 // registers EDX:EAX.
17239 if (Subtarget->is64Bit()) {
17240 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17241 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17244 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17245 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17248 Chain = HI.getValue(1);
17250 if (Subtarget->is64Bit()) {
17251 // The EAX register is loaded with the low-order 32 bits. The EDX register
17252 // is loaded with the supported high-order bits of the counter.
17253 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17254 DAG.getConstant(32, DL, MVT::i8));
17255 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17256 Results.push_back(Chain);
17260 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17261 SDValue Ops[] = { LO, HI };
17262 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17263 Results.push_back(Pair);
17264 Results.push_back(Chain);
17267 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17268 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17269 // also used to custom lower READCYCLECOUNTER nodes.
17270 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17271 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17272 SmallVectorImpl<SDValue> &Results) {
17273 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17274 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17277 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17278 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17279 // and the EAX register is loaded with the low-order 32 bits.
17280 if (Subtarget->is64Bit()) {
17281 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17282 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17285 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17286 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17289 SDValue Chain = HI.getValue(1);
17291 if (Opcode == X86ISD::RDTSCP_DAG) {
17292 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17294 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17295 // the ECX register. Add 'ecx' explicitly to the chain.
17296 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17298 // Explicitly store the content of ECX at the location passed in input
17299 // to the 'rdtscp' intrinsic.
17300 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17301 MachinePointerInfo(), false, false, 0);
17304 if (Subtarget->is64Bit()) {
17305 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17306 // the EAX register is loaded with the low-order 32 bits.
17307 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17308 DAG.getConstant(32, DL, MVT::i8));
17309 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17310 Results.push_back(Chain);
17314 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17315 SDValue Ops[] = { LO, HI };
17316 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17317 Results.push_back(Pair);
17318 Results.push_back(Chain);
17321 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17322 SelectionDAG &DAG) {
17323 SmallVector<SDValue, 2> Results;
17325 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17327 return DAG.getMergeValues(Results, DL);
17330 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
17331 MachineFunction &MF = DAG.getMachineFunction();
17332 SDValue Chain = Op.getOperand(0);
17333 SDValue RegNode = Op.getOperand(2);
17334 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
17336 report_fatal_error("EH registrations only live in functions using WinEH");
17338 // Cast the operand to an alloca, and remember the frame index.
17339 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
17341 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
17342 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
17344 // Return the chain operand without making any DAG nodes.
17348 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
17349 /// return truncate Store/MaskedStore Node
17350 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17354 SDValue Mask = Op.getOperand(4);
17355 SDValue DataToTruncate = Op.getOperand(3);
17356 SDValue Addr = Op.getOperand(2);
17357 SDValue Chain = Op.getOperand(0);
17359 MVT VT = DataToTruncate.getSimpleValueType();
17360 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17362 if (isAllOnesConstant(Mask)) // return just a truncate store
17363 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17364 MachinePointerInfo(), SVT, false, false,
17365 SVT.getScalarSizeInBits()/8);
17367 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17368 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17369 Mask.getSimpleValueType().getSizeInBits());
17370 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17371 // are extracted by EXTRACT_SUBVECTOR.
17372 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17373 DAG.getBitcast(BitcastVT, Mask),
17374 DAG.getIntPtrConstant(0, dl));
17376 MachineMemOperand *MMO = DAG.getMachineFunction().
17377 getMachineMemOperand(MachinePointerInfo(),
17378 MachineMemOperand::MOStore, SVT.getStoreSize(),
17379 SVT.getScalarSizeInBits()/8);
17381 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17382 VMask, SVT, MMO, true);
17385 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17386 SelectionDAG &DAG) {
17387 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17389 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17391 if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17392 return MarkEHRegistrationNode(Op, DAG);
17393 if (IntNo == llvm::Intrinsic::x86_flags_read_u32 ||
17394 IntNo == llvm::Intrinsic::x86_flags_read_u64 ||
17395 IntNo == llvm::Intrinsic::x86_flags_write_u32 ||
17396 IntNo == llvm::Intrinsic::x86_flags_write_u64) {
17397 // We need a frame pointer because this will get lowered to a PUSH/POP
17399 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17400 MFI->setHasOpaqueSPAdjustment(true);
17401 // Don't do anything here, we will expand these intrinsics out later
17402 // during ExpandISelPseudos in EmitInstrWithCustomInserter.
17409 switch(IntrData->Type) {
17410 default: llvm_unreachable("Unknown Intrinsic Type");
17413 // Emit the node with the right value type.
17414 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17415 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17417 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17418 // Otherwise return the value from Rand, which is always 0, casted to i32.
17419 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17420 DAG.getConstant(1, dl, Op->getValueType(1)),
17421 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17422 SDValue(Result.getNode(), 1) };
17423 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17424 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17427 // Return { result, isValid, chain }.
17428 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17429 SDValue(Result.getNode(), 2));
17432 //gather(v1, mask, index, base, scale);
17433 SDValue Chain = Op.getOperand(0);
17434 SDValue Src = Op.getOperand(2);
17435 SDValue Base = Op.getOperand(3);
17436 SDValue Index = Op.getOperand(4);
17437 SDValue Mask = Op.getOperand(5);
17438 SDValue Scale = Op.getOperand(6);
17439 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17443 //scatter(base, mask, index, v1, scale);
17444 SDValue Chain = Op.getOperand(0);
17445 SDValue Base = Op.getOperand(2);
17446 SDValue Mask = Op.getOperand(3);
17447 SDValue Index = Op.getOperand(4);
17448 SDValue Src = Op.getOperand(5);
17449 SDValue Scale = Op.getOperand(6);
17450 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17454 SDValue Hint = Op.getOperand(6);
17455 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17456 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17457 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17458 SDValue Chain = Op.getOperand(0);
17459 SDValue Mask = Op.getOperand(2);
17460 SDValue Index = Op.getOperand(3);
17461 SDValue Base = Op.getOperand(4);
17462 SDValue Scale = Op.getOperand(5);
17463 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17465 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17467 SmallVector<SDValue, 2> Results;
17468 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17470 return DAG.getMergeValues(Results, dl);
17472 // Read Performance Monitoring Counters.
17474 SmallVector<SDValue, 2> Results;
17475 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17476 return DAG.getMergeValues(Results, dl);
17478 // XTEST intrinsics.
17480 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17481 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17482 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17483 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17485 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17486 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17487 Ret, SDValue(InTrans.getNode(), 1));
17491 SmallVector<SDValue, 2> Results;
17492 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17493 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17494 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17495 DAG.getConstant(-1, dl, MVT::i8));
17496 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17497 Op.getOperand(4), GenCF.getValue(1));
17498 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17499 Op.getOperand(5), MachinePointerInfo(),
17501 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17502 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17504 Results.push_back(SetCC);
17505 Results.push_back(Store);
17506 return DAG.getMergeValues(Results, dl);
17508 case COMPRESS_TO_MEM: {
17510 SDValue Mask = Op.getOperand(4);
17511 SDValue DataToCompress = Op.getOperand(3);
17512 SDValue Addr = Op.getOperand(2);
17513 SDValue Chain = Op.getOperand(0);
17515 MVT VT = DataToCompress.getSimpleValueType();
17516 if (isAllOnesConstant(Mask)) // return just a store
17517 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17518 MachinePointerInfo(), false, false,
17519 VT.getScalarSizeInBits()/8);
17521 SDValue Compressed =
17522 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17523 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17524 return DAG.getStore(Chain, dl, Compressed, Addr,
17525 MachinePointerInfo(), false, false,
17526 VT.getScalarSizeInBits()/8);
17528 case TRUNCATE_TO_MEM_VI8:
17529 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17530 case TRUNCATE_TO_MEM_VI16:
17531 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17532 case TRUNCATE_TO_MEM_VI32:
17533 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17534 case EXPAND_FROM_MEM: {
17536 SDValue Mask = Op.getOperand(4);
17537 SDValue PassThru = Op.getOperand(3);
17538 SDValue Addr = Op.getOperand(2);
17539 SDValue Chain = Op.getOperand(0);
17540 MVT VT = Op.getSimpleValueType();
17542 if (isAllOnesConstant(Mask)) // return just a load
17543 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17544 false, VT.getScalarSizeInBits()/8);
17546 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17547 false, false, false,
17548 VT.getScalarSizeInBits()/8);
17550 SDValue Results[] = {
17551 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17552 Mask, PassThru, Subtarget, DAG), Chain};
17553 return DAG.getMergeValues(Results, dl);
17558 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17559 SelectionDAG &DAG) const {
17560 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17561 MFI->setReturnAddressIsTaken(true);
17563 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17566 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17568 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17571 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17572 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17573 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17574 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17575 DAG.getNode(ISD::ADD, dl, PtrVT,
17576 FrameAddr, Offset),
17577 MachinePointerInfo(), false, false, false, 0);
17580 // Just load the return address.
17581 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17582 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17583 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17586 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17587 MachineFunction &MF = DAG.getMachineFunction();
17588 MachineFrameInfo *MFI = MF.getFrameInfo();
17589 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17590 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17591 EVT VT = Op.getValueType();
17593 MFI->setFrameAddressIsTaken(true);
17595 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17596 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17597 // is not possible to crawl up the stack without looking at the unwind codes
17599 int FrameAddrIndex = FuncInfo->getFAIndex();
17600 if (!FrameAddrIndex) {
17601 // Set up a frame object for the return address.
17602 unsigned SlotSize = RegInfo->getSlotSize();
17603 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17604 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17605 FuncInfo->setFAIndex(FrameAddrIndex);
17607 return DAG.getFrameIndex(FrameAddrIndex, VT);
17610 unsigned FrameReg =
17611 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17612 SDLoc dl(Op); // FIXME probably not meaningful
17613 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17614 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17615 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17616 "Invalid Frame Register!");
17617 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17619 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17620 MachinePointerInfo(),
17621 false, false, false, 0);
17625 // FIXME? Maybe this could be a TableGen attribute on some registers and
17626 // this table could be generated automatically from RegInfo.
17627 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17628 SelectionDAG &DAG) const {
17629 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17630 const MachineFunction &MF = DAG.getMachineFunction();
17632 unsigned Reg = StringSwitch<unsigned>(RegName)
17633 .Case("esp", X86::ESP)
17634 .Case("rsp", X86::RSP)
17635 .Case("ebp", X86::EBP)
17636 .Case("rbp", X86::RBP)
17639 if (Reg == X86::EBP || Reg == X86::RBP) {
17640 if (!TFI.hasFP(MF))
17641 report_fatal_error("register " + StringRef(RegName) +
17642 " is allocatable: function has no frame pointer");
17645 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17646 unsigned FrameReg =
17647 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17648 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17649 "Invalid Frame Register!");
17657 report_fatal_error("Invalid register name global variable");
17660 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17661 SelectionDAG &DAG) const {
17662 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17663 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17666 unsigned X86TargetLowering::getExceptionPointerRegister(
17667 const Constant *PersonalityFn) const {
17668 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17669 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17671 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17674 unsigned X86TargetLowering::getExceptionSelectorRegister(
17675 const Constant *PersonalityFn) const {
17676 // Funclet personalities don't use selectors (the runtime does the selection).
17677 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17678 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17681 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17682 SDValue Chain = Op.getOperand(0);
17683 SDValue Offset = Op.getOperand(1);
17684 SDValue Handler = Op.getOperand(2);
17687 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17688 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17689 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17690 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17691 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17692 "Invalid Frame Register!");
17693 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17694 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17696 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17697 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17699 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17700 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17702 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17704 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17705 DAG.getRegister(StoreAddrReg, PtrVT));
17708 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17709 SelectionDAG &DAG) const {
17711 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17712 DAG.getVTList(MVT::i32, MVT::Other),
17713 Op.getOperand(0), Op.getOperand(1));
17716 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17717 SelectionDAG &DAG) const {
17719 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17720 Op.getOperand(0), Op.getOperand(1));
17723 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17724 return Op.getOperand(0);
17727 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17728 SelectionDAG &DAG) const {
17729 SDValue Root = Op.getOperand(0);
17730 SDValue Trmp = Op.getOperand(1); // trampoline
17731 SDValue FPtr = Op.getOperand(2); // nested function
17732 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17735 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17736 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17738 if (Subtarget->is64Bit()) {
17739 SDValue OutChains[6];
17741 // Large code-model.
17742 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17743 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17745 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17746 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17748 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17750 // Load the pointer to the nested function into R11.
17751 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17752 SDValue Addr = Trmp;
17753 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17754 Addr, MachinePointerInfo(TrmpAddr),
17757 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17758 DAG.getConstant(2, dl, MVT::i64));
17759 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17760 MachinePointerInfo(TrmpAddr, 2),
17763 // Load the 'nest' parameter value into R10.
17764 // R10 is specified in X86CallingConv.td
17765 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17766 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17767 DAG.getConstant(10, dl, MVT::i64));
17768 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17769 Addr, MachinePointerInfo(TrmpAddr, 10),
17772 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17773 DAG.getConstant(12, dl, MVT::i64));
17774 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17775 MachinePointerInfo(TrmpAddr, 12),
17778 // Jump to the nested function.
17779 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17780 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17781 DAG.getConstant(20, dl, MVT::i64));
17782 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17783 Addr, MachinePointerInfo(TrmpAddr, 20),
17786 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17787 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17788 DAG.getConstant(22, dl, MVT::i64));
17789 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17790 Addr, MachinePointerInfo(TrmpAddr, 22),
17793 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17795 const Function *Func =
17796 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17797 CallingConv::ID CC = Func->getCallingConv();
17802 llvm_unreachable("Unsupported calling convention");
17803 case CallingConv::C:
17804 case CallingConv::X86_StdCall: {
17805 // Pass 'nest' parameter in ECX.
17806 // Must be kept in sync with X86CallingConv.td
17807 NestReg = X86::ECX;
17809 // Check that ECX wasn't needed by an 'inreg' parameter.
17810 FunctionType *FTy = Func->getFunctionType();
17811 const AttributeSet &Attrs = Func->getAttributes();
17813 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17814 unsigned InRegCount = 0;
17817 for (FunctionType::param_iterator I = FTy->param_begin(),
17818 E = FTy->param_end(); I != E; ++I, ++Idx)
17819 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17820 auto &DL = DAG.getDataLayout();
17821 // FIXME: should only count parameters that are lowered to integers.
17822 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17825 if (InRegCount > 2) {
17826 report_fatal_error("Nest register in use - reduce number of inreg"
17832 case CallingConv::X86_FastCall:
17833 case CallingConv::X86_ThisCall:
17834 case CallingConv::Fast:
17835 // Pass 'nest' parameter in EAX.
17836 // Must be kept in sync with X86CallingConv.td
17837 NestReg = X86::EAX;
17841 SDValue OutChains[4];
17842 SDValue Addr, Disp;
17844 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17845 DAG.getConstant(10, dl, MVT::i32));
17846 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17848 // This is storing the opcode for MOV32ri.
17849 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17850 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17851 OutChains[0] = DAG.getStore(Root, dl,
17852 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17853 Trmp, MachinePointerInfo(TrmpAddr),
17856 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17857 DAG.getConstant(1, dl, MVT::i32));
17858 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17859 MachinePointerInfo(TrmpAddr, 1),
17862 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17863 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17864 DAG.getConstant(5, dl, MVT::i32));
17865 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17866 Addr, MachinePointerInfo(TrmpAddr, 5),
17869 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17870 DAG.getConstant(6, dl, MVT::i32));
17871 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17872 MachinePointerInfo(TrmpAddr, 6),
17875 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17879 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17880 SelectionDAG &DAG) const {
17882 The rounding mode is in bits 11:10 of FPSR, and has the following
17884 00 Round to nearest
17889 FLT_ROUNDS, on the other hand, expects the following:
17896 To perform the conversion, we do:
17897 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17900 MachineFunction &MF = DAG.getMachineFunction();
17901 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17902 unsigned StackAlignment = TFI.getStackAlignment();
17903 MVT VT = Op.getSimpleValueType();
17906 // Save FP Control Word to stack slot
17907 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17908 SDValue StackSlot =
17909 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17911 MachineMemOperand *MMO =
17912 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17913 MachineMemOperand::MOStore, 2, 2);
17915 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17916 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17917 DAG.getVTList(MVT::Other),
17918 Ops, MVT::i16, MMO);
17920 // Load FP Control Word from stack slot
17921 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17922 MachinePointerInfo(), false, false, false, 0);
17924 // Transform as necessary
17926 DAG.getNode(ISD::SRL, DL, MVT::i16,
17927 DAG.getNode(ISD::AND, DL, MVT::i16,
17928 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17929 DAG.getConstant(11, DL, MVT::i8));
17931 DAG.getNode(ISD::SRL, DL, MVT::i16,
17932 DAG.getNode(ISD::AND, DL, MVT::i16,
17933 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17934 DAG.getConstant(9, DL, MVT::i8));
17937 DAG.getNode(ISD::AND, DL, MVT::i16,
17938 DAG.getNode(ISD::ADD, DL, MVT::i16,
17939 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17940 DAG.getConstant(1, DL, MVT::i16)),
17941 DAG.getConstant(3, DL, MVT::i16));
17943 return DAG.getNode((VT.getSizeInBits() < 16 ?
17944 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17947 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17949 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17950 // to 512-bit vector.
17951 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17952 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17953 // split the vector, perform operation on it's Lo a Hi part and
17954 // concatenate the results.
17955 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17957 MVT VT = Op.getSimpleValueType();
17958 MVT EltVT = VT.getVectorElementType();
17959 unsigned NumElems = VT.getVectorNumElements();
17961 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17962 // Extend to 512 bit vector.
17963 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17964 "Unsupported value type for operation");
17966 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17967 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17968 DAG.getUNDEF(NewVT),
17970 DAG.getIntPtrConstant(0, dl));
17971 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17973 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17974 DAG.getIntPtrConstant(0, dl));
17977 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17978 "Unsupported element type");
17980 if (16 < NumElems) {
17981 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17983 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17984 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17986 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17987 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17989 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17992 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17994 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17995 "Unsupported value type for operation");
17997 // Use native supported vector instruction vplzcntd.
17998 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17999 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
18000 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
18001 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
18003 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
18006 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
18007 SelectionDAG &DAG) {
18008 MVT VT = Op.getSimpleValueType();
18010 unsigned NumBits = VT.getSizeInBits();
18013 if (VT.isVector() && Subtarget->hasAVX512())
18014 return LowerVectorCTLZ_AVX512(Op, DAG);
18016 Op = Op.getOperand(0);
18017 if (VT == MVT::i8) {
18018 // Zero extend to i32 since there is not an i8 bsr.
18020 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18023 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
18024 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18025 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18027 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18030 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
18031 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18034 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18036 // Finally xor with NumBits-1.
18037 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18038 DAG.getConstant(NumBits - 1, dl, OpVT));
18041 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18045 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
18046 SelectionDAG &DAG) {
18047 MVT VT = Op.getSimpleValueType();
18049 unsigned NumBits = VT.getSizeInBits();
18052 Op = Op.getOperand(0);
18053 if (VT == MVT::i8) {
18054 // Zero extend to i32 since there is not an i8 bsr.
18056 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18059 // Issue a bsr (scan bits in reverse).
18060 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18061 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18063 // And xor with NumBits-1.
18064 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18065 DAG.getConstant(NumBits - 1, dl, OpVT));
18068 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18072 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18073 MVT VT = Op.getSimpleValueType();
18074 unsigned NumBits = VT.getScalarSizeInBits();
18077 if (VT.isVector()) {
18078 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18080 SDValue N0 = Op.getOperand(0);
18081 SDValue Zero = DAG.getConstant(0, dl, VT);
18083 // lsb(x) = (x & -x)
18084 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
18085 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
18087 // cttz_undef(x) = (width - 1) - ctlz(lsb)
18088 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
18089 TLI.isOperationLegal(ISD::CTLZ, VT)) {
18090 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
18091 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
18092 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
18095 // cttz(x) = ctpop(lsb - 1)
18096 SDValue One = DAG.getConstant(1, dl, VT);
18097 return DAG.getNode(ISD::CTPOP, dl, VT,
18098 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
18101 assert(Op.getOpcode() == ISD::CTTZ &&
18102 "Only scalar CTTZ requires custom lowering");
18104 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18105 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18106 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
18108 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18111 DAG.getConstant(NumBits, dl, VT),
18112 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18115 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18118 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18119 // ones, and then concatenate the result back.
18120 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18121 MVT VT = Op.getSimpleValueType();
18123 assert(VT.is256BitVector() && VT.isInteger() &&
18124 "Unsupported value type for operation");
18126 unsigned NumElems = VT.getVectorNumElements();
18129 // Extract the LHS vectors
18130 SDValue LHS = Op.getOperand(0);
18131 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18132 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18134 // Extract the RHS vectors
18135 SDValue RHS = Op.getOperand(1);
18136 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18137 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18139 MVT EltVT = VT.getVectorElementType();
18140 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18142 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18143 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18144 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18147 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18148 if (Op.getValueType() == MVT::i1)
18149 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18150 Op.getOperand(0), Op.getOperand(1));
18151 assert(Op.getSimpleValueType().is256BitVector() &&
18152 Op.getSimpleValueType().isInteger() &&
18153 "Only handle AVX 256-bit vector integer operation");
18154 return Lower256IntArith(Op, DAG);
18157 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18158 if (Op.getValueType() == MVT::i1)
18159 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18160 Op.getOperand(0), Op.getOperand(1));
18161 assert(Op.getSimpleValueType().is256BitVector() &&
18162 Op.getSimpleValueType().isInteger() &&
18163 "Only handle AVX 256-bit vector integer operation");
18164 return Lower256IntArith(Op, DAG);
18167 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
18168 assert(Op.getSimpleValueType().is256BitVector() &&
18169 Op.getSimpleValueType().isInteger() &&
18170 "Only handle AVX 256-bit vector integer operation");
18171 return Lower256IntArith(Op, DAG);
18174 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18175 SelectionDAG &DAG) {
18177 MVT VT = Op.getSimpleValueType();
18180 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
18182 // Decompose 256-bit ops into smaller 128-bit ops.
18183 if (VT.is256BitVector() && !Subtarget->hasInt256())
18184 return Lower256IntArith(Op, DAG);
18186 SDValue A = Op.getOperand(0);
18187 SDValue B = Op.getOperand(1);
18189 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
18190 // pairs, multiply and truncate.
18191 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
18192 if (Subtarget->hasInt256()) {
18193 if (VT == MVT::v32i8) {
18194 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
18195 SDValue Lo = DAG.getIntPtrConstant(0, dl);
18196 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
18197 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
18198 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
18199 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
18200 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
18201 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18202 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
18203 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
18206 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
18207 return DAG.getNode(
18208 ISD::TRUNCATE, dl, VT,
18209 DAG.getNode(ISD::MUL, dl, ExVT,
18210 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
18211 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
18214 assert(VT == MVT::v16i8 &&
18215 "Pre-AVX2 support only supports v16i8 multiplication");
18216 MVT ExVT = MVT::v8i16;
18218 // Extract the lo parts and sign extend to i16
18220 if (Subtarget->hasSSE41()) {
18221 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
18222 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
18224 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
18225 -1, 4, -1, 5, -1, 6, -1, 7};
18226 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18227 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18228 ALo = DAG.getBitcast(ExVT, ALo);
18229 BLo = DAG.getBitcast(ExVT, BLo);
18230 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
18231 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
18234 // Extract the hi parts and sign extend to i16
18236 if (Subtarget->hasSSE41()) {
18237 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
18238 -1, -1, -1, -1, -1, -1, -1, -1};
18239 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18240 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18241 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
18242 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
18244 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
18245 -1, 12, -1, 13, -1, 14, -1, 15};
18246 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18247 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18248 AHi = DAG.getBitcast(ExVT, AHi);
18249 BHi = DAG.getBitcast(ExVT, BHi);
18250 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
18251 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
18254 // Multiply, mask the lower 8bits of the lo/hi results and pack
18255 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
18256 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
18257 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
18258 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
18259 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18262 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18263 if (VT == MVT::v4i32) {
18264 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18265 "Should not custom lower when pmuldq is available!");
18267 // Extract the odd parts.
18268 static const int UnpackMask[] = { 1, -1, 3, -1 };
18269 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18270 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18272 // Multiply the even parts.
18273 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18274 // Now multiply odd parts.
18275 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18277 Evens = DAG.getBitcast(VT, Evens);
18278 Odds = DAG.getBitcast(VT, Odds);
18280 // Merge the two vectors back together with a shuffle. This expands into 2
18282 static const int ShufMask[] = { 0, 4, 2, 6 };
18283 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18286 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18287 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18289 // Ahi = psrlqi(a, 32);
18290 // Bhi = psrlqi(b, 32);
18292 // AloBlo = pmuludq(a, b);
18293 // AloBhi = pmuludq(a, Bhi);
18294 // AhiBlo = pmuludq(Ahi, b);
18296 // AloBhi = psllqi(AloBhi, 32);
18297 // AhiBlo = psllqi(AhiBlo, 32);
18298 // return AloBlo + AloBhi + AhiBlo;
18300 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18301 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18303 SDValue AhiBlo = Ahi;
18304 SDValue AloBhi = Bhi;
18305 // Bit cast to 32-bit vectors for MULUDQ
18306 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18307 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18308 A = DAG.getBitcast(MulVT, A);
18309 B = DAG.getBitcast(MulVT, B);
18310 Ahi = DAG.getBitcast(MulVT, Ahi);
18311 Bhi = DAG.getBitcast(MulVT, Bhi);
18313 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18314 // After shifting right const values the result may be all-zero.
18315 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
18316 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18317 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18319 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
18320 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18321 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18324 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18325 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18328 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18329 assert(Subtarget->isTargetWin64() && "Unexpected target");
18330 EVT VT = Op.getValueType();
18331 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18332 "Unexpected return type for lowering");
18336 switch (Op->getOpcode()) {
18337 default: llvm_unreachable("Unexpected request for libcall!");
18338 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18339 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18340 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18341 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18342 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18343 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18347 SDValue InChain = DAG.getEntryNode();
18349 TargetLowering::ArgListTy Args;
18350 TargetLowering::ArgListEntry Entry;
18351 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18352 EVT ArgVT = Op->getOperand(i).getValueType();
18353 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18354 "Unexpected argument type for lowering");
18355 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18356 Entry.Node = StackPtr;
18357 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18359 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18360 Entry.Ty = PointerType::get(ArgTy,0);
18361 Entry.isSExt = false;
18362 Entry.isZExt = false;
18363 Args.push_back(Entry);
18366 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18367 getPointerTy(DAG.getDataLayout()));
18369 TargetLowering::CallLoweringInfo CLI(DAG);
18370 CLI.setDebugLoc(dl).setChain(InChain)
18371 .setCallee(getLibcallCallingConv(LC),
18372 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18373 Callee, std::move(Args), 0)
18374 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18376 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18377 return DAG.getBitcast(VT, CallInfo.first);
18380 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18381 SelectionDAG &DAG) {
18382 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18383 MVT VT = Op0.getSimpleValueType();
18386 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18387 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18389 // PMULxD operations multiply each even value (starting at 0) of LHS with
18390 // the related value of RHS and produce a widen result.
18391 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18392 // => <2 x i64> <ae|cg>
18394 // In other word, to have all the results, we need to perform two PMULxD:
18395 // 1. one with the even values.
18396 // 2. one with the odd values.
18397 // To achieve #2, with need to place the odd values at an even position.
18399 // Place the odd value at an even position (basically, shift all values 1
18400 // step to the left):
18401 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18402 // <a|b|c|d> => <b|undef|d|undef>
18403 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18404 // <e|f|g|h> => <f|undef|h|undef>
18405 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18407 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18409 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18410 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18412 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18413 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18414 // => <2 x i64> <ae|cg>
18415 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18416 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18417 // => <2 x i64> <bf|dh>
18418 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18420 // Shuffle it back into the right order.
18421 SDValue Highs, Lows;
18422 if (VT == MVT::v8i32) {
18423 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18424 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18425 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18426 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18428 const int HighMask[] = {1, 5, 3, 7};
18429 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18430 const int LowMask[] = {0, 4, 2, 6};
18431 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18434 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18435 // unsigned multiply.
18436 if (IsSigned && !Subtarget->hasSSE41()) {
18437 SDValue ShAmt = DAG.getConstant(
18439 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18440 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18441 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18442 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18443 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18445 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18446 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18449 // The first result of MUL_LOHI is actually the low value, followed by the
18451 SDValue Ops[] = {Lows, Highs};
18452 return DAG.getMergeValues(Ops, dl);
18455 // Return true if the required (according to Opcode) shift-imm form is natively
18456 // supported by the Subtarget
18457 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18459 if (VT.getScalarSizeInBits() < 16)
18462 if (VT.is512BitVector() &&
18463 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18466 bool LShift = VT.is128BitVector() ||
18467 (VT.is256BitVector() && Subtarget->hasInt256());
18469 bool AShift = LShift && (Subtarget->hasVLX() ||
18470 (VT != MVT::v2i64 && VT != MVT::v4i64));
18471 return (Opcode == ISD::SRA) ? AShift : LShift;
18474 // The shift amount is a variable, but it is the same for all vector lanes.
18475 // These instructions are defined together with shift-immediate.
18477 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18479 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18482 // Return true if the required (according to Opcode) variable-shift form is
18483 // natively supported by the Subtarget
18484 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18487 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18490 // vXi16 supported only on AVX-512, BWI
18491 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18494 if (VT.is512BitVector() || Subtarget->hasVLX())
18497 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18498 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18499 return (Opcode == ISD::SRA) ? AShift : LShift;
18502 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18503 const X86Subtarget *Subtarget) {
18504 MVT VT = Op.getSimpleValueType();
18506 SDValue R = Op.getOperand(0);
18507 SDValue Amt = Op.getOperand(1);
18509 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18510 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18512 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18513 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18514 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18515 SDValue Ex = DAG.getBitcast(ExVT, R);
18517 if (ShiftAmt >= 32) {
18518 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18520 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18521 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18522 ShiftAmt - 32, DAG);
18523 if (VT == MVT::v2i64)
18524 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18525 if (VT == MVT::v4i64)
18526 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18527 {9, 1, 11, 3, 13, 5, 15, 7});
18529 // SRA upper i32, SHL whole i64 and select lower i32.
18530 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18533 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18534 Lower = DAG.getBitcast(ExVT, Lower);
18535 if (VT == MVT::v2i64)
18536 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18537 if (VT == MVT::v4i64)
18538 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18539 {8, 1, 10, 3, 12, 5, 14, 7});
18541 return DAG.getBitcast(VT, Ex);
18544 // Optimize shl/srl/sra with constant shift amount.
18545 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18546 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18547 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18549 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18550 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18552 // i64 SRA needs to be performed as partial shifts.
18553 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18554 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18555 return ArithmeticShiftRight64(ShiftAmt);
18557 if (VT == MVT::v16i8 ||
18558 (Subtarget->hasInt256() && VT == MVT::v32i8) ||
18559 VT == MVT::v64i8) {
18560 unsigned NumElts = VT.getVectorNumElements();
18561 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18563 // Simple i8 add case
18564 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18565 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18567 // ashr(R, 7) === cmp_slt(R, 0)
18568 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18569 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18570 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18573 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18574 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18577 if (Op.getOpcode() == ISD::SHL) {
18578 // Make a large shift.
18579 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18581 SHL = DAG.getBitcast(VT, SHL);
18582 // Zero out the rightmost bits.
18583 return DAG.getNode(ISD::AND, dl, VT, SHL,
18584 DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, VT));
18586 if (Op.getOpcode() == ISD::SRL) {
18587 // Make a large shift.
18588 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18590 SRL = DAG.getBitcast(VT, SRL);
18591 // Zero out the leftmost bits.
18592 return DAG.getNode(ISD::AND, dl, VT, SRL,
18593 DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, VT));
18595 if (Op.getOpcode() == ISD::SRA) {
18596 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18597 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18599 SDValue Mask = DAG.getConstant(128 >> ShiftAmt, dl, VT);
18600 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18601 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18604 llvm_unreachable("Unknown shift opcode.");
18609 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18610 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18611 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18613 // Peek through any splat that was introduced for i64 shift vectorization.
18614 int SplatIndex = -1;
18615 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18616 if (SVN->isSplat()) {
18617 SplatIndex = SVN->getSplatIndex();
18618 Amt = Amt.getOperand(0);
18619 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18620 "Splat shuffle referencing second operand");
18623 if (Amt.getOpcode() != ISD::BITCAST ||
18624 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18627 Amt = Amt.getOperand(0);
18628 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18629 VT.getVectorNumElements();
18630 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18631 uint64_t ShiftAmt = 0;
18632 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18633 for (unsigned i = 0; i != Ratio; ++i) {
18634 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18638 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18641 // Check remaining shift amounts (if not a splat).
18642 if (SplatIndex < 0) {
18643 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18644 uint64_t ShAmt = 0;
18645 for (unsigned j = 0; j != Ratio; ++j) {
18646 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18650 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18652 if (ShAmt != ShiftAmt)
18657 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18658 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18660 if (Op.getOpcode() == ISD::SRA)
18661 return ArithmeticShiftRight64(ShiftAmt);
18667 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18668 const X86Subtarget* Subtarget) {
18669 MVT VT = Op.getSimpleValueType();
18671 SDValue R = Op.getOperand(0);
18672 SDValue Amt = Op.getOperand(1);
18674 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18675 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18677 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18678 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18680 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18682 MVT EltVT = VT.getVectorElementType();
18684 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18685 // Check if this build_vector node is doing a splat.
18686 // If so, then set BaseShAmt equal to the splat value.
18687 BaseShAmt = BV->getSplatValue();
18688 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18689 BaseShAmt = SDValue();
18691 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18692 Amt = Amt.getOperand(0);
18694 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18695 if (SVN && SVN->isSplat()) {
18696 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18697 SDValue InVec = Amt.getOperand(0);
18698 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18699 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18700 "Unexpected shuffle index found!");
18701 BaseShAmt = InVec.getOperand(SplatIdx);
18702 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18703 if (ConstantSDNode *C =
18704 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18705 if (C->getZExtValue() == SplatIdx)
18706 BaseShAmt = InVec.getOperand(1);
18711 // Avoid introducing an extract element from a shuffle.
18712 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18713 DAG.getIntPtrConstant(SplatIdx, dl));
18717 if (BaseShAmt.getNode()) {
18718 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18719 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18720 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18721 else if (EltVT.bitsLT(MVT::i32))
18722 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18724 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18728 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18729 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18730 Amt.getOpcode() == ISD::BITCAST &&
18731 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18732 Amt = Amt.getOperand(0);
18733 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18734 VT.getVectorNumElements();
18735 std::vector<SDValue> Vals(Ratio);
18736 for (unsigned i = 0; i != Ratio; ++i)
18737 Vals[i] = Amt.getOperand(i);
18738 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18739 for (unsigned j = 0; j != Ratio; ++j)
18740 if (Vals[j] != Amt.getOperand(i + j))
18744 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18745 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18750 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18751 SelectionDAG &DAG) {
18752 MVT VT = Op.getSimpleValueType();
18754 SDValue R = Op.getOperand(0);
18755 SDValue Amt = Op.getOperand(1);
18757 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18758 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18760 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18763 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18766 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18769 // XOP has 128-bit variable logical/arithmetic shifts.
18770 // +ve/-ve Amt = shift left/right.
18771 if (Subtarget->hasXOP() &&
18772 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18773 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18774 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18775 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18776 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18778 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18779 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18780 if (Op.getOpcode() == ISD::SRA)
18781 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18784 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18785 // shifts per-lane and then shuffle the partial results back together.
18786 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18787 // Splat the shift amounts so the scalar shifts above will catch it.
18788 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18789 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18790 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18791 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18792 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18795 // i64 vector arithmetic shift can be emulated with the transform:
18796 // M = lshr(SIGN_BIT, Amt)
18797 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18798 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18799 Op.getOpcode() == ISD::SRA) {
18800 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18801 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18802 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18803 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18804 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18808 // If possible, lower this packed shift into a vector multiply instead of
18809 // expanding it into a sequence of scalar shifts.
18810 // Do this only if the vector shift count is a constant build_vector.
18811 if (Op.getOpcode() == ISD::SHL &&
18812 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18813 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18814 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18815 SmallVector<SDValue, 8> Elts;
18816 MVT SVT = VT.getVectorElementType();
18817 unsigned SVTBits = SVT.getSizeInBits();
18818 APInt One(SVTBits, 1);
18819 unsigned NumElems = VT.getVectorNumElements();
18821 for (unsigned i=0; i !=NumElems; ++i) {
18822 SDValue Op = Amt->getOperand(i);
18823 if (Op->getOpcode() == ISD::UNDEF) {
18824 Elts.push_back(Op);
18828 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18829 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18830 uint64_t ShAmt = C.getZExtValue();
18831 if (ShAmt >= SVTBits) {
18832 Elts.push_back(DAG.getUNDEF(SVT));
18835 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18837 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18838 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18841 // Lower SHL with variable shift amount.
18842 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18843 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18845 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18846 DAG.getConstant(0x3f800000U, dl, VT));
18847 Op = DAG.getBitcast(MVT::v4f32, Op);
18848 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18849 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18852 // If possible, lower this shift as a sequence of two shifts by
18853 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18855 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18857 // Could be rewritten as:
18858 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18860 // The advantage is that the two shifts from the example would be
18861 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18862 // the vector shift into four scalar shifts plus four pairs of vector
18864 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18865 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18866 unsigned TargetOpcode = X86ISD::MOVSS;
18867 bool CanBeSimplified;
18868 // The splat value for the first packed shift (the 'X' from the example).
18869 SDValue Amt1 = Amt->getOperand(0);
18870 // The splat value for the second packed shift (the 'Y' from the example).
18871 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18872 Amt->getOperand(2);
18874 // See if it is possible to replace this node with a sequence of
18875 // two shifts followed by a MOVSS/MOVSD
18876 if (VT == MVT::v4i32) {
18877 // Check if it is legal to use a MOVSS.
18878 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18879 Amt2 == Amt->getOperand(3);
18880 if (!CanBeSimplified) {
18881 // Otherwise, check if we can still simplify this node using a MOVSD.
18882 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18883 Amt->getOperand(2) == Amt->getOperand(3);
18884 TargetOpcode = X86ISD::MOVSD;
18885 Amt2 = Amt->getOperand(2);
18888 // Do similar checks for the case where the machine value type
18890 CanBeSimplified = Amt1 == Amt->getOperand(1);
18891 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18892 CanBeSimplified = Amt2 == Amt->getOperand(i);
18894 if (!CanBeSimplified) {
18895 TargetOpcode = X86ISD::MOVSD;
18896 CanBeSimplified = true;
18897 Amt2 = Amt->getOperand(4);
18898 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18899 CanBeSimplified = Amt1 == Amt->getOperand(i);
18900 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18901 CanBeSimplified = Amt2 == Amt->getOperand(j);
18905 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18906 isa<ConstantSDNode>(Amt2)) {
18907 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18908 MVT CastVT = MVT::v4i32;
18910 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18911 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18913 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18914 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18915 if (TargetOpcode == X86ISD::MOVSD)
18916 CastVT = MVT::v2i64;
18917 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18918 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18919 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18921 return DAG.getBitcast(VT, Result);
18925 // v4i32 Non Uniform Shifts.
18926 // If the shift amount is constant we can shift each lane using the SSE2
18927 // immediate shifts, else we need to zero-extend each lane to the lower i64
18928 // and shift using the SSE2 variable shifts.
18929 // The separate results can then be blended together.
18930 if (VT == MVT::v4i32) {
18931 unsigned Opc = Op.getOpcode();
18932 SDValue Amt0, Amt1, Amt2, Amt3;
18933 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18934 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18935 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18936 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18937 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18939 // ISD::SHL is handled above but we include it here for completeness.
18942 llvm_unreachable("Unknown target vector shift node");
18944 Opc = X86ISD::VSHL;
18947 Opc = X86ISD::VSRL;
18950 Opc = X86ISD::VSRA;
18953 // The SSE2 shifts use the lower i64 as the same shift amount for
18954 // all lanes and the upper i64 is ignored. These shuffle masks
18955 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18956 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18957 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18958 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18959 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18960 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18963 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18964 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18965 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18966 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18967 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18968 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18969 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18972 if (VT == MVT::v16i8 ||
18973 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18974 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18975 unsigned ShiftOpcode = Op->getOpcode();
18977 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18978 // On SSE41 targets we make use of the fact that VSELECT lowers
18979 // to PBLENDVB which selects bytes based just on the sign bit.
18980 if (Subtarget->hasSSE41()) {
18981 V0 = DAG.getBitcast(VT, V0);
18982 V1 = DAG.getBitcast(VT, V1);
18983 Sel = DAG.getBitcast(VT, Sel);
18984 return DAG.getBitcast(SelVT,
18985 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18987 // On pre-SSE41 targets we test for the sign bit by comparing to
18988 // zero - a negative value will set all bits of the lanes to true
18989 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18990 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18991 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18992 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18995 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18996 // We can safely do this using i16 shifts as we're only interested in
18997 // the 3 lower bits of each byte.
18998 Amt = DAG.getBitcast(ExtVT, Amt);
18999 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
19000 Amt = DAG.getBitcast(VT, Amt);
19002 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
19003 // r = VSELECT(r, shift(r, 4), a);
19005 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
19006 R = SignBitSelect(VT, Amt, M, R);
19009 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19011 // r = VSELECT(r, shift(r, 2), a);
19012 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
19013 R = SignBitSelect(VT, Amt, M, R);
19016 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19018 // return VSELECT(r, shift(r, 1), a);
19019 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
19020 R = SignBitSelect(VT, Amt, M, R);
19024 if (Op->getOpcode() == ISD::SRA) {
19025 // For SRA we need to unpack each byte to the higher byte of a i16 vector
19026 // so we can correctly sign extend. We don't care what happens to the
19028 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
19029 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
19030 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
19031 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
19032 ALo = DAG.getBitcast(ExtVT, ALo);
19033 AHi = DAG.getBitcast(ExtVT, AHi);
19034 RLo = DAG.getBitcast(ExtVT, RLo);
19035 RHi = DAG.getBitcast(ExtVT, RHi);
19037 // r = VSELECT(r, shift(r, 4), a);
19038 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19039 DAG.getConstant(4, dl, ExtVT));
19040 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19041 DAG.getConstant(4, dl, ExtVT));
19042 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19043 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19046 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19047 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19049 // r = VSELECT(r, shift(r, 2), a);
19050 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19051 DAG.getConstant(2, dl, ExtVT));
19052 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19053 DAG.getConstant(2, dl, ExtVT));
19054 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19055 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19058 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19059 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19061 // r = VSELECT(r, shift(r, 1), a);
19062 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19063 DAG.getConstant(1, dl, ExtVT));
19064 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19065 DAG.getConstant(1, dl, ExtVT));
19066 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19067 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19069 // Logical shift the result back to the lower byte, leaving a zero upper
19071 // meaning that we can safely pack with PACKUSWB.
19073 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
19075 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
19076 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
19080 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
19081 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
19082 // solution better.
19083 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
19084 MVT ExtVT = MVT::v8i32;
19086 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
19087 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
19088 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
19089 return DAG.getNode(ISD::TRUNCATE, dl, VT,
19090 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
19093 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
19094 MVT ExtVT = MVT::v8i32;
19095 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
19096 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
19097 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
19098 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
19099 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
19100 ALo = DAG.getBitcast(ExtVT, ALo);
19101 AHi = DAG.getBitcast(ExtVT, AHi);
19102 RLo = DAG.getBitcast(ExtVT, RLo);
19103 RHi = DAG.getBitcast(ExtVT, RHi);
19104 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
19105 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
19106 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
19107 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
19108 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
19111 if (VT == MVT::v8i16) {
19112 unsigned ShiftOpcode = Op->getOpcode();
19114 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
19115 // On SSE41 targets we make use of the fact that VSELECT lowers
19116 // to PBLENDVB which selects bytes based just on the sign bit.
19117 if (Subtarget->hasSSE41()) {
19118 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
19119 V0 = DAG.getBitcast(ExtVT, V0);
19120 V1 = DAG.getBitcast(ExtVT, V1);
19121 Sel = DAG.getBitcast(ExtVT, Sel);
19122 return DAG.getBitcast(
19123 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
19125 // On pre-SSE41 targets we splat the sign bit - a negative value will
19126 // set all bits of the lanes to true and VSELECT uses that in
19127 // its OR(AND(V0,C),AND(V1,~C)) lowering.
19129 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
19130 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
19133 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
19134 if (Subtarget->hasSSE41()) {
19135 // On SSE41 targets we need to replicate the shift mask in both
19136 // bytes for PBLENDVB.
19139 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
19140 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
19142 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
19145 // r = VSELECT(r, shift(r, 8), a);
19146 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
19147 R = SignBitSelect(Amt, M, R);
19150 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19152 // r = VSELECT(r, shift(r, 4), a);
19153 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
19154 R = SignBitSelect(Amt, M, R);
19157 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19159 // r = VSELECT(r, shift(r, 2), a);
19160 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
19161 R = SignBitSelect(Amt, M, R);
19164 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19166 // return VSELECT(r, shift(r, 1), a);
19167 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
19168 R = SignBitSelect(Amt, M, R);
19172 // Decompose 256-bit shifts into smaller 128-bit shifts.
19173 if (VT.is256BitVector()) {
19174 unsigned NumElems = VT.getVectorNumElements();
19175 MVT EltVT = VT.getVectorElementType();
19176 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19178 // Extract the two vectors
19179 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
19180 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
19182 // Recreate the shift amount vectors
19183 SDValue Amt1, Amt2;
19184 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
19185 // Constant shift amount
19186 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
19187 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
19188 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
19190 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
19191 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19193 // Variable shift amount
19194 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19195 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19198 // Issue new vector shifts for the smaller types
19199 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19200 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19202 // Concatenate the result back
19203 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19209 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
19210 SelectionDAG &DAG) {
19211 MVT VT = Op.getSimpleValueType();
19213 SDValue R = Op.getOperand(0);
19214 SDValue Amt = Op.getOperand(1);
19216 assert(VT.isVector() && "Custom lowering only for vector rotates!");
19217 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
19218 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
19220 // XOP has 128-bit vector variable + immediate rotates.
19221 // +ve/-ve Amt = rotate left/right.
19223 // Split 256-bit integers.
19224 if (VT.is256BitVector())
19225 return Lower256IntArith(Op, DAG);
19227 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
19229 // Attempt to rotate by immediate.
19230 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
19231 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
19232 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
19233 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
19234 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
19235 DAG.getConstant(RotateAmt, DL, MVT::i8));
19239 // Use general rotate by variable (per-element).
19240 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
19243 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19244 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19245 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19246 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19247 // has only one use.
19248 SDNode *N = Op.getNode();
19249 SDValue LHS = N->getOperand(0);
19250 SDValue RHS = N->getOperand(1);
19251 unsigned BaseOp = 0;
19254 switch (Op.getOpcode()) {
19255 default: llvm_unreachable("Unknown ovf instruction!");
19257 // A subtract of one will be selected as a INC. Note that INC doesn't
19258 // set CF, so we can't do this for UADDO.
19259 if (isOneConstant(RHS)) {
19260 BaseOp = X86ISD::INC;
19261 Cond = X86::COND_O;
19264 BaseOp = X86ISD::ADD;
19265 Cond = X86::COND_O;
19268 BaseOp = X86ISD::ADD;
19269 Cond = X86::COND_B;
19272 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19273 // set CF, so we can't do this for USUBO.
19274 if (isOneConstant(RHS)) {
19275 BaseOp = X86ISD::DEC;
19276 Cond = X86::COND_O;
19279 BaseOp = X86ISD::SUB;
19280 Cond = X86::COND_O;
19283 BaseOp = X86ISD::SUB;
19284 Cond = X86::COND_B;
19287 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19288 Cond = X86::COND_O;
19290 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19291 if (N->getValueType(0) == MVT::i8) {
19292 BaseOp = X86ISD::UMUL8;
19293 Cond = X86::COND_O;
19296 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19298 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19301 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19302 DAG.getConstant(X86::COND_O, DL, MVT::i32),
19303 SDValue(Sum.getNode(), 2));
19305 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19309 // Also sets EFLAGS.
19310 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19311 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19314 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19315 DAG.getConstant(Cond, DL, MVT::i32),
19316 SDValue(Sum.getNode(), 1));
19318 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19321 /// Returns true if the operand type is exactly twice the native width, and
19322 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19323 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19324 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19325 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
19326 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19329 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19330 else if (OpWidth == 128)
19331 return Subtarget->hasCmpxchg16b();
19336 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19337 return needsCmpXchgNb(SI->getValueOperand()->getType());
19340 // Note: this turns large loads into lock cmpxchg8b/16b.
19341 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19342 TargetLowering::AtomicExpansionKind
19343 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19344 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19345 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
19346 : AtomicExpansionKind::None;
19349 TargetLowering::AtomicExpansionKind
19350 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19351 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19352 Type *MemType = AI->getType();
19354 // If the operand is too big, we must see if cmpxchg8/16b is available
19355 // and default to library calls otherwise.
19356 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
19357 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19358 : AtomicExpansionKind::None;
19361 AtomicRMWInst::BinOp Op = AI->getOperation();
19364 llvm_unreachable("Unknown atomic operation");
19365 case AtomicRMWInst::Xchg:
19366 case AtomicRMWInst::Add:
19367 case AtomicRMWInst::Sub:
19368 // It's better to use xadd, xsub or xchg for these in all cases.
19369 return AtomicExpansionKind::None;
19370 case AtomicRMWInst::Or:
19371 case AtomicRMWInst::And:
19372 case AtomicRMWInst::Xor:
19373 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19374 // prefix to a normal instruction for these operations.
19375 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19376 : AtomicExpansionKind::None;
19377 case AtomicRMWInst::Nand:
19378 case AtomicRMWInst::Max:
19379 case AtomicRMWInst::Min:
19380 case AtomicRMWInst::UMax:
19381 case AtomicRMWInst::UMin:
19382 // These always require a non-trivial set of data operations on x86. We must
19383 // use a cmpxchg loop.
19384 return AtomicExpansionKind::CmpXChg;
19388 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19389 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19390 // no-sse2). There isn't any reason to disable it if the target processor
19392 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19396 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19397 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19398 Type *MemType = AI->getType();
19399 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19400 // there is no benefit in turning such RMWs into loads, and it is actually
19401 // harmful as it introduces a mfence.
19402 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19405 auto Builder = IRBuilder<>(AI);
19406 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19407 auto SynchScope = AI->getSynchScope();
19408 // We must restrict the ordering to avoid generating loads with Release or
19409 // ReleaseAcquire orderings.
19410 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19411 auto Ptr = AI->getPointerOperand();
19413 // Before the load we need a fence. Here is an example lifted from
19414 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19417 // x.store(1, relaxed);
19418 // r1 = y.fetch_add(0, release);
19420 // y.fetch_add(42, acquire);
19421 // r2 = x.load(relaxed);
19422 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19423 // lowered to just a load without a fence. A mfence flushes the store buffer,
19424 // making the optimization clearly correct.
19425 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19426 // otherwise, we might be able to be more aggressive on relaxed idempotent
19427 // rmw. In practice, they do not look useful, so we don't try to be
19428 // especially clever.
19429 if (SynchScope == SingleThread)
19430 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19431 // the IR level, so we must wrap it in an intrinsic.
19434 if (!hasMFENCE(*Subtarget))
19435 // FIXME: it might make sense to use a locked operation here but on a
19436 // different cache-line to prevent cache-line bouncing. In practice it
19437 // is probably a small win, and x86 processors without mfence are rare
19438 // enough that we do not bother.
19442 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19443 Builder.CreateCall(MFence, {});
19445 // Finally we can emit the atomic load.
19446 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19447 AI->getType()->getPrimitiveSizeInBits());
19448 Loaded->setAtomic(Order, SynchScope);
19449 AI->replaceAllUsesWith(Loaded);
19450 AI->eraseFromParent();
19454 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19455 SelectionDAG &DAG) {
19457 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19458 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19459 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19460 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19462 // The only fence that needs an instruction is a sequentially-consistent
19463 // cross-thread fence.
19464 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19465 if (hasMFENCE(*Subtarget))
19466 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19468 SDValue Chain = Op.getOperand(0);
19469 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19471 DAG.getRegister(X86::ESP, MVT::i32), // Base
19472 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19473 DAG.getRegister(0, MVT::i32), // Index
19474 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19475 DAG.getRegister(0, MVT::i32), // Segment.
19479 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19480 return SDValue(Res, 0);
19483 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19484 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19487 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19488 SelectionDAG &DAG) {
19489 MVT T = Op.getSimpleValueType();
19493 switch(T.SimpleTy) {
19494 default: llvm_unreachable("Invalid value type!");
19495 case MVT::i8: Reg = X86::AL; size = 1; break;
19496 case MVT::i16: Reg = X86::AX; size = 2; break;
19497 case MVT::i32: Reg = X86::EAX; size = 4; break;
19499 assert(Subtarget->is64Bit() && "Node not type legal!");
19500 Reg = X86::RAX; size = 8;
19503 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19504 Op.getOperand(2), SDValue());
19505 SDValue Ops[] = { cpIn.getValue(0),
19508 DAG.getTargetConstant(size, DL, MVT::i8),
19509 cpIn.getValue(1) };
19510 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19511 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19512 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19516 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19517 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19518 MVT::i32, cpOut.getValue(2));
19519 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19520 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19523 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19524 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19525 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19529 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19530 SelectionDAG &DAG) {
19531 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19532 MVT DstVT = Op.getSimpleValueType();
19534 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19535 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19536 if (DstVT != MVT::f64)
19537 // This conversion needs to be expanded.
19540 SDValue InVec = Op->getOperand(0);
19542 unsigned NumElts = SrcVT.getVectorNumElements();
19543 MVT SVT = SrcVT.getVectorElementType();
19545 // Widen the vector in input in the case of MVT::v2i32.
19546 // Example: from MVT::v2i32 to MVT::v4i32.
19547 SmallVector<SDValue, 16> Elts;
19548 for (unsigned i = 0, e = NumElts; i != e; ++i)
19549 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19550 DAG.getIntPtrConstant(i, dl)));
19552 // Explicitly mark the extra elements as Undef.
19553 Elts.append(NumElts, DAG.getUNDEF(SVT));
19555 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19556 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19557 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19558 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19559 DAG.getIntPtrConstant(0, dl));
19562 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19563 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19564 assert((DstVT == MVT::i64 ||
19565 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19566 "Unexpected custom BITCAST");
19567 // i64 <=> MMX conversions are Legal.
19568 if (SrcVT==MVT::i64 && DstVT.isVector())
19570 if (DstVT==MVT::i64 && SrcVT.isVector())
19572 // MMX <=> MMX conversions are Legal.
19573 if (SrcVT.isVector() && DstVT.isVector())
19575 // All other conversions need to be expanded.
19579 /// Compute the horizontal sum of bytes in V for the elements of VT.
19581 /// Requires V to be a byte vector and VT to be an integer vector type with
19582 /// wider elements than V's type. The width of the elements of VT determines
19583 /// how many bytes of V are summed horizontally to produce each element of the
19585 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19586 const X86Subtarget *Subtarget,
19587 SelectionDAG &DAG) {
19589 MVT ByteVecVT = V.getSimpleValueType();
19590 MVT EltVT = VT.getVectorElementType();
19591 int NumElts = VT.getVectorNumElements();
19592 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19593 "Expected value to have byte element type.");
19594 assert(EltVT != MVT::i8 &&
19595 "Horizontal byte sum only makes sense for wider elements!");
19596 unsigned VecSize = VT.getSizeInBits();
19597 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19599 // PSADBW instruction horizontally add all bytes and leave the result in i64
19600 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19601 if (EltVT == MVT::i64) {
19602 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19603 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19604 V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
19605 return DAG.getBitcast(VT, V);
19608 if (EltVT == MVT::i32) {
19609 // We unpack the low half and high half into i32s interleaved with zeros so
19610 // that we can use PSADBW to horizontally sum them. The most useful part of
19611 // this is that it lines up the results of two PSADBW instructions to be
19612 // two v2i64 vectors which concatenated are the 4 population counts. We can
19613 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19614 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19615 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19616 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19618 // Do the horizontal sums into two v2i64s.
19619 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19620 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19621 Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19622 DAG.getBitcast(ByteVecVT, Low), Zeros);
19623 High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19624 DAG.getBitcast(ByteVecVT, High), Zeros);
19626 // Merge them together.
19627 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19628 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19629 DAG.getBitcast(ShortVecVT, Low),
19630 DAG.getBitcast(ShortVecVT, High));
19632 return DAG.getBitcast(VT, V);
19635 // The only element type left is i16.
19636 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19638 // To obtain pop count for each i16 element starting from the pop count for
19639 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19640 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19641 // directly supported.
19642 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19643 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19644 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19645 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19646 DAG.getBitcast(ByteVecVT, V));
19647 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19650 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19651 const X86Subtarget *Subtarget,
19652 SelectionDAG &DAG) {
19653 MVT VT = Op.getSimpleValueType();
19654 MVT EltVT = VT.getVectorElementType();
19655 unsigned VecSize = VT.getSizeInBits();
19657 // Implement a lookup table in register by using an algorithm based on:
19658 // http://wm.ite.pl/articles/sse-popcount.html
19660 // The general idea is that every lower byte nibble in the input vector is an
19661 // index into a in-register pre-computed pop count table. We then split up the
19662 // input vector in two new ones: (1) a vector with only the shifted-right
19663 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19664 // masked out higher ones) for each byte. PSHUB is used separately with both
19665 // to index the in-register table. Next, both are added and the result is a
19666 // i8 vector where each element contains the pop count for input byte.
19668 // To obtain the pop count for elements != i8, we follow up with the same
19669 // approach and use additional tricks as described below.
19671 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19672 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19673 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19674 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19676 int NumByteElts = VecSize / 8;
19677 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19678 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19679 SmallVector<SDValue, 16> LUTVec;
19680 for (int i = 0; i < NumByteElts; ++i)
19681 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19682 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19683 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19684 DAG.getConstant(0x0F, DL, MVT::i8));
19685 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19688 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19689 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19690 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19693 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19695 // The input vector is used as the shuffle mask that index elements into the
19696 // LUT. After counting low and high nibbles, add the vector to obtain the
19697 // final pop count per i8 element.
19698 SDValue HighPopCnt =
19699 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19700 SDValue LowPopCnt =
19701 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19702 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19704 if (EltVT == MVT::i8)
19707 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19710 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19711 const X86Subtarget *Subtarget,
19712 SelectionDAG &DAG) {
19713 MVT VT = Op.getSimpleValueType();
19714 assert(VT.is128BitVector() &&
19715 "Only 128-bit vector bitmath lowering supported.");
19717 int VecSize = VT.getSizeInBits();
19718 MVT EltVT = VT.getVectorElementType();
19719 int Len = EltVT.getSizeInBits();
19721 // This is the vectorized version of the "best" algorithm from
19722 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19723 // with a minor tweak to use a series of adds + shifts instead of vector
19724 // multiplications. Implemented for all integer vector types. We only use
19725 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19726 // much faster, even faster than using native popcnt instructions.
19728 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19729 MVT VT = V.getSimpleValueType();
19730 SmallVector<SDValue, 32> Shifters(
19731 VT.getVectorNumElements(),
19732 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19733 return DAG.getNode(OpCode, DL, VT, V,
19734 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19736 auto GetMask = [&](SDValue V, APInt Mask) {
19737 MVT VT = V.getSimpleValueType();
19738 SmallVector<SDValue, 32> Masks(
19739 VT.getVectorNumElements(),
19740 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19741 return DAG.getNode(ISD::AND, DL, VT, V,
19742 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19745 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19746 // x86, so set the SRL type to have elements at least i16 wide. This is
19747 // correct because all of our SRLs are followed immediately by a mask anyways
19748 // that handles any bits that sneak into the high bits of the byte elements.
19749 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19753 // v = v - ((v >> 1) & 0x55555555...)
19755 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19756 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19757 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19759 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19760 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19761 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19762 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19763 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19765 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19766 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19767 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19768 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19770 // At this point, V contains the byte-wise population count, and we are
19771 // merely doing a horizontal sum if necessary to get the wider element
19773 if (EltVT == MVT::i8)
19776 return LowerHorizontalByteSum(
19777 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19781 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19782 SelectionDAG &DAG) {
19783 MVT VT = Op.getSimpleValueType();
19784 // FIXME: Need to add AVX-512 support here!
19785 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19786 "Unknown CTPOP type to handle");
19787 SDLoc DL(Op.getNode());
19788 SDValue Op0 = Op.getOperand(0);
19790 if (!Subtarget->hasSSSE3()) {
19791 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19792 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19793 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19796 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19797 unsigned NumElems = VT.getVectorNumElements();
19799 // Extract each 128-bit vector, compute pop count and concat the result.
19800 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19801 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19803 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19804 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19805 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19808 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19811 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19812 SelectionDAG &DAG) {
19813 assert(Op.getSimpleValueType().isVector() &&
19814 "We only do custom lowering for vector population count.");
19815 return LowerVectorCTPOP(Op, Subtarget, DAG);
19818 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19819 SDNode *Node = Op.getNode();
19821 EVT T = Node->getValueType(0);
19822 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19823 DAG.getConstant(0, dl, T), Node->getOperand(2));
19824 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19825 cast<AtomicSDNode>(Node)->getMemoryVT(),
19826 Node->getOperand(0),
19827 Node->getOperand(1), negOp,
19828 cast<AtomicSDNode>(Node)->getMemOperand(),
19829 cast<AtomicSDNode>(Node)->getOrdering(),
19830 cast<AtomicSDNode>(Node)->getSynchScope());
19833 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19834 SDNode *Node = Op.getNode();
19836 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19838 // Convert seq_cst store -> xchg
19839 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19840 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19841 // (The only way to get a 16-byte store is cmpxchg16b)
19842 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19843 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19844 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19845 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19846 cast<AtomicSDNode>(Node)->getMemoryVT(),
19847 Node->getOperand(0),
19848 Node->getOperand(1), Node->getOperand(2),
19849 cast<AtomicSDNode>(Node)->getMemOperand(),
19850 cast<AtomicSDNode>(Node)->getOrdering(),
19851 cast<AtomicSDNode>(Node)->getSynchScope());
19852 return Swap.getValue(1);
19854 // Other atomic stores have a simple pattern.
19858 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19859 MVT VT = Op.getNode()->getSimpleValueType(0);
19861 // Let legalize expand this if it isn't a legal type yet.
19862 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19865 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19868 bool ExtraOp = false;
19869 switch (Op.getOpcode()) {
19870 default: llvm_unreachable("Invalid code");
19871 case ISD::ADDC: Opc = X86ISD::ADD; break;
19872 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19873 case ISD::SUBC: Opc = X86ISD::SUB; break;
19874 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19878 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19880 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19881 Op.getOperand(1), Op.getOperand(2));
19884 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19885 SelectionDAG &DAG) {
19886 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19888 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19889 // which returns the values as { float, float } (in XMM0) or
19890 // { double, double } (which is returned in XMM0, XMM1).
19892 SDValue Arg = Op.getOperand(0);
19893 EVT ArgVT = Arg.getValueType();
19894 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19896 TargetLowering::ArgListTy Args;
19897 TargetLowering::ArgListEntry Entry;
19901 Entry.isSExt = false;
19902 Entry.isZExt = false;
19903 Args.push_back(Entry);
19905 bool isF64 = ArgVT == MVT::f64;
19906 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19907 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19908 // the results are returned via SRet in memory.
19909 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19910 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19912 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19914 Type *RetTy = isF64
19915 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19916 : (Type*)VectorType::get(ArgTy, 4);
19918 TargetLowering::CallLoweringInfo CLI(DAG);
19919 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19920 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19922 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19925 // Returned in xmm0 and xmm1.
19926 return CallResult.first;
19928 // Returned in bits 0:31 and 32:64 xmm0.
19929 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19930 CallResult.first, DAG.getIntPtrConstant(0, dl));
19931 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19932 CallResult.first, DAG.getIntPtrConstant(1, dl));
19933 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19934 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19937 /// Widen a vector input to a vector of NVT. The
19938 /// input vector must have the same element type as NVT.
19939 static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG,
19940 bool FillWithZeroes = false) {
19941 // Check if InOp already has the right width.
19942 MVT InVT = InOp.getSimpleValueType();
19946 if (InOp.isUndef())
19947 return DAG.getUNDEF(NVT);
19949 assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
19950 "input and widen element type must match");
19952 unsigned InNumElts = InVT.getVectorNumElements();
19953 unsigned WidenNumElts = NVT.getVectorNumElements();
19954 assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&
19955 "Unexpected request for vector widening");
19957 EVT EltVT = NVT.getVectorElementType();
19960 if (InOp.getOpcode() == ISD::CONCAT_VECTORS &&
19961 InOp.getNumOperands() == 2) {
19962 SDValue N1 = InOp.getOperand(1);
19963 if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) ||
19965 InOp = InOp.getOperand(0);
19966 InVT = InOp.getSimpleValueType();
19967 InNumElts = InVT.getVectorNumElements();
19970 if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) ||
19971 ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) {
19972 SmallVector<SDValue, 16> Ops;
19973 for (unsigned i = 0; i < InNumElts; ++i)
19974 Ops.push_back(InOp.getOperand(i));
19976 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) :
19977 DAG.getUNDEF(EltVT);
19978 for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i)
19979 Ops.push_back(FillVal);
19980 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
19982 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) :
19984 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal,
19985 InOp, DAG.getIntPtrConstant(0, dl));
19988 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19989 SelectionDAG &DAG) {
19990 assert(Subtarget->hasAVX512() &&
19991 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19993 // X86 scatter kills mask register, so its type should be added to
19994 // the list of return values.
19995 // If the "scatter" has 2 return values, it is already handled.
19996 if (Op.getNode()->getNumValues() == 2)
19999 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
20000 SDValue Src = N->getValue();
20001 MVT VT = Src.getSimpleValueType();
20002 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
20005 SDValue NewScatter;
20006 SDValue Index = N->getIndex();
20007 SDValue Mask = N->getMask();
20008 SDValue Chain = N->getChain();
20009 SDValue BasePtr = N->getBasePtr();
20010 MVT MemVT = N->getMemoryVT().getSimpleVT();
20011 MVT IndexVT = Index.getSimpleValueType();
20012 MVT MaskVT = Mask.getSimpleValueType();
20014 if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) {
20015 // The v2i32 value was promoted to v2i64.
20016 // Now we "redo" the type legalizer's work and widen the original
20017 // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64
20019 assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&
20020 "Unexpected memory type");
20021 int ShuffleMask[] = {0, 2, -1, -1};
20022 Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src),
20023 DAG.getUNDEF(MVT::v4i32), ShuffleMask);
20024 // Now we have 4 elements instead of 2.
20025 // Expand the index.
20026 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4);
20027 Index = ExtendToType(Index, NewIndexVT, DAG);
20029 // Expand the mask with zeroes
20030 // Mask may be <2 x i64> or <2 x i1> at this moment
20031 assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&
20032 "Unexpected mask type");
20033 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4);
20034 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20038 unsigned NumElts = VT.getVectorNumElements();
20039 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20040 !Index.getSimpleValueType().is512BitVector()) {
20041 // AVX512F supports only 512-bit vectors. Or data or index should
20042 // be 512 bit wide. If now the both index and data are 256-bit, but
20043 // the vector contains 8 elements, we just sign-extend the index
20044 if (IndexVT == MVT::v8i32)
20045 // Just extend index
20046 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20048 // The minimal number of elts in scatter is 8
20051 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20052 // Use original index here, do not modify the index twice
20053 Index = ExtendToType(N->getIndex(), NewIndexVT, DAG);
20054 if (IndexVT.getScalarType() == MVT::i32)
20055 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20058 // At this point we have promoted mask operand
20059 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20060 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20061 // Use the original mask here, do not modify the mask twice
20062 Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true);
20064 // The value that should be stored
20065 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20066 Src = ExtendToType(Src, NewVT, DAG);
20069 // If the mask is "wide" at this point - truncate it to i1 vector
20070 MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts);
20071 Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask);
20073 // The mask is killed by scatter, add it to the values
20074 SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other);
20075 SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
20076 NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops,
20077 N->getMemOperand());
20078 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
20079 return SDValue(NewScatter.getNode(), 0);
20082 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget *Subtarget,
20083 SelectionDAG &DAG) {
20085 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
20086 MVT VT = Op.getSimpleValueType();
20087 SDValue Mask = N->getMask();
20090 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20091 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20092 // This operation is legal for targets with VLX, but without
20093 // VLX the vector should be widened to 512 bit
20094 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20095 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20096 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20097 SDValue Src0 = N->getSrc0();
20098 Src0 = ExtendToType(Src0, WideDataVT, DAG);
20099 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20100 SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(),
20101 N->getBasePtr(), Mask, Src0,
20102 N->getMemoryVT(), N->getMemOperand(),
20103 N->getExtensionType());
20105 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20106 NewLoad.getValue(0),
20107 DAG.getIntPtrConstant(0, dl));
20108 SDValue RetOps[] = {Exract, NewLoad.getValue(1)};
20109 return DAG.getMergeValues(RetOps, dl);
20114 static SDValue LowerMSTORE(SDValue Op, const X86Subtarget *Subtarget,
20115 SelectionDAG &DAG) {
20116 MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode());
20117 SDValue DataToStore = N->getValue();
20118 MVT VT = DataToStore.getSimpleValueType();
20119 SDValue Mask = N->getMask();
20122 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20123 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20124 // This operation is legal for targets with VLX, but without
20125 // VLX the vector should be widened to 512 bit
20126 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20127 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20128 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20129 DataToStore = ExtendToType(DataToStore, WideDataVT, DAG);
20130 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20131 return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(),
20132 Mask, N->getMemoryVT(), N->getMemOperand(),
20133 N->isTruncatingStore());
20138 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
20139 SelectionDAG &DAG) {
20140 assert(Subtarget->hasAVX512() &&
20141 "MGATHER/MSCATTER are supported on AVX-512 arch only");
20143 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
20145 MVT VT = Op.getSimpleValueType();
20146 SDValue Index = N->getIndex();
20147 SDValue Mask = N->getMask();
20148 SDValue Src0 = N->getValue();
20149 MVT IndexVT = Index.getSimpleValueType();
20150 MVT MaskVT = Mask.getSimpleValueType();
20152 unsigned NumElts = VT.getVectorNumElements();
20153 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
20155 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20156 !Index.getSimpleValueType().is512BitVector()) {
20157 // AVX512F supports only 512-bit vectors. Or data or index should
20158 // be 512 bit wide. If now the both index and data are 256-bit, but
20159 // the vector contains 8 elements, we just sign-extend the index
20160 if (NumElts == 8) {
20161 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20162 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
20163 N->getOperand(3), Index };
20164 DAG.UpdateNodeOperands(N, Ops);
20168 // Minimal number of elements in Gather
20171 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20172 Index = ExtendToType(Index, NewIndexVT, DAG);
20173 if (IndexVT.getScalarType() == MVT::i32)
20174 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20177 MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts);
20178 // At this point we have promoted mask operand
20179 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20180 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20181 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20182 Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask);
20184 // The pass-thru value
20185 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20186 Src0 = ExtendToType(Src0, NewVT, DAG);
20188 SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
20189 SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other),
20190 N->getMemoryVT(), dl, Ops,
20191 N->getMemOperand());
20192 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20193 NewGather.getValue(0),
20194 DAG.getIntPtrConstant(0, dl));
20195 SDValue RetOps[] = {Exract, NewGather.getValue(1)};
20196 return DAG.getMergeValues(RetOps, dl);
20201 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
20202 SelectionDAG &DAG) const {
20203 // TODO: Eventually, the lowering of these nodes should be informed by or
20204 // deferred to the GC strategy for the function in which they appear. For
20205 // now, however, they must be lowered to something. Since they are logically
20206 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20207 // require special handling for these nodes), lower them as literal NOOPs for
20209 SmallVector<SDValue, 2> Ops;
20211 Ops.push_back(Op.getOperand(0));
20212 if (Op->getGluedNode())
20213 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20216 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20217 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20222 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
20223 SelectionDAG &DAG) const {
20224 // TODO: Eventually, the lowering of these nodes should be informed by or
20225 // deferred to the GC strategy for the function in which they appear. For
20226 // now, however, they must be lowered to something. Since they are logically
20227 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20228 // require special handling for these nodes), lower them as literal NOOPs for
20230 SmallVector<SDValue, 2> Ops;
20232 Ops.push_back(Op.getOperand(0));
20233 if (Op->getGluedNode())
20234 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20237 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20238 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20243 /// LowerOperation - Provide custom lowering hooks for some operations.
20245 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20246 switch (Op.getOpcode()) {
20247 default: llvm_unreachable("Should not custom lower this!");
20248 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20249 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20250 return LowerCMP_SWAP(Op, Subtarget, DAG);
20251 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20252 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20253 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20254 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20255 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
20256 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
20257 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20258 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20259 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20260 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20261 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20262 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20263 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20264 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20265 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20266 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20267 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20268 case ISD::SHL_PARTS:
20269 case ISD::SRA_PARTS:
20270 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20271 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20272 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20273 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20274 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20275 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20276 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20277 case ISD::SIGN_EXTEND_VECTOR_INREG:
20278 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
20279 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20280 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20281 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20282 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20284 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20285 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20286 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20287 case ISD::SETCC: return LowerSETCC(Op, DAG);
20288 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
20289 case ISD::SELECT: return LowerSELECT(Op, DAG);
20290 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20291 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20292 case ISD::VASTART: return LowerVASTART(Op, DAG);
20293 case ISD::VAARG: return LowerVAARG(Op, DAG);
20294 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20295 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20296 case ISD::INTRINSIC_VOID:
20297 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20298 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20299 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20300 case ISD::FRAME_TO_ARGS_OFFSET:
20301 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20302 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20303 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20304 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20305 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20306 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20307 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20308 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20309 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
20310 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
20312 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
20313 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20314 case ISD::UMUL_LOHI:
20315 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20316 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
20319 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20325 case ISD::UMULO: return LowerXALUO(Op, DAG);
20326 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20327 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20331 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20332 case ISD::ADD: return LowerADD(Op, DAG);
20333 case ISD::SUB: return LowerSUB(Op, DAG);
20337 case ISD::UMIN: return LowerMINMAX(Op, DAG);
20338 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20339 case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
20340 case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);
20341 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
20342 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
20343 case ISD::GC_TRANSITION_START:
20344 return LowerGC_TRANSITION_START(Op, DAG);
20345 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
20349 /// ReplaceNodeResults - Replace a node with an illegal result type
20350 /// with a new node built out of custom code.
20351 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20352 SmallVectorImpl<SDValue>&Results,
20353 SelectionDAG &DAG) const {
20355 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20356 switch (N->getOpcode()) {
20358 llvm_unreachable("Do not know how to custom type legalize this operation!");
20359 case X86ISD::AVG: {
20360 // Legalize types for X86ISD::AVG by expanding vectors.
20361 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20363 auto InVT = N->getValueType(0);
20364 auto InVTSize = InVT.getSizeInBits();
20365 const unsigned RegSize =
20366 (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
20367 assert((!Subtarget->hasAVX512() || RegSize < 512) &&
20368 "512-bit vector requires AVX512");
20369 assert((!Subtarget->hasAVX2() || RegSize < 256) &&
20370 "256-bit vector requires AVX2");
20372 auto ElemVT = InVT.getVectorElementType();
20373 auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
20374 RegSize / ElemVT.getSizeInBits());
20375 assert(RegSize % InVT.getSizeInBits() == 0);
20376 unsigned NumConcat = RegSize / InVT.getSizeInBits();
20378 SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
20379 Ops[0] = N->getOperand(0);
20380 SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20381 Ops[0] = N->getOperand(1);
20382 SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20384 SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
20385 Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
20386 DAG.getIntPtrConstant(0, dl)));
20389 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20390 case X86ISD::FMINC:
20392 case X86ISD::FMAXC:
20393 case X86ISD::FMAX: {
20394 EVT VT = N->getValueType(0);
20395 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
20396 SDValue UNDEF = DAG.getUNDEF(VT);
20397 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20398 N->getOperand(0), UNDEF);
20399 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20400 N->getOperand(1), UNDEF);
20401 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20404 case ISD::SIGN_EXTEND_INREG:
20409 // We don't want to expand or promote these.
20416 case ISD::UDIVREM: {
20417 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20418 Results.push_back(V);
20421 case ISD::FP_TO_SINT:
20422 case ISD::FP_TO_UINT: {
20423 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20425 std::pair<SDValue,SDValue> Vals =
20426 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20427 SDValue FIST = Vals.first, StackSlot = Vals.second;
20428 if (FIST.getNode()) {
20429 EVT VT = N->getValueType(0);
20430 // Return a load from the stack slot.
20431 if (StackSlot.getNode())
20432 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20433 MachinePointerInfo(),
20434 false, false, false, 0));
20436 Results.push_back(FIST);
20440 case ISD::UINT_TO_FP: {
20441 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20442 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20443 N->getValueType(0) != MVT::v2f32)
20445 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20447 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
20449 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20450 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20451 DAG.getBitcast(MVT::v2i64, VBias));
20452 Or = DAG.getBitcast(MVT::v2f64, Or);
20453 // TODO: Are there any fast-math-flags to propagate here?
20454 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20455 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20458 case ISD::FP_ROUND: {
20459 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20461 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20462 Results.push_back(V);
20465 case ISD::FP_EXTEND: {
20466 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
20467 // No other ValueType for FP_EXTEND should reach this point.
20468 assert(N->getValueType(0) == MVT::v2f32 &&
20469 "Do not know how to legalize this Node");
20472 case ISD::INTRINSIC_W_CHAIN: {
20473 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20475 default : llvm_unreachable("Do not know how to custom type "
20476 "legalize this intrinsic operation!");
20477 case Intrinsic::x86_rdtsc:
20478 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20480 case Intrinsic::x86_rdtscp:
20481 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20483 case Intrinsic::x86_rdpmc:
20484 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20487 case ISD::INTRINSIC_WO_CHAIN: {
20488 if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG))
20489 Results.push_back(V);
20492 case ISD::READCYCLECOUNTER: {
20493 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20496 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20497 EVT T = N->getValueType(0);
20498 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20499 bool Regs64bit = T == MVT::i128;
20500 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20501 SDValue cpInL, cpInH;
20502 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20503 DAG.getConstant(0, dl, HalfT));
20504 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20505 DAG.getConstant(1, dl, HalfT));
20506 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20507 Regs64bit ? X86::RAX : X86::EAX,
20509 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20510 Regs64bit ? X86::RDX : X86::EDX,
20511 cpInH, cpInL.getValue(1));
20512 SDValue swapInL, swapInH;
20513 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20514 DAG.getConstant(0, dl, HalfT));
20515 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20516 DAG.getConstant(1, dl, HalfT));
20517 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20518 Regs64bit ? X86::RBX : X86::EBX,
20519 swapInL, cpInH.getValue(1));
20520 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20521 Regs64bit ? X86::RCX : X86::ECX,
20522 swapInH, swapInL.getValue(1));
20523 SDValue Ops[] = { swapInH.getValue(0),
20525 swapInH.getValue(1) };
20526 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20527 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20528 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20529 X86ISD::LCMPXCHG8_DAG;
20530 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20531 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20532 Regs64bit ? X86::RAX : X86::EAX,
20533 HalfT, Result.getValue(1));
20534 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20535 Regs64bit ? X86::RDX : X86::EDX,
20536 HalfT, cpOutL.getValue(2));
20537 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20539 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20540 MVT::i32, cpOutH.getValue(2));
20542 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20543 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
20544 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20546 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20547 Results.push_back(Success);
20548 Results.push_back(EFLAGS.getValue(1));
20551 case ISD::ATOMIC_SWAP:
20552 case ISD::ATOMIC_LOAD_ADD:
20553 case ISD::ATOMIC_LOAD_SUB:
20554 case ISD::ATOMIC_LOAD_AND:
20555 case ISD::ATOMIC_LOAD_OR:
20556 case ISD::ATOMIC_LOAD_XOR:
20557 case ISD::ATOMIC_LOAD_NAND:
20558 case ISD::ATOMIC_LOAD_MIN:
20559 case ISD::ATOMIC_LOAD_MAX:
20560 case ISD::ATOMIC_LOAD_UMIN:
20561 case ISD::ATOMIC_LOAD_UMAX:
20562 case ISD::ATOMIC_LOAD: {
20563 // Delegate to generic TypeLegalization. Situations we can really handle
20564 // should have already been dealt with by AtomicExpandPass.cpp.
20567 case ISD::BITCAST: {
20568 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20569 EVT DstVT = N->getValueType(0);
20570 EVT SrcVT = N->getOperand(0)->getValueType(0);
20572 if (SrcVT != MVT::f64 ||
20573 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20576 unsigned NumElts = DstVT.getVectorNumElements();
20577 EVT SVT = DstVT.getVectorElementType();
20578 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20579 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20580 MVT::v2f64, N->getOperand(0));
20581 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
20583 if (ExperimentalVectorWideningLegalization) {
20584 // If we are legalizing vectors by widening, we already have the desired
20585 // legal vector type, just return it.
20586 Results.push_back(ToVecInt);
20590 SmallVector<SDValue, 8> Elts;
20591 for (unsigned i = 0, e = NumElts; i != e; ++i)
20592 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20593 ToVecInt, DAG.getIntPtrConstant(i, dl)));
20595 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20600 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20601 switch ((X86ISD::NodeType)Opcode) {
20602 case X86ISD::FIRST_NUMBER: break;
20603 case X86ISD::BSF: return "X86ISD::BSF";
20604 case X86ISD::BSR: return "X86ISD::BSR";
20605 case X86ISD::SHLD: return "X86ISD::SHLD";
20606 case X86ISD::SHRD: return "X86ISD::SHRD";
20607 case X86ISD::FAND: return "X86ISD::FAND";
20608 case X86ISD::FANDN: return "X86ISD::FANDN";
20609 case X86ISD::FOR: return "X86ISD::FOR";
20610 case X86ISD::FXOR: return "X86ISD::FXOR";
20611 case X86ISD::FILD: return "X86ISD::FILD";
20612 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20613 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20614 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20615 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20616 case X86ISD::FLD: return "X86ISD::FLD";
20617 case X86ISD::FST: return "X86ISD::FST";
20618 case X86ISD::CALL: return "X86ISD::CALL";
20619 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20620 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20621 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20622 case X86ISD::BT: return "X86ISD::BT";
20623 case X86ISD::CMP: return "X86ISD::CMP";
20624 case X86ISD::COMI: return "X86ISD::COMI";
20625 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20626 case X86ISD::CMPM: return "X86ISD::CMPM";
20627 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20628 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20629 case X86ISD::SETCC: return "X86ISD::SETCC";
20630 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20631 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20632 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20633 case X86ISD::CMOV: return "X86ISD::CMOV";
20634 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20635 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20636 case X86ISD::IRET: return "X86ISD::IRET";
20637 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20638 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20639 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20640 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20641 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20642 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20643 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20644 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20645 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20646 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20647 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20648 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20649 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20650 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20651 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20652 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20653 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20654 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20655 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20656 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20657 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20658 case X86ISD::HADD: return "X86ISD::HADD";
20659 case X86ISD::HSUB: return "X86ISD::HSUB";
20660 case X86ISD::FHADD: return "X86ISD::FHADD";
20661 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20662 case X86ISD::ABS: return "X86ISD::ABS";
20663 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20664 case X86ISD::FMAX: return "X86ISD::FMAX";
20665 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20666 case X86ISD::FMIN: return "X86ISD::FMIN";
20667 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20668 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20669 case X86ISD::FMINC: return "X86ISD::FMINC";
20670 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20671 case X86ISD::FRCP: return "X86ISD::FRCP";
20672 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20673 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20674 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20675 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20676 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20677 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20678 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20679 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20680 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20681 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20682 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20683 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20684 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20685 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20686 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20687 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20688 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20689 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20690 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20691 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20692 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20693 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20694 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20695 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20696 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20697 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20698 case X86ISD::CVT2MASK: return "X86ISD::CVT2MASK";
20699 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20700 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20701 case X86ISD::VSHL: return "X86ISD::VSHL";
20702 case X86ISD::VSRL: return "X86ISD::VSRL";
20703 case X86ISD::VSRA: return "X86ISD::VSRA";
20704 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20705 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20706 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20707 case X86ISD::CMPP: return "X86ISD::CMPP";
20708 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20709 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20710 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20711 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20712 case X86ISD::ADD: return "X86ISD::ADD";
20713 case X86ISD::SUB: return "X86ISD::SUB";
20714 case X86ISD::ADC: return "X86ISD::ADC";
20715 case X86ISD::SBB: return "X86ISD::SBB";
20716 case X86ISD::SMUL: return "X86ISD::SMUL";
20717 case X86ISD::UMUL: return "X86ISD::UMUL";
20718 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20719 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20720 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20721 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20722 case X86ISD::INC: return "X86ISD::INC";
20723 case X86ISD::DEC: return "X86ISD::DEC";
20724 case X86ISD::OR: return "X86ISD::OR";
20725 case X86ISD::XOR: return "X86ISD::XOR";
20726 case X86ISD::AND: return "X86ISD::AND";
20727 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20728 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20729 case X86ISD::PTEST: return "X86ISD::PTEST";
20730 case X86ISD::TESTP: return "X86ISD::TESTP";
20731 case X86ISD::TESTM: return "X86ISD::TESTM";
20732 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20733 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20734 case X86ISD::KTEST: return "X86ISD::KTEST";
20735 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20736 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20737 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20738 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20739 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20740 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20741 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20742 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20743 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20744 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20745 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20746 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20747 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20748 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20749 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20750 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20751 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20752 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20753 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20754 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20755 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20756 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20757 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20758 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20759 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20760 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20761 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20762 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20763 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20764 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20765 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20766 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20767 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20768 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20769 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20770 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20771 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20772 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20773 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20774 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20775 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20776 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20777 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20778 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20779 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20780 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20781 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20782 case X86ISD::SAHF: return "X86ISD::SAHF";
20783 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20784 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20785 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20786 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20787 case X86ISD::VPROT: return "X86ISD::VPROT";
20788 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20789 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20790 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20791 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20792 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20793 case X86ISD::FMADD: return "X86ISD::FMADD";
20794 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20795 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20796 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20797 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20798 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20799 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20800 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20801 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20802 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20803 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20804 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20805 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20806 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20807 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20808 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20809 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20810 case X86ISD::XTEST: return "X86ISD::XTEST";
20811 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20812 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20813 case X86ISD::SELECT: return "X86ISD::SELECT";
20814 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20815 case X86ISD::RCP28: return "X86ISD::RCP28";
20816 case X86ISD::EXP2: return "X86ISD::EXP2";
20817 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20818 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20819 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20820 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20821 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20822 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20823 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20824 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20825 case X86ISD::ADDS: return "X86ISD::ADDS";
20826 case X86ISD::SUBS: return "X86ISD::SUBS";
20827 case X86ISD::AVG: return "X86ISD::AVG";
20828 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20829 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20830 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20831 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20832 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20833 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20834 case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
20839 // isLegalAddressingMode - Return true if the addressing mode represented
20840 // by AM is legal for this target, for a load/store of the specified type.
20841 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20842 const AddrMode &AM, Type *Ty,
20843 unsigned AS) const {
20844 // X86 supports extremely general addressing modes.
20845 CodeModel::Model M = getTargetMachine().getCodeModel();
20846 Reloc::Model R = getTargetMachine().getRelocationModel();
20848 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20849 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20854 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20856 // If a reference to this global requires an extra load, we can't fold it.
20857 if (isGlobalStubReference(GVFlags))
20860 // If BaseGV requires a register for the PIC base, we cannot also have a
20861 // BaseReg specified.
20862 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20865 // If lower 4G is not available, then we must use rip-relative addressing.
20866 if ((M != CodeModel::Small || R != Reloc::Static) &&
20867 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20871 switch (AM.Scale) {
20877 // These scales always work.
20882 // These scales are formed with basereg+scalereg. Only accept if there is
20887 default: // Other stuff never works.
20894 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20895 unsigned Bits = Ty->getScalarSizeInBits();
20897 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20898 // particularly cheaper than those without.
20902 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20903 // variable shifts just as cheap as scalar ones.
20904 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20907 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20908 // fully general vector.
20912 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20913 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20915 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20916 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20917 return NumBits1 > NumBits2;
20920 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20921 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20924 if (!isTypeLegal(EVT::getEVT(Ty1)))
20927 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20929 // Assuming the caller doesn't have a zeroext or signext return parameter,
20930 // truncation all the way down to i1 is valid.
20934 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20935 return isInt<32>(Imm);
20938 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20939 // Can also use sub to handle negated immediates.
20940 return isInt<32>(Imm);
20943 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20944 if (!VT1.isInteger() || !VT2.isInteger())
20946 unsigned NumBits1 = VT1.getSizeInBits();
20947 unsigned NumBits2 = VT2.getSizeInBits();
20948 return NumBits1 > NumBits2;
20951 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20952 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20953 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20956 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20957 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20958 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20961 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20962 EVT VT1 = Val.getValueType();
20963 if (isZExtFree(VT1, VT2))
20966 if (Val.getOpcode() != ISD::LOAD)
20969 if (!VT1.isSimple() || !VT1.isInteger() ||
20970 !VT2.isSimple() || !VT2.isInteger())
20973 switch (VT1.getSimpleVT().SimpleTy) {
20978 // X86 has 8, 16, and 32-bit zero-extending loads.
20985 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20988 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20989 if (!Subtarget->hasAnyFMA())
20992 VT = VT.getScalarType();
20994 if (!VT.isSimple())
20997 switch (VT.getSimpleVT().SimpleTy) {
21008 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
21009 // i16 instructions are longer (0x66 prefix) and potentially slower.
21010 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
21013 /// isShuffleMaskLegal - Targets can use this to indicate that they only
21014 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
21015 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
21016 /// are assumed to be legal.
21018 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
21020 if (!VT.isSimple())
21023 // Not for i1 vectors
21024 if (VT.getSimpleVT().getScalarType() == MVT::i1)
21027 // Very little shuffling can be done for 64-bit vectors right now.
21028 if (VT.getSimpleVT().getSizeInBits() == 64)
21031 // We only care that the types being shuffled are legal. The lowering can
21032 // handle any possible shuffle mask that results.
21033 return isTypeLegal(VT.getSimpleVT());
21037 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
21039 // Just delegate to the generic legality, clear masks aren't special.
21040 return isShuffleMaskLegal(Mask, VT);
21043 //===----------------------------------------------------------------------===//
21044 // X86 Scheduler Hooks
21045 //===----------------------------------------------------------------------===//
21047 /// Utility function to emit xbegin specifying the start of an RTM region.
21048 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
21049 const TargetInstrInfo *TII) {
21050 DebugLoc DL = MI->getDebugLoc();
21052 const BasicBlock *BB = MBB->getBasicBlock();
21053 MachineFunction::iterator I = ++MBB->getIterator();
21055 // For the v = xbegin(), we generate
21066 MachineBasicBlock *thisMBB = MBB;
21067 MachineFunction *MF = MBB->getParent();
21068 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21069 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21070 MF->insert(I, mainMBB);
21071 MF->insert(I, sinkMBB);
21073 // Transfer the remainder of BB and its successor edges to sinkMBB.
21074 sinkMBB->splice(sinkMBB->begin(), MBB,
21075 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21076 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21080 // # fallthrough to mainMBB
21081 // # abortion to sinkMBB
21082 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
21083 thisMBB->addSuccessor(mainMBB);
21084 thisMBB->addSuccessor(sinkMBB);
21088 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
21089 mainMBB->addSuccessor(sinkMBB);
21092 // EAX is live into the sinkMBB
21093 sinkMBB->addLiveIn(X86::EAX);
21094 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21095 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21098 MI->eraseFromParent();
21102 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
21103 // or XMM0_V32I8 in AVX all of this code can be replaced with that
21104 // in the .td file.
21105 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
21106 const TargetInstrInfo *TII) {
21108 switch (MI->getOpcode()) {
21109 default: llvm_unreachable("illegal opcode!");
21110 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
21111 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
21112 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
21113 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
21114 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
21115 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
21116 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
21117 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
21120 DebugLoc dl = MI->getDebugLoc();
21121 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21123 unsigned NumArgs = MI->getNumOperands();
21124 for (unsigned i = 1; i < NumArgs; ++i) {
21125 MachineOperand &Op = MI->getOperand(i);
21126 if (!(Op.isReg() && Op.isImplicit()))
21127 MIB.addOperand(Op);
21129 if (MI->hasOneMemOperand())
21130 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21132 BuildMI(*BB, MI, dl,
21133 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21134 .addReg(X86::XMM0);
21136 MI->eraseFromParent();
21140 // FIXME: Custom handling because TableGen doesn't support multiple implicit
21141 // defs in an instruction pattern
21142 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
21143 const TargetInstrInfo *TII) {
21145 switch (MI->getOpcode()) {
21146 default: llvm_unreachable("illegal opcode!");
21147 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
21148 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
21149 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
21150 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
21151 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
21152 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
21153 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
21154 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
21157 DebugLoc dl = MI->getDebugLoc();
21158 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21160 unsigned NumArgs = MI->getNumOperands(); // remove the results
21161 for (unsigned i = 1; i < NumArgs; ++i) {
21162 MachineOperand &Op = MI->getOperand(i);
21163 if (!(Op.isReg() && Op.isImplicit()))
21164 MIB.addOperand(Op);
21166 if (MI->hasOneMemOperand())
21167 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21169 BuildMI(*BB, MI, dl,
21170 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21173 MI->eraseFromParent();
21177 static MachineBasicBlock *EmitWRPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21178 const X86Subtarget *Subtarget) {
21179 DebugLoc dl = MI->getDebugLoc();
21180 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21182 // insert input VAL into EAX
21183 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
21184 .addReg(MI->getOperand(0).getReg());
21185 // insert zero to ECX
21186 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21189 // insert zero to EDX
21190 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::EDX)
21193 // insert WRPKRU instruction
21194 BuildMI(*BB, MI, dl, TII->get(X86::WRPKRUr));
21196 MI->eraseFromParent(); // The pseudo is gone now.
21200 static MachineBasicBlock *EmitRDPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21201 const X86Subtarget *Subtarget) {
21202 DebugLoc dl = MI->getDebugLoc();
21203 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21205 // insert zero to ECX
21206 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21209 // insert RDPKRU instruction
21210 BuildMI(*BB, MI, dl, TII->get(X86::RDPKRUr));
21211 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21214 MI->eraseFromParent(); // The pseudo is gone now.
21218 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
21219 const X86Subtarget *Subtarget) {
21220 DebugLoc dl = MI->getDebugLoc();
21221 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21222 // Address into RAX/EAX, other two args into ECX, EDX.
21223 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
21224 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
21225 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
21226 for (int i = 0; i < X86::AddrNumOperands; ++i)
21227 MIB.addOperand(MI->getOperand(i));
21229 unsigned ValOps = X86::AddrNumOperands;
21230 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
21231 .addReg(MI->getOperand(ValOps).getReg());
21232 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
21233 .addReg(MI->getOperand(ValOps+1).getReg());
21235 // The instruction doesn't actually take any operands though.
21236 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
21238 MI->eraseFromParent(); // The pseudo is gone now.
21242 MachineBasicBlock *
21243 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
21244 MachineBasicBlock *MBB) const {
21245 // Emit va_arg instruction on X86-64.
21247 // Operands to this pseudo-instruction:
21248 // 0 ) Output : destination address (reg)
21249 // 1-5) Input : va_list address (addr, i64mem)
21250 // 6 ) ArgSize : Size (in bytes) of vararg type
21251 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
21252 // 8 ) Align : Alignment of type
21253 // 9 ) EFLAGS (implicit-def)
21255 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
21256 static_assert(X86::AddrNumOperands == 5,
21257 "VAARG_64 assumes 5 address operands");
21259 unsigned DestReg = MI->getOperand(0).getReg();
21260 MachineOperand &Base = MI->getOperand(1);
21261 MachineOperand &Scale = MI->getOperand(2);
21262 MachineOperand &Index = MI->getOperand(3);
21263 MachineOperand &Disp = MI->getOperand(4);
21264 MachineOperand &Segment = MI->getOperand(5);
21265 unsigned ArgSize = MI->getOperand(6).getImm();
21266 unsigned ArgMode = MI->getOperand(7).getImm();
21267 unsigned Align = MI->getOperand(8).getImm();
21269 // Memory Reference
21270 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
21271 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21272 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21274 // Machine Information
21275 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21276 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
21277 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
21278 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
21279 DebugLoc DL = MI->getDebugLoc();
21281 // struct va_list {
21284 // i64 overflow_area (address)
21285 // i64 reg_save_area (address)
21287 // sizeof(va_list) = 24
21288 // alignment(va_list) = 8
21290 unsigned TotalNumIntRegs = 6;
21291 unsigned TotalNumXMMRegs = 8;
21292 bool UseGPOffset = (ArgMode == 1);
21293 bool UseFPOffset = (ArgMode == 2);
21294 unsigned MaxOffset = TotalNumIntRegs * 8 +
21295 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21297 /* Align ArgSize to a multiple of 8 */
21298 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21299 bool NeedsAlign = (Align > 8);
21301 MachineBasicBlock *thisMBB = MBB;
21302 MachineBasicBlock *overflowMBB;
21303 MachineBasicBlock *offsetMBB;
21304 MachineBasicBlock *endMBB;
21306 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21307 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21308 unsigned OffsetReg = 0;
21310 if (!UseGPOffset && !UseFPOffset) {
21311 // If we only pull from the overflow region, we don't create a branch.
21312 // We don't need to alter control flow.
21313 OffsetDestReg = 0; // unused
21314 OverflowDestReg = DestReg;
21316 offsetMBB = nullptr;
21317 overflowMBB = thisMBB;
21320 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21321 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21322 // If not, pull from overflow_area. (branch to overflowMBB)
21327 // offsetMBB overflowMBB
21332 // Registers for the PHI in endMBB
21333 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21334 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21336 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21337 MachineFunction *MF = MBB->getParent();
21338 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21339 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21340 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21342 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21344 // Insert the new basic blocks
21345 MF->insert(MBBIter, offsetMBB);
21346 MF->insert(MBBIter, overflowMBB);
21347 MF->insert(MBBIter, endMBB);
21349 // Transfer the remainder of MBB and its successor edges to endMBB.
21350 endMBB->splice(endMBB->begin(), thisMBB,
21351 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21352 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21354 // Make offsetMBB and overflowMBB successors of thisMBB
21355 thisMBB->addSuccessor(offsetMBB);
21356 thisMBB->addSuccessor(overflowMBB);
21358 // endMBB is a successor of both offsetMBB and overflowMBB
21359 offsetMBB->addSuccessor(endMBB);
21360 overflowMBB->addSuccessor(endMBB);
21362 // Load the offset value into a register
21363 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21364 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21368 .addDisp(Disp, UseFPOffset ? 4 : 0)
21369 .addOperand(Segment)
21370 .setMemRefs(MMOBegin, MMOEnd);
21372 // Check if there is enough room left to pull this argument.
21373 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21375 .addImm(MaxOffset + 8 - ArgSizeA8);
21377 // Branch to "overflowMBB" if offset >= max
21378 // Fall through to "offsetMBB" otherwise
21379 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21380 .addMBB(overflowMBB);
21383 // In offsetMBB, emit code to use the reg_save_area.
21385 assert(OffsetReg != 0);
21387 // Read the reg_save_area address.
21388 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21389 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21394 .addOperand(Segment)
21395 .setMemRefs(MMOBegin, MMOEnd);
21397 // Zero-extend the offset
21398 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21399 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21402 .addImm(X86::sub_32bit);
21404 // Add the offset to the reg_save_area to get the final address.
21405 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21406 .addReg(OffsetReg64)
21407 .addReg(RegSaveReg);
21409 // Compute the offset for the next argument
21410 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21411 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21413 .addImm(UseFPOffset ? 16 : 8);
21415 // Store it back into the va_list.
21416 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21420 .addDisp(Disp, UseFPOffset ? 4 : 0)
21421 .addOperand(Segment)
21422 .addReg(NextOffsetReg)
21423 .setMemRefs(MMOBegin, MMOEnd);
21426 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21431 // Emit code to use overflow area
21434 // Load the overflow_area address into a register.
21435 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21436 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21441 .addOperand(Segment)
21442 .setMemRefs(MMOBegin, MMOEnd);
21444 // If we need to align it, do so. Otherwise, just copy the address
21445 // to OverflowDestReg.
21447 // Align the overflow address
21448 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21449 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21451 // aligned_addr = (addr + (align-1)) & ~(align-1)
21452 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21453 .addReg(OverflowAddrReg)
21456 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21458 .addImm(~(uint64_t)(Align-1));
21460 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21461 .addReg(OverflowAddrReg);
21464 // Compute the next overflow address after this argument.
21465 // (the overflow address should be kept 8-byte aligned)
21466 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21467 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21468 .addReg(OverflowDestReg)
21469 .addImm(ArgSizeA8);
21471 // Store the new overflow address.
21472 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21477 .addOperand(Segment)
21478 .addReg(NextAddrReg)
21479 .setMemRefs(MMOBegin, MMOEnd);
21481 // If we branched, emit the PHI to the front of endMBB.
21483 BuildMI(*endMBB, endMBB->begin(), DL,
21484 TII->get(X86::PHI), DestReg)
21485 .addReg(OffsetDestReg).addMBB(offsetMBB)
21486 .addReg(OverflowDestReg).addMBB(overflowMBB);
21489 // Erase the pseudo instruction
21490 MI->eraseFromParent();
21495 MachineBasicBlock *
21496 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21498 MachineBasicBlock *MBB) const {
21499 // Emit code to save XMM registers to the stack. The ABI says that the
21500 // number of registers to save is given in %al, so it's theoretically
21501 // possible to do an indirect jump trick to avoid saving all of them,
21502 // however this code takes a simpler approach and just executes all
21503 // of the stores if %al is non-zero. It's less code, and it's probably
21504 // easier on the hardware branch predictor, and stores aren't all that
21505 // expensive anyway.
21507 // Create the new basic blocks. One block contains all the XMM stores,
21508 // and one block is the final destination regardless of whether any
21509 // stores were performed.
21510 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21511 MachineFunction *F = MBB->getParent();
21512 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21513 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21514 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21515 F->insert(MBBIter, XMMSaveMBB);
21516 F->insert(MBBIter, EndMBB);
21518 // Transfer the remainder of MBB and its successor edges to EndMBB.
21519 EndMBB->splice(EndMBB->begin(), MBB,
21520 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21521 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21523 // The original block will now fall through to the XMM save block.
21524 MBB->addSuccessor(XMMSaveMBB);
21525 // The XMMSaveMBB will fall through to the end block.
21526 XMMSaveMBB->addSuccessor(EndMBB);
21528 // Now add the instructions.
21529 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21530 DebugLoc DL = MI->getDebugLoc();
21532 unsigned CountReg = MI->getOperand(0).getReg();
21533 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21534 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21536 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
21537 // If %al is 0, branch around the XMM save block.
21538 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21539 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21540 MBB->addSuccessor(EndMBB);
21543 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21544 // that was just emitted, but clearly shouldn't be "saved".
21545 assert((MI->getNumOperands() <= 3 ||
21546 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21547 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21548 && "Expected last argument to be EFLAGS");
21549 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21550 // In the XMM save block, save all the XMM argument registers.
21551 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21552 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21553 MachineMemOperand *MMO = F->getMachineMemOperand(
21554 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
21555 MachineMemOperand::MOStore,
21556 /*Size=*/16, /*Align=*/16);
21557 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21558 .addFrameIndex(RegSaveFrameIndex)
21559 .addImm(/*Scale=*/1)
21560 .addReg(/*IndexReg=*/0)
21561 .addImm(/*Disp=*/Offset)
21562 .addReg(/*Segment=*/0)
21563 .addReg(MI->getOperand(i).getReg())
21564 .addMemOperand(MMO);
21567 MI->eraseFromParent(); // The pseudo instruction is gone now.
21572 // The EFLAGS operand of SelectItr might be missing a kill marker
21573 // because there were multiple uses of EFLAGS, and ISel didn't know
21574 // which to mark. Figure out whether SelectItr should have had a
21575 // kill marker, and set it if it should. Returns the correct kill
21577 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21578 MachineBasicBlock* BB,
21579 const TargetRegisterInfo* TRI) {
21580 // Scan forward through BB for a use/def of EFLAGS.
21581 MachineBasicBlock::iterator miI(std::next(SelectItr));
21582 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21583 const MachineInstr& mi = *miI;
21584 if (mi.readsRegister(X86::EFLAGS))
21586 if (mi.definesRegister(X86::EFLAGS))
21587 break; // Should have kill-flag - update below.
21590 // If we hit the end of the block, check whether EFLAGS is live into a
21592 if (miI == BB->end()) {
21593 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21594 sEnd = BB->succ_end();
21595 sItr != sEnd; ++sItr) {
21596 MachineBasicBlock* succ = *sItr;
21597 if (succ->isLiveIn(X86::EFLAGS))
21602 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21603 // out. SelectMI should have a kill flag on EFLAGS.
21604 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21608 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
21609 // together with other CMOV pseudo-opcodes into a single basic-block with
21610 // conditional jump around it.
21611 static bool isCMOVPseudo(MachineInstr *MI) {
21612 switch (MI->getOpcode()) {
21613 case X86::CMOV_FR32:
21614 case X86::CMOV_FR64:
21615 case X86::CMOV_GR8:
21616 case X86::CMOV_GR16:
21617 case X86::CMOV_GR32:
21618 case X86::CMOV_RFP32:
21619 case X86::CMOV_RFP64:
21620 case X86::CMOV_RFP80:
21621 case X86::CMOV_V2F64:
21622 case X86::CMOV_V2I64:
21623 case X86::CMOV_V4F32:
21624 case X86::CMOV_V4F64:
21625 case X86::CMOV_V4I64:
21626 case X86::CMOV_V16F32:
21627 case X86::CMOV_V8F32:
21628 case X86::CMOV_V8F64:
21629 case X86::CMOV_V8I64:
21630 case X86::CMOV_V8I1:
21631 case X86::CMOV_V16I1:
21632 case X86::CMOV_V32I1:
21633 case X86::CMOV_V64I1:
21641 MachineBasicBlock *
21642 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21643 MachineBasicBlock *BB) const {
21644 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21645 DebugLoc DL = MI->getDebugLoc();
21647 // To "insert" a SELECT_CC instruction, we actually have to insert the
21648 // diamond control-flow pattern. The incoming instruction knows the
21649 // destination vreg to set, the condition code register to branch on, the
21650 // true/false values to select between, and a branch opcode to use.
21651 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21652 MachineFunction::iterator It = ++BB->getIterator();
21657 // cmpTY ccX, r1, r2
21659 // fallthrough --> copy0MBB
21660 MachineBasicBlock *thisMBB = BB;
21661 MachineFunction *F = BB->getParent();
21663 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21664 // as described above, by inserting a BB, and then making a PHI at the join
21665 // point to select the true and false operands of the CMOV in the PHI.
21667 // The code also handles two different cases of multiple CMOV opcodes
21671 // In this case, there are multiple CMOVs in a row, all which are based on
21672 // the same condition setting (or the exact opposite condition setting).
21673 // In this case we can lower all the CMOVs using a single inserted BB, and
21674 // then make a number of PHIs at the join point to model the CMOVs. The only
21675 // trickiness here, is that in a case like:
21677 // t2 = CMOV cond1 t1, f1
21678 // t3 = CMOV cond1 t2, f2
21680 // when rewriting this into PHIs, we have to perform some renaming on the
21681 // temps since you cannot have a PHI operand refer to a PHI result earlier
21682 // in the same block. The "simple" but wrong lowering would be:
21684 // t2 = PHI t1(BB1), f1(BB2)
21685 // t3 = PHI t2(BB1), f2(BB2)
21687 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21688 // renaming is to note that on the path through BB1, t2 is really just a
21689 // copy of t1, and do that renaming, properly generating:
21691 // t2 = PHI t1(BB1), f1(BB2)
21692 // t3 = PHI t1(BB1), f2(BB2)
21694 // Case 2, we lower cascaded CMOVs such as
21696 // (CMOV (CMOV F, T, cc1), T, cc2)
21698 // to two successives branches. For that, we look for another CMOV as the
21699 // following instruction.
21701 // Without this, we would add a PHI between the two jumps, which ends up
21702 // creating a few copies all around. For instance, for
21704 // (sitofp (zext (fcmp une)))
21706 // we would generate:
21708 // ucomiss %xmm1, %xmm0
21709 // movss <1.0f>, %xmm0
21710 // movaps %xmm0, %xmm1
21712 // xorps %xmm1, %xmm1
21715 // movaps %xmm1, %xmm0
21719 // because this custom-inserter would have generated:
21731 // A: X = ...; Y = ...
21733 // C: Z = PHI [X, A], [Y, B]
21735 // E: PHI [X, C], [Z, D]
21737 // If we lower both CMOVs in a single step, we can instead generate:
21749 // A: X = ...; Y = ...
21751 // E: PHI [X, A], [X, C], [Y, D]
21753 // Which, in our sitofp/fcmp example, gives us something like:
21755 // ucomiss %xmm1, %xmm0
21756 // movss <1.0f>, %xmm0
21759 // xorps %xmm0, %xmm0
21763 MachineInstr *CascadedCMOV = nullptr;
21764 MachineInstr *LastCMOV = MI;
21765 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21766 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21767 MachineBasicBlock::iterator NextMIIt =
21768 std::next(MachineBasicBlock::iterator(MI));
21770 // Check for case 1, where there are multiple CMOVs with the same condition
21771 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21772 // number of jumps the most.
21774 if (isCMOVPseudo(MI)) {
21775 // See if we have a string of CMOVS with the same condition.
21776 while (NextMIIt != BB->end() &&
21777 isCMOVPseudo(NextMIIt) &&
21778 (NextMIIt->getOperand(3).getImm() == CC ||
21779 NextMIIt->getOperand(3).getImm() == OppCC)) {
21780 LastCMOV = &*NextMIIt;
21785 // This checks for case 2, but only do this if we didn't already find
21786 // case 1, as indicated by LastCMOV == MI.
21787 if (LastCMOV == MI &&
21788 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21789 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21790 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21791 CascadedCMOV = &*NextMIIt;
21794 MachineBasicBlock *jcc1MBB = nullptr;
21796 // If we have a cascaded CMOV, we lower it to two successive branches to
21797 // the same block. EFLAGS is used by both, so mark it as live in the second.
21798 if (CascadedCMOV) {
21799 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21800 F->insert(It, jcc1MBB);
21801 jcc1MBB->addLiveIn(X86::EFLAGS);
21804 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21805 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21806 F->insert(It, copy0MBB);
21807 F->insert(It, sinkMBB);
21809 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21810 // live into the sink and copy blocks.
21811 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21813 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21814 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21815 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21816 copy0MBB->addLiveIn(X86::EFLAGS);
21817 sinkMBB->addLiveIn(X86::EFLAGS);
21820 // Transfer the remainder of BB and its successor edges to sinkMBB.
21821 sinkMBB->splice(sinkMBB->begin(), BB,
21822 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21823 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21825 // Add the true and fallthrough blocks as its successors.
21826 if (CascadedCMOV) {
21827 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21828 BB->addSuccessor(jcc1MBB);
21830 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21831 // jump to the sinkMBB.
21832 jcc1MBB->addSuccessor(copy0MBB);
21833 jcc1MBB->addSuccessor(sinkMBB);
21835 BB->addSuccessor(copy0MBB);
21838 // The true block target of the first (or only) branch is always sinkMBB.
21839 BB->addSuccessor(sinkMBB);
21841 // Create the conditional branch instruction.
21842 unsigned Opc = X86::GetCondBranchFromCond(CC);
21843 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21845 if (CascadedCMOV) {
21846 unsigned Opc2 = X86::GetCondBranchFromCond(
21847 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21848 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21852 // %FalseValue = ...
21853 // # fallthrough to sinkMBB
21854 copy0MBB->addSuccessor(sinkMBB);
21857 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21859 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21860 MachineBasicBlock::iterator MIItEnd =
21861 std::next(MachineBasicBlock::iterator(LastCMOV));
21862 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21863 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21864 MachineInstrBuilder MIB;
21866 // As we are creating the PHIs, we have to be careful if there is more than
21867 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21868 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21869 // That also means that PHI construction must work forward from earlier to
21870 // later, and that the code must maintain a mapping from earlier PHI's
21871 // destination registers, and the registers that went into the PHI.
21873 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21874 unsigned DestReg = MIIt->getOperand(0).getReg();
21875 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21876 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21878 // If this CMOV we are generating is the opposite condition from
21879 // the jump we generated, then we have to swap the operands for the
21880 // PHI that is going to be generated.
21881 if (MIIt->getOperand(3).getImm() == OppCC)
21882 std::swap(Op1Reg, Op2Reg);
21884 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21885 Op1Reg = RegRewriteTable[Op1Reg].first;
21887 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21888 Op2Reg = RegRewriteTable[Op2Reg].second;
21890 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21891 TII->get(X86::PHI), DestReg)
21892 .addReg(Op1Reg).addMBB(copy0MBB)
21893 .addReg(Op2Reg).addMBB(thisMBB);
21895 // Add this PHI to the rewrite table.
21896 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21899 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21900 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21901 if (CascadedCMOV) {
21902 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21903 // Copy the PHI result to the register defined by the second CMOV.
21904 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21905 DL, TII->get(TargetOpcode::COPY),
21906 CascadedCMOV->getOperand(0).getReg())
21907 .addReg(MI->getOperand(0).getReg());
21908 CascadedCMOV->eraseFromParent();
21911 // Now remove the CMOV(s).
21912 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21913 (MIIt++)->eraseFromParent();
21918 MachineBasicBlock *
21919 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21920 MachineBasicBlock *BB) const {
21921 // Combine the following atomic floating-point modification pattern:
21922 // a.store(reg OP a.load(acquire), release)
21923 // Transform them into:
21924 // OPss (%gpr), %xmm
21925 // movss %xmm, (%gpr)
21926 // Or sd equivalent for 64-bit operations.
21928 switch (MI->getOpcode()) {
21929 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21930 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21931 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21933 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21934 DebugLoc DL = MI->getDebugLoc();
21935 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21936 MachineOperand MSrc = MI->getOperand(0);
21937 unsigned VSrc = MI->getOperand(5).getReg();
21938 const MachineOperand &Disp = MI->getOperand(3);
21939 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21940 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21941 if (hasDisp && MSrc.isReg())
21942 MSrc.setIsKill(false);
21943 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21944 .addOperand(/*Base=*/MSrc)
21945 .addImm(/*Scale=*/1)
21946 .addReg(/*Index=*/0)
21947 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21949 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21950 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21952 .addOperand(/*Base=*/MSrc)
21953 .addImm(/*Scale=*/1)
21954 .addReg(/*Index=*/0)
21955 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21956 .addReg(/*Segment=*/0);
21957 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21958 MI->eraseFromParent(); // The pseudo instruction is gone now.
21962 MachineBasicBlock *
21963 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21964 MachineBasicBlock *BB) const {
21965 MachineFunction *MF = BB->getParent();
21966 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21967 DebugLoc DL = MI->getDebugLoc();
21968 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21970 assert(MF->shouldSplitStack());
21972 const bool Is64Bit = Subtarget->is64Bit();
21973 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21975 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21976 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21979 // ... [Till the alloca]
21980 // If stacklet is not large enough, jump to mallocMBB
21983 // Allocate by subtracting from RSP
21984 // Jump to continueMBB
21987 // Allocate by call to runtime
21991 // [rest of original BB]
21994 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21995 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21996 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21998 MachineRegisterInfo &MRI = MF->getRegInfo();
21999 const TargetRegisterClass *AddrRegClass =
22000 getRegClassFor(getPointerTy(MF->getDataLayout()));
22002 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
22003 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
22004 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
22005 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
22006 sizeVReg = MI->getOperand(1).getReg(),
22007 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
22009 MachineFunction::iterator MBBIter = ++BB->getIterator();
22011 MF->insert(MBBIter, bumpMBB);
22012 MF->insert(MBBIter, mallocMBB);
22013 MF->insert(MBBIter, continueMBB);
22015 continueMBB->splice(continueMBB->begin(), BB,
22016 std::next(MachineBasicBlock::iterator(MI)), BB->end());
22017 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
22019 // Add code to the main basic block to check if the stack limit has been hit,
22020 // and if so, jump to mallocMBB otherwise to bumpMBB.
22021 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
22022 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
22023 .addReg(tmpSPVReg).addReg(sizeVReg);
22024 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
22025 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
22026 .addReg(SPLimitVReg);
22027 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
22029 // bumpMBB simply decreases the stack pointer, since we know the current
22030 // stacklet has enough space.
22031 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
22032 .addReg(SPLimitVReg);
22033 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
22034 .addReg(SPLimitVReg);
22035 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22037 // Calls into a routine in libgcc to allocate more space from the heap.
22038 const uint32_t *RegMask =
22039 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
22041 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
22043 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22044 .addExternalSymbol("__morestack_allocate_stack_space")
22045 .addRegMask(RegMask)
22046 .addReg(X86::RDI, RegState::Implicit)
22047 .addReg(X86::RAX, RegState::ImplicitDefine);
22048 } else if (Is64Bit) {
22049 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
22051 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22052 .addExternalSymbol("__morestack_allocate_stack_space")
22053 .addRegMask(RegMask)
22054 .addReg(X86::EDI, RegState::Implicit)
22055 .addReg(X86::EAX, RegState::ImplicitDefine);
22057 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
22059 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
22060 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
22061 .addExternalSymbol("__morestack_allocate_stack_space")
22062 .addRegMask(RegMask)
22063 .addReg(X86::EAX, RegState::ImplicitDefine);
22067 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
22070 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
22071 .addReg(IsLP64 ? X86::RAX : X86::EAX);
22072 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22074 // Set up the CFG correctly.
22075 BB->addSuccessor(bumpMBB);
22076 BB->addSuccessor(mallocMBB);
22077 mallocMBB->addSuccessor(continueMBB);
22078 bumpMBB->addSuccessor(continueMBB);
22080 // Take care of the PHI nodes.
22081 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
22082 MI->getOperand(0).getReg())
22083 .addReg(mallocPtrVReg).addMBB(mallocMBB)
22084 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
22086 // Delete the original pseudo instruction.
22087 MI->eraseFromParent();
22090 return continueMBB;
22093 MachineBasicBlock *
22094 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
22095 MachineBasicBlock *BB) const {
22096 assert(!Subtarget->isTargetMachO());
22097 DebugLoc DL = MI->getDebugLoc();
22098 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
22099 *BB->getParent(), *BB, MI, DL, false);
22100 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
22101 MI->eraseFromParent(); // The pseudo instruction is gone now.
22105 MachineBasicBlock *
22106 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
22107 MachineBasicBlock *BB) const {
22108 MachineFunction *MF = BB->getParent();
22109 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22110 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
22111 DebugLoc DL = MI->getDebugLoc();
22113 assert(!isAsynchronousEHPersonality(
22114 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
22115 "SEH does not use catchret!");
22117 // Only 32-bit EH needs to worry about manually restoring stack pointers.
22118 if (!Subtarget->is32Bit())
22121 // C++ EH creates a new target block to hold the restore code, and wires up
22122 // the new block to the return destination with a normal JMP_4.
22123 MachineBasicBlock *RestoreMBB =
22124 MF->CreateMachineBasicBlock(BB->getBasicBlock());
22125 assert(BB->succ_size() == 1);
22126 MF->insert(std::next(BB->getIterator()), RestoreMBB);
22127 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
22128 BB->addSuccessor(RestoreMBB);
22129 MI->getOperand(0).setMBB(RestoreMBB);
22131 auto RestoreMBBI = RestoreMBB->begin();
22132 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
22133 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
22137 MachineBasicBlock *
22138 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
22139 MachineBasicBlock *BB) const {
22140 MachineFunction *MF = BB->getParent();
22141 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
22142 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
22143 // Only 32-bit SEH requires special handling for catchpad.
22144 if (IsSEH && Subtarget->is32Bit()) {
22145 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22146 DebugLoc DL = MI->getDebugLoc();
22147 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
22149 MI->eraseFromParent();
22153 MachineBasicBlock *
22154 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
22155 MachineBasicBlock *BB) const {
22156 // This is pretty easy. We're taking the value that we received from
22157 // our load from the relocation, sticking it in either RDI (x86-64)
22158 // or EAX and doing an indirect call. The return value will then
22159 // be in the normal return register.
22160 MachineFunction *F = BB->getParent();
22161 const X86InstrInfo *TII = Subtarget->getInstrInfo();
22162 DebugLoc DL = MI->getDebugLoc();
22164 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
22165 assert(MI->getOperand(3).isGlobal() && "This should be a global");
22167 // Get a register mask for the lowered call.
22168 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
22169 // proper register mask.
22170 const uint32_t *RegMask =
22171 Subtarget->is64Bit() ?
22172 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
22173 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
22174 if (Subtarget->is64Bit()) {
22175 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22176 TII->get(X86::MOV64rm), X86::RDI)
22178 .addImm(0).addReg(0)
22179 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22180 MI->getOperand(3).getTargetFlags())
22182 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
22183 addDirectMem(MIB, X86::RDI);
22184 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
22185 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
22186 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22187 TII->get(X86::MOV32rm), X86::EAX)
22189 .addImm(0).addReg(0)
22190 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22191 MI->getOperand(3).getTargetFlags())
22193 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22194 addDirectMem(MIB, X86::EAX);
22195 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22197 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22198 TII->get(X86::MOV32rm), X86::EAX)
22199 .addReg(TII->getGlobalBaseReg(F))
22200 .addImm(0).addReg(0)
22201 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22202 MI->getOperand(3).getTargetFlags())
22204 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22205 addDirectMem(MIB, X86::EAX);
22206 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22209 MI->eraseFromParent(); // The pseudo instruction is gone now.
22213 MachineBasicBlock *
22214 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
22215 MachineBasicBlock *MBB) const {
22216 DebugLoc DL = MI->getDebugLoc();
22217 MachineFunction *MF = MBB->getParent();
22218 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22219 MachineRegisterInfo &MRI = MF->getRegInfo();
22221 const BasicBlock *BB = MBB->getBasicBlock();
22222 MachineFunction::iterator I = ++MBB->getIterator();
22224 // Memory Reference
22225 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22226 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22229 unsigned MemOpndSlot = 0;
22231 unsigned CurOp = 0;
22233 DstReg = MI->getOperand(CurOp++).getReg();
22234 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
22235 assert(RC->hasType(MVT::i32) && "Invalid destination!");
22236 unsigned mainDstReg = MRI.createVirtualRegister(RC);
22237 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
22239 MemOpndSlot = CurOp;
22241 MVT PVT = getPointerTy(MF->getDataLayout());
22242 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22243 "Invalid Pointer Size!");
22245 // For v = setjmp(buf), we generate
22248 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
22249 // SjLjSetup restoreMBB
22255 // v = phi(main, restore)
22258 // if base pointer being used, load it from frame
22261 MachineBasicBlock *thisMBB = MBB;
22262 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
22263 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
22264 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
22265 MF->insert(I, mainMBB);
22266 MF->insert(I, sinkMBB);
22267 MF->push_back(restoreMBB);
22268 restoreMBB->setHasAddressTaken();
22270 MachineInstrBuilder MIB;
22272 // Transfer the remainder of BB and its successor edges to sinkMBB.
22273 sinkMBB->splice(sinkMBB->begin(), MBB,
22274 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
22275 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
22278 unsigned PtrStoreOpc = 0;
22279 unsigned LabelReg = 0;
22280 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22281 Reloc::Model RM = MF->getTarget().getRelocationModel();
22282 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
22283 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
22285 // Prepare IP either in reg or imm.
22286 if (!UseImmLabel) {
22287 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
22288 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
22289 LabelReg = MRI.createVirtualRegister(PtrRC);
22290 if (Subtarget->is64Bit()) {
22291 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
22295 .addMBB(restoreMBB)
22298 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
22299 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
22300 .addReg(XII->getGlobalBaseReg(MF))
22303 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
22307 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
22309 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
22310 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22311 if (i == X86::AddrDisp)
22312 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
22314 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
22317 MIB.addReg(LabelReg);
22319 MIB.addMBB(restoreMBB);
22320 MIB.setMemRefs(MMOBegin, MMOEnd);
22322 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
22323 .addMBB(restoreMBB);
22325 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22326 MIB.addRegMask(RegInfo->getNoPreservedMask());
22327 thisMBB->addSuccessor(mainMBB);
22328 thisMBB->addSuccessor(restoreMBB);
22332 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
22333 mainMBB->addSuccessor(sinkMBB);
22336 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
22337 TII->get(X86::PHI), DstReg)
22338 .addReg(mainDstReg).addMBB(mainMBB)
22339 .addReg(restoreDstReg).addMBB(restoreMBB);
22342 if (RegInfo->hasBasePointer(*MF)) {
22343 const bool Uses64BitFramePtr =
22344 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
22345 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
22346 X86FI->setRestoreBasePointer(MF);
22347 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
22348 unsigned BasePtr = RegInfo->getBaseRegister();
22349 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
22350 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
22351 FramePtr, true, X86FI->getRestoreBasePointerOffset())
22352 .setMIFlag(MachineInstr::FrameSetup);
22354 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
22355 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
22356 restoreMBB->addSuccessor(sinkMBB);
22358 MI->eraseFromParent();
22362 MachineBasicBlock *
22363 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
22364 MachineBasicBlock *MBB) const {
22365 DebugLoc DL = MI->getDebugLoc();
22366 MachineFunction *MF = MBB->getParent();
22367 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22368 MachineRegisterInfo &MRI = MF->getRegInfo();
22370 // Memory Reference
22371 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22372 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22374 MVT PVT = getPointerTy(MF->getDataLayout());
22375 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22376 "Invalid Pointer Size!");
22378 const TargetRegisterClass *RC =
22379 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
22380 unsigned Tmp = MRI.createVirtualRegister(RC);
22381 // Since FP is only updated here but NOT referenced, it's treated as GPR.
22382 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22383 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
22384 unsigned SP = RegInfo->getStackRegister();
22386 MachineInstrBuilder MIB;
22388 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22389 const int64_t SPOffset = 2 * PVT.getStoreSize();
22391 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
22392 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
22395 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
22396 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
22397 MIB.addOperand(MI->getOperand(i));
22398 MIB.setMemRefs(MMOBegin, MMOEnd);
22400 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
22401 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22402 if (i == X86::AddrDisp)
22403 MIB.addDisp(MI->getOperand(i), LabelOffset);
22405 MIB.addOperand(MI->getOperand(i));
22407 MIB.setMemRefs(MMOBegin, MMOEnd);
22409 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
22410 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22411 if (i == X86::AddrDisp)
22412 MIB.addDisp(MI->getOperand(i), SPOffset);
22414 MIB.addOperand(MI->getOperand(i));
22416 MIB.setMemRefs(MMOBegin, MMOEnd);
22418 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
22420 MI->eraseFromParent();
22424 // Replace 213-type (isel default) FMA3 instructions with 231-type for
22425 // accumulator loops. Writing back to the accumulator allows the coalescer
22426 // to remove extra copies in the loop.
22427 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
22428 MachineBasicBlock *
22429 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
22430 MachineBasicBlock *MBB) const {
22431 MachineOperand &AddendOp = MI->getOperand(3);
22433 // Bail out early if the addend isn't a register - we can't switch these.
22434 if (!AddendOp.isReg())
22437 MachineFunction &MF = *MBB->getParent();
22438 MachineRegisterInfo &MRI = MF.getRegInfo();
22440 // Check whether the addend is defined by a PHI:
22441 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
22442 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
22443 if (!AddendDef.isPHI())
22446 // Look for the following pattern:
22448 // %addend = phi [%entry, 0], [%loop, %result]
22450 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
22454 // %addend = phi [%entry, 0], [%loop, %result]
22456 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
22458 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
22459 assert(AddendDef.getOperand(i).isReg());
22460 MachineOperand PHISrcOp = AddendDef.getOperand(i);
22461 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
22462 if (&PHISrcInst == MI) {
22463 // Found a matching instruction.
22464 unsigned NewFMAOpc = 0;
22465 switch (MI->getOpcode()) {
22466 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
22467 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
22468 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
22469 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
22470 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
22471 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
22472 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
22473 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
22474 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
22475 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
22476 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
22477 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
22478 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
22479 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
22480 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
22481 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
22482 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
22483 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
22484 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
22485 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
22487 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
22488 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
22489 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
22490 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
22491 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
22492 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
22493 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
22494 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
22495 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
22496 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
22497 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
22498 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
22499 default: llvm_unreachable("Unrecognized FMA variant.");
22502 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22503 MachineInstrBuilder MIB =
22504 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
22505 .addOperand(MI->getOperand(0))
22506 .addOperand(MI->getOperand(3))
22507 .addOperand(MI->getOperand(2))
22508 .addOperand(MI->getOperand(1));
22509 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
22510 MI->eraseFromParent();
22517 MachineBasicBlock *
22518 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
22519 MachineBasicBlock *BB) const {
22520 switch (MI->getOpcode()) {
22521 default: llvm_unreachable("Unexpected instr type to insert");
22522 case X86::TAILJMPd64:
22523 case X86::TAILJMPr64:
22524 case X86::TAILJMPm64:
22525 case X86::TAILJMPd64_REX:
22526 case X86::TAILJMPr64_REX:
22527 case X86::TAILJMPm64_REX:
22528 llvm_unreachable("TAILJMP64 would not be touched here.");
22529 case X86::TCRETURNdi64:
22530 case X86::TCRETURNri64:
22531 case X86::TCRETURNmi64:
22533 case X86::WIN_ALLOCA:
22534 return EmitLoweredWinAlloca(MI, BB);
22535 case X86::CATCHRET:
22536 return EmitLoweredCatchRet(MI, BB);
22537 case X86::CATCHPAD:
22538 return EmitLoweredCatchPad(MI, BB);
22539 case X86::SEG_ALLOCA_32:
22540 case X86::SEG_ALLOCA_64:
22541 return EmitLoweredSegAlloca(MI, BB);
22542 case X86::TLSCall_32:
22543 case X86::TLSCall_64:
22544 return EmitLoweredTLSCall(MI, BB);
22545 case X86::CMOV_FR32:
22546 case X86::CMOV_FR64:
22547 case X86::CMOV_FR128:
22548 case X86::CMOV_GR8:
22549 case X86::CMOV_GR16:
22550 case X86::CMOV_GR32:
22551 case X86::CMOV_RFP32:
22552 case X86::CMOV_RFP64:
22553 case X86::CMOV_RFP80:
22554 case X86::CMOV_V2F64:
22555 case X86::CMOV_V2I64:
22556 case X86::CMOV_V4F32:
22557 case X86::CMOV_V4F64:
22558 case X86::CMOV_V4I64:
22559 case X86::CMOV_V16F32:
22560 case X86::CMOV_V8F32:
22561 case X86::CMOV_V8F64:
22562 case X86::CMOV_V8I64:
22563 case X86::CMOV_V8I1:
22564 case X86::CMOV_V16I1:
22565 case X86::CMOV_V32I1:
22566 case X86::CMOV_V64I1:
22567 return EmitLoweredSelect(MI, BB);
22569 case X86::RDFLAGS32:
22570 case X86::RDFLAGS64: {
22571 DebugLoc DL = MI->getDebugLoc();
22572 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22574 MI->getOpcode() == X86::RDFLAGS32 ? X86::PUSHF32 : X86::PUSHF64;
22576 MI->getOpcode() == X86::RDFLAGS32 ? X86::POP32r : X86::POP64r;
22577 BuildMI(*BB, MI, DL, TII->get(PushF));
22578 BuildMI(*BB, MI, DL, TII->get(Pop), MI->getOperand(0).getReg());
22580 MI->eraseFromParent(); // The pseudo is gone now.
22584 case X86::WRFLAGS32:
22585 case X86::WRFLAGS64: {
22586 DebugLoc DL = MI->getDebugLoc();
22587 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22589 MI->getOpcode() == X86::WRFLAGS32 ? X86::PUSH32r : X86::PUSH64r;
22591 MI->getOpcode() == X86::WRFLAGS32 ? X86::POPF32 : X86::POPF64;
22592 BuildMI(*BB, MI, DL, TII->get(Push)).addReg(MI->getOperand(0).getReg());
22593 BuildMI(*BB, MI, DL, TII->get(PopF));
22595 MI->eraseFromParent(); // The pseudo is gone now.
22599 case X86::RELEASE_FADD32mr:
22600 case X86::RELEASE_FADD64mr:
22601 return EmitLoweredAtomicFP(MI, BB);
22603 case X86::FP32_TO_INT16_IN_MEM:
22604 case X86::FP32_TO_INT32_IN_MEM:
22605 case X86::FP32_TO_INT64_IN_MEM:
22606 case X86::FP64_TO_INT16_IN_MEM:
22607 case X86::FP64_TO_INT32_IN_MEM:
22608 case X86::FP64_TO_INT64_IN_MEM:
22609 case X86::FP80_TO_INT16_IN_MEM:
22610 case X86::FP80_TO_INT32_IN_MEM:
22611 case X86::FP80_TO_INT64_IN_MEM: {
22612 MachineFunction *F = BB->getParent();
22613 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22614 DebugLoc DL = MI->getDebugLoc();
22616 // Change the floating point control register to use "round towards zero"
22617 // mode when truncating to an integer value.
22618 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
22619 addFrameReference(BuildMI(*BB, MI, DL,
22620 TII->get(X86::FNSTCW16m)), CWFrameIdx);
22622 // Load the old value of the high byte of the control word...
22624 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
22625 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
22628 // Set the high part to be round to zero...
22629 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
22632 // Reload the modified control word now...
22633 addFrameReference(BuildMI(*BB, MI, DL,
22634 TII->get(X86::FLDCW16m)), CWFrameIdx);
22636 // Restore the memory image of control word to original value
22637 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
22640 // Get the X86 opcode to use.
22642 switch (MI->getOpcode()) {
22643 default: llvm_unreachable("illegal opcode!");
22644 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
22645 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22646 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22647 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22648 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22649 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22650 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22651 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22652 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22656 MachineOperand &Op = MI->getOperand(0);
22658 AM.BaseType = X86AddressMode::RegBase;
22659 AM.Base.Reg = Op.getReg();
22661 AM.BaseType = X86AddressMode::FrameIndexBase;
22662 AM.Base.FrameIndex = Op.getIndex();
22664 Op = MI->getOperand(1);
22666 AM.Scale = Op.getImm();
22667 Op = MI->getOperand(2);
22669 AM.IndexReg = Op.getImm();
22670 Op = MI->getOperand(3);
22671 if (Op.isGlobal()) {
22672 AM.GV = Op.getGlobal();
22674 AM.Disp = Op.getImm();
22676 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22677 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22679 // Reload the original control word now.
22680 addFrameReference(BuildMI(*BB, MI, DL,
22681 TII->get(X86::FLDCW16m)), CWFrameIdx);
22683 MI->eraseFromParent(); // The pseudo instruction is gone now.
22686 // String/text processing lowering.
22687 case X86::PCMPISTRM128REG:
22688 case X86::VPCMPISTRM128REG:
22689 case X86::PCMPISTRM128MEM:
22690 case X86::VPCMPISTRM128MEM:
22691 case X86::PCMPESTRM128REG:
22692 case X86::VPCMPESTRM128REG:
22693 case X86::PCMPESTRM128MEM:
22694 case X86::VPCMPESTRM128MEM:
22695 assert(Subtarget->hasSSE42() &&
22696 "Target must have SSE4.2 or AVX features enabled");
22697 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22699 // String/text processing lowering.
22700 case X86::PCMPISTRIREG:
22701 case X86::VPCMPISTRIREG:
22702 case X86::PCMPISTRIMEM:
22703 case X86::VPCMPISTRIMEM:
22704 case X86::PCMPESTRIREG:
22705 case X86::VPCMPESTRIREG:
22706 case X86::PCMPESTRIMEM:
22707 case X86::VPCMPESTRIMEM:
22708 assert(Subtarget->hasSSE42() &&
22709 "Target must have SSE4.2 or AVX features enabled");
22710 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22712 // Thread synchronization.
22714 return EmitMonitor(MI, BB, Subtarget);
22717 return EmitWRPKRU(MI, BB, Subtarget);
22719 return EmitRDPKRU(MI, BB, Subtarget);
22722 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22724 case X86::VASTART_SAVE_XMM_REGS:
22725 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22727 case X86::VAARG_64:
22728 return EmitVAARG64WithCustomInserter(MI, BB);
22730 case X86::EH_SjLj_SetJmp32:
22731 case X86::EH_SjLj_SetJmp64:
22732 return emitEHSjLjSetJmp(MI, BB);
22734 case X86::EH_SjLj_LongJmp32:
22735 case X86::EH_SjLj_LongJmp64:
22736 return emitEHSjLjLongJmp(MI, BB);
22738 case TargetOpcode::STATEPOINT:
22739 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22740 // this point in the process. We diverge later.
22741 return emitPatchPoint(MI, BB);
22743 case TargetOpcode::STACKMAP:
22744 case TargetOpcode::PATCHPOINT:
22745 return emitPatchPoint(MI, BB);
22747 case X86::VFMADDPDr213r:
22748 case X86::VFMADDPSr213r:
22749 case X86::VFMADDSDr213r:
22750 case X86::VFMADDSSr213r:
22751 case X86::VFMSUBPDr213r:
22752 case X86::VFMSUBPSr213r:
22753 case X86::VFMSUBSDr213r:
22754 case X86::VFMSUBSSr213r:
22755 case X86::VFNMADDPDr213r:
22756 case X86::VFNMADDPSr213r:
22757 case X86::VFNMADDSDr213r:
22758 case X86::VFNMADDSSr213r:
22759 case X86::VFNMSUBPDr213r:
22760 case X86::VFNMSUBPSr213r:
22761 case X86::VFNMSUBSDr213r:
22762 case X86::VFNMSUBSSr213r:
22763 case X86::VFMADDSUBPDr213r:
22764 case X86::VFMADDSUBPSr213r:
22765 case X86::VFMSUBADDPDr213r:
22766 case X86::VFMSUBADDPSr213r:
22767 case X86::VFMADDPDr213rY:
22768 case X86::VFMADDPSr213rY:
22769 case X86::VFMSUBPDr213rY:
22770 case X86::VFMSUBPSr213rY:
22771 case X86::VFNMADDPDr213rY:
22772 case X86::VFNMADDPSr213rY:
22773 case X86::VFNMSUBPDr213rY:
22774 case X86::VFNMSUBPSr213rY:
22775 case X86::VFMADDSUBPDr213rY:
22776 case X86::VFMADDSUBPSr213rY:
22777 case X86::VFMSUBADDPDr213rY:
22778 case X86::VFMSUBADDPSr213rY:
22779 return emitFMA3Instr(MI, BB);
22783 //===----------------------------------------------------------------------===//
22784 // X86 Optimization Hooks
22785 //===----------------------------------------------------------------------===//
22787 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22790 const SelectionDAG &DAG,
22791 unsigned Depth) const {
22792 unsigned BitWidth = KnownZero.getBitWidth();
22793 unsigned Opc = Op.getOpcode();
22794 assert((Opc >= ISD::BUILTIN_OP_END ||
22795 Opc == ISD::INTRINSIC_WO_CHAIN ||
22796 Opc == ISD::INTRINSIC_W_CHAIN ||
22797 Opc == ISD::INTRINSIC_VOID) &&
22798 "Should use MaskedValueIsZero if you don't know whether Op"
22799 " is a target node!");
22801 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22815 // These nodes' second result is a boolean.
22816 if (Op.getResNo() == 0)
22819 case X86ISD::SETCC:
22820 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22822 case ISD::INTRINSIC_WO_CHAIN: {
22823 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22824 unsigned NumLoBits = 0;
22827 case Intrinsic::x86_sse_movmsk_ps:
22828 case Intrinsic::x86_avx_movmsk_ps_256:
22829 case Intrinsic::x86_sse2_movmsk_pd:
22830 case Intrinsic::x86_avx_movmsk_pd_256:
22831 case Intrinsic::x86_mmx_pmovmskb:
22832 case Intrinsic::x86_sse2_pmovmskb_128:
22833 case Intrinsic::x86_avx2_pmovmskb: {
22834 // High bits of movmskp{s|d}, pmovmskb are known zero.
22836 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22837 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22838 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22839 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22840 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22841 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22842 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22843 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22845 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22854 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22856 const SelectionDAG &,
22857 unsigned Depth) const {
22858 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22859 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22860 return Op.getValueType().getScalarSizeInBits();
22866 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22867 /// node is a GlobalAddress + offset.
22868 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22869 const GlobalValue* &GA,
22870 int64_t &Offset) const {
22871 if (N->getOpcode() == X86ISD::Wrapper) {
22872 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22873 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22874 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22878 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22881 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22882 /// FIXME: This could be expanded to support 512 bit vectors as well.
22883 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22884 TargetLowering::DAGCombinerInfo &DCI,
22885 const X86Subtarget* Subtarget) {
22887 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22888 SDValue V1 = SVOp->getOperand(0);
22889 SDValue V2 = SVOp->getOperand(1);
22890 MVT VT = SVOp->getSimpleValueType(0);
22891 unsigned NumElems = VT.getVectorNumElements();
22893 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22894 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22898 // V UNDEF BUILD_VECTOR UNDEF
22900 // CONCAT_VECTOR CONCAT_VECTOR
22903 // RESULT: V + zero extended
22905 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22906 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22907 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22910 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22913 // To match the shuffle mask, the first half of the mask should
22914 // be exactly the first vector, and all the rest a splat with the
22915 // first element of the second one.
22916 for (unsigned i = 0; i != NumElems/2; ++i)
22917 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22918 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22921 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22922 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22923 if (Ld->hasNUsesOfValue(1, 0)) {
22924 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22925 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22927 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22929 Ld->getPointerInfo(),
22930 Ld->getAlignment(),
22931 false/*isVolatile*/, true/*ReadMem*/,
22932 false/*WriteMem*/);
22934 // Make sure the newly-created LOAD is in the same position as Ld in
22935 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22936 // and update uses of Ld's output chain to use the TokenFactor.
22937 if (Ld->hasAnyUseOfValue(1)) {
22938 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22939 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22940 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22941 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22942 SDValue(ResNode.getNode(), 1));
22945 return DAG.getBitcast(VT, ResNode);
22949 // Emit a zeroed vector and insert the desired subvector on its
22951 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22952 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22953 return DCI.CombineTo(N, InsV);
22959 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22962 /// This is the leaf of the recursive combinine below. When we have found some
22963 /// chain of single-use x86 shuffle instructions and accumulated the combined
22964 /// shuffle mask represented by them, this will try to pattern match that mask
22965 /// into either a single instruction if there is a special purpose instruction
22966 /// for this operation, or into a PSHUFB instruction which is a fully general
22967 /// instruction but should only be used to replace chains over a certain depth.
22968 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22969 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22970 TargetLowering::DAGCombinerInfo &DCI,
22971 const X86Subtarget *Subtarget) {
22972 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22974 // Find the operand that enters the chain. Note that multiple uses are OK
22975 // here, we're not going to remove the operand we find.
22976 SDValue Input = Op.getOperand(0);
22977 while (Input.getOpcode() == ISD::BITCAST)
22978 Input = Input.getOperand(0);
22980 MVT VT = Input.getSimpleValueType();
22981 MVT RootVT = Root.getSimpleValueType();
22984 if (Mask.size() == 1) {
22985 int Index = Mask[0];
22986 assert((Index >= 0 || Index == SM_SentinelUndef ||
22987 Index == SM_SentinelZero) &&
22988 "Invalid shuffle index found!");
22990 // We may end up with an accumulated mask of size 1 as a result of
22991 // widening of shuffle operands (see function canWidenShuffleElements).
22992 // If the only shuffle index is equal to SM_SentinelZero then propagate
22993 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22994 // mask, and therefore the entire chain of shuffles can be folded away.
22995 if (Index == SM_SentinelZero)
22996 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22998 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
23003 // Use the float domain if the operand type is a floating point type.
23004 bool FloatDomain = VT.isFloatingPoint();
23006 // For floating point shuffles, we don't have free copies in the shuffle
23007 // instructions or the ability to load as part of the instruction, so
23008 // canonicalize their shuffles to UNPCK or MOV variants.
23010 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
23011 // vectors because it can have a load folded into it that UNPCK cannot. This
23012 // doesn't preclude something switching to the shorter encoding post-RA.
23014 // FIXME: Should teach these routines about AVX vector widths.
23015 if (FloatDomain && VT.is128BitVector()) {
23016 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
23017 bool Lo = Mask.equals({0, 0});
23020 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
23021 // is no slower than UNPCKLPD but has the option to fold the input operand
23022 // into even an unaligned memory load.
23023 if (Lo && Subtarget->hasSSE3()) {
23024 Shuffle = X86ISD::MOVDDUP;
23025 ShuffleVT = MVT::v2f64;
23027 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
23028 // than the UNPCK variants.
23029 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
23030 ShuffleVT = MVT::v4f32;
23032 if (Depth == 1 && Root->getOpcode() == Shuffle)
23033 return false; // Nothing to do!
23034 Op = DAG.getBitcast(ShuffleVT, Input);
23035 DCI.AddToWorklist(Op.getNode());
23036 if (Shuffle == X86ISD::MOVDDUP)
23037 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23039 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23040 DCI.AddToWorklist(Op.getNode());
23041 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23045 if (Subtarget->hasSSE3() &&
23046 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
23047 bool Lo = Mask.equals({0, 0, 2, 2});
23048 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
23049 MVT ShuffleVT = MVT::v4f32;
23050 if (Depth == 1 && Root->getOpcode() == Shuffle)
23051 return false; // Nothing to do!
23052 Op = DAG.getBitcast(ShuffleVT, Input);
23053 DCI.AddToWorklist(Op.getNode());
23054 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23055 DCI.AddToWorklist(Op.getNode());
23056 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23060 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
23061 bool Lo = Mask.equals({0, 0, 1, 1});
23062 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23063 MVT ShuffleVT = MVT::v4f32;
23064 if (Depth == 1 && Root->getOpcode() == Shuffle)
23065 return false; // Nothing to do!
23066 Op = DAG.getBitcast(ShuffleVT, Input);
23067 DCI.AddToWorklist(Op.getNode());
23068 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23069 DCI.AddToWorklist(Op.getNode());
23070 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23076 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
23077 // variants as none of these have single-instruction variants that are
23078 // superior to the UNPCK formulation.
23079 if (!FloatDomain && VT.is128BitVector() &&
23080 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23081 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
23082 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
23084 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
23085 bool Lo = Mask[0] == 0;
23086 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23087 if (Depth == 1 && Root->getOpcode() == Shuffle)
23088 return false; // Nothing to do!
23090 switch (Mask.size()) {
23092 ShuffleVT = MVT::v8i16;
23095 ShuffleVT = MVT::v16i8;
23098 llvm_unreachable("Impossible mask size!");
23100 Op = DAG.getBitcast(ShuffleVT, Input);
23101 DCI.AddToWorklist(Op.getNode());
23102 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23103 DCI.AddToWorklist(Op.getNode());
23104 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23109 // Don't try to re-form single instruction chains under any circumstances now
23110 // that we've done encoding canonicalization for them.
23114 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
23115 // can replace them with a single PSHUFB instruction profitably. Intel's
23116 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
23117 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
23118 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
23119 SmallVector<SDValue, 16> PSHUFBMask;
23120 int NumBytes = VT.getSizeInBits() / 8;
23121 int Ratio = NumBytes / Mask.size();
23122 for (int i = 0; i < NumBytes; ++i) {
23123 if (Mask[i / Ratio] == SM_SentinelUndef) {
23124 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
23127 int M = Mask[i / Ratio] != SM_SentinelZero
23128 ? Ratio * Mask[i / Ratio] + i % Ratio
23130 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
23132 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
23133 Op = DAG.getBitcast(ByteVT, Input);
23134 DCI.AddToWorklist(Op.getNode());
23135 SDValue PSHUFBMaskOp =
23136 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
23137 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
23138 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
23139 DCI.AddToWorklist(Op.getNode());
23140 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23145 // Failed to find any combines.
23149 /// \brief Fully generic combining of x86 shuffle instructions.
23151 /// This should be the last combine run over the x86 shuffle instructions. Once
23152 /// they have been fully optimized, this will recursively consider all chains
23153 /// of single-use shuffle instructions, build a generic model of the cumulative
23154 /// shuffle operation, and check for simpler instructions which implement this
23155 /// operation. We use this primarily for two purposes:
23157 /// 1) Collapse generic shuffles to specialized single instructions when
23158 /// equivalent. In most cases, this is just an encoding size win, but
23159 /// sometimes we will collapse multiple generic shuffles into a single
23160 /// special-purpose shuffle.
23161 /// 2) Look for sequences of shuffle instructions with 3 or more total
23162 /// instructions, and replace them with the slightly more expensive SSSE3
23163 /// PSHUFB instruction if available. We do this as the last combining step
23164 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
23165 /// a suitable short sequence of other instructions. The PHUFB will either
23166 /// use a register or have to read from memory and so is slightly (but only
23167 /// slightly) more expensive than the other shuffle instructions.
23169 /// Because this is inherently a quadratic operation (for each shuffle in
23170 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
23171 /// This should never be an issue in practice as the shuffle lowering doesn't
23172 /// produce sequences of more than 8 instructions.
23174 /// FIXME: We will currently miss some cases where the redundant shuffling
23175 /// would simplify under the threshold for PSHUFB formation because of
23176 /// combine-ordering. To fix this, we should do the redundant instruction
23177 /// combining in this recursive walk.
23178 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
23179 ArrayRef<int> RootMask,
23180 int Depth, bool HasPSHUFB,
23182 TargetLowering::DAGCombinerInfo &DCI,
23183 const X86Subtarget *Subtarget) {
23184 // Bound the depth of our recursive combine because this is ultimately
23185 // quadratic in nature.
23189 // Directly rip through bitcasts to find the underlying operand.
23190 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
23191 Op = Op.getOperand(0);
23193 MVT VT = Op.getSimpleValueType();
23194 if (!VT.isVector())
23195 return false; // Bail if we hit a non-vector.
23197 assert(Root.getSimpleValueType().isVector() &&
23198 "Shuffles operate on vector types!");
23199 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
23200 "Can only combine shuffles of the same vector register size.");
23202 if (!isTargetShuffle(Op.getOpcode()))
23204 SmallVector<int, 16> OpMask;
23206 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, true, OpMask, IsUnary);
23207 // We only can combine unary shuffles which we can decode the mask for.
23208 if (!HaveMask || !IsUnary)
23211 assert(VT.getVectorNumElements() == OpMask.size() &&
23212 "Different mask size from vector size!");
23213 assert(((RootMask.size() > OpMask.size() &&
23214 RootMask.size() % OpMask.size() == 0) ||
23215 (OpMask.size() > RootMask.size() &&
23216 OpMask.size() % RootMask.size() == 0) ||
23217 OpMask.size() == RootMask.size()) &&
23218 "The smaller number of elements must divide the larger.");
23219 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
23220 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
23221 assert(((RootRatio == 1 && OpRatio == 1) ||
23222 (RootRatio == 1) != (OpRatio == 1)) &&
23223 "Must not have a ratio for both incoming and op masks!");
23225 SmallVector<int, 16> Mask;
23226 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
23228 // Merge this shuffle operation's mask into our accumulated mask. Note that
23229 // this shuffle's mask will be the first applied to the input, followed by the
23230 // root mask to get us all the way to the root value arrangement. The reason
23231 // for this order is that we are recursing up the operation chain.
23232 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
23233 int RootIdx = i / RootRatio;
23234 if (RootMask[RootIdx] < 0) {
23235 // This is a zero or undef lane, we're done.
23236 Mask.push_back(RootMask[RootIdx]);
23240 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
23241 int OpIdx = RootMaskedIdx / OpRatio;
23242 if (OpMask[OpIdx] < 0) {
23243 // The incoming lanes are zero or undef, it doesn't matter which ones we
23245 Mask.push_back(OpMask[OpIdx]);
23249 // Ok, we have non-zero lanes, map them through.
23250 Mask.push_back(OpMask[OpIdx] * OpRatio +
23251 RootMaskedIdx % OpRatio);
23254 // See if we can recurse into the operand to combine more things.
23255 switch (Op.getOpcode()) {
23256 case X86ISD::PSHUFB:
23258 case X86ISD::PSHUFD:
23259 case X86ISD::PSHUFHW:
23260 case X86ISD::PSHUFLW:
23261 if (Op.getOperand(0).hasOneUse() &&
23262 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23263 HasPSHUFB, DAG, DCI, Subtarget))
23267 case X86ISD::UNPCKL:
23268 case X86ISD::UNPCKH:
23269 assert(Op.getOperand(0) == Op.getOperand(1) &&
23270 "We only combine unary shuffles!");
23271 // We can't check for single use, we have to check that this shuffle is the
23273 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
23274 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23275 HasPSHUFB, DAG, DCI, Subtarget))
23280 // Minor canonicalization of the accumulated shuffle mask to make it easier
23281 // to match below. All this does is detect masks with squential pairs of
23282 // elements, and shrink them to the half-width mask. It does this in a loop
23283 // so it will reduce the size of the mask to the minimal width mask which
23284 // performs an equivalent shuffle.
23285 SmallVector<int, 16> WidenedMask;
23286 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
23287 Mask = std::move(WidenedMask);
23288 WidenedMask.clear();
23291 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
23295 /// \brief Get the PSHUF-style mask from PSHUF node.
23297 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
23298 /// PSHUF-style masks that can be reused with such instructions.
23299 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
23300 MVT VT = N.getSimpleValueType();
23301 SmallVector<int, 4> Mask;
23303 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, false, Mask, IsUnary);
23307 // If we have more than 128-bits, only the low 128-bits of shuffle mask
23308 // matter. Check that the upper masks are repeats and remove them.
23309 if (VT.getSizeInBits() > 128) {
23310 int LaneElts = 128 / VT.getScalarSizeInBits();
23312 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
23313 for (int j = 0; j < LaneElts; ++j)
23314 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
23315 "Mask doesn't repeat in high 128-bit lanes!");
23317 Mask.resize(LaneElts);
23320 switch (N.getOpcode()) {
23321 case X86ISD::PSHUFD:
23323 case X86ISD::PSHUFLW:
23326 case X86ISD::PSHUFHW:
23327 Mask.erase(Mask.begin(), Mask.begin() + 4);
23328 for (int &M : Mask)
23332 llvm_unreachable("No valid shuffle instruction found!");
23336 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
23338 /// We walk up the chain and look for a combinable shuffle, skipping over
23339 /// shuffles that we could hoist this shuffle's transformation past without
23340 /// altering anything.
23342 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
23344 TargetLowering::DAGCombinerInfo &DCI) {
23345 assert(N.getOpcode() == X86ISD::PSHUFD &&
23346 "Called with something other than an x86 128-bit half shuffle!");
23349 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
23350 // of the shuffles in the chain so that we can form a fresh chain to replace
23352 SmallVector<SDValue, 8> Chain;
23353 SDValue V = N.getOperand(0);
23354 for (; V.hasOneUse(); V = V.getOperand(0)) {
23355 switch (V.getOpcode()) {
23357 return SDValue(); // Nothing combined!
23360 // Skip bitcasts as we always know the type for the target specific
23364 case X86ISD::PSHUFD:
23365 // Found another dword shuffle.
23368 case X86ISD::PSHUFLW:
23369 // Check that the low words (being shuffled) are the identity in the
23370 // dword shuffle, and the high words are self-contained.
23371 if (Mask[0] != 0 || Mask[1] != 1 ||
23372 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
23375 Chain.push_back(V);
23378 case X86ISD::PSHUFHW:
23379 // Check that the high words (being shuffled) are the identity in the
23380 // dword shuffle, and the low words are self-contained.
23381 if (Mask[2] != 2 || Mask[3] != 3 ||
23382 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
23385 Chain.push_back(V);
23388 case X86ISD::UNPCKL:
23389 case X86ISD::UNPCKH:
23390 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
23391 // shuffle into a preceding word shuffle.
23392 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
23393 V.getSimpleValueType().getVectorElementType() != MVT::i16)
23396 // Search for a half-shuffle which we can combine with.
23397 unsigned CombineOp =
23398 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
23399 if (V.getOperand(0) != V.getOperand(1) ||
23400 !V->isOnlyUserOf(V.getOperand(0).getNode()))
23402 Chain.push_back(V);
23403 V = V.getOperand(0);
23405 switch (V.getOpcode()) {
23407 return SDValue(); // Nothing to combine.
23409 case X86ISD::PSHUFLW:
23410 case X86ISD::PSHUFHW:
23411 if (V.getOpcode() == CombineOp)
23414 Chain.push_back(V);
23418 V = V.getOperand(0);
23422 } while (V.hasOneUse());
23425 // Break out of the loop if we break out of the switch.
23429 if (!V.hasOneUse())
23430 // We fell out of the loop without finding a viable combining instruction.
23433 // Merge this node's mask and our incoming mask.
23434 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23435 for (int &M : Mask)
23437 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
23438 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23440 // Rebuild the chain around this new shuffle.
23441 while (!Chain.empty()) {
23442 SDValue W = Chain.pop_back_val();
23444 if (V.getValueType() != W.getOperand(0).getValueType())
23445 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
23447 switch (W.getOpcode()) {
23449 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
23451 case X86ISD::UNPCKL:
23452 case X86ISD::UNPCKH:
23453 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
23456 case X86ISD::PSHUFD:
23457 case X86ISD::PSHUFLW:
23458 case X86ISD::PSHUFHW:
23459 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
23463 if (V.getValueType() != N.getValueType())
23464 V = DAG.getBitcast(N.getValueType(), V);
23466 // Return the new chain to replace N.
23470 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
23473 /// We walk up the chain, skipping shuffles of the other half and looking
23474 /// through shuffles which switch halves trying to find a shuffle of the same
23475 /// pair of dwords.
23476 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
23478 TargetLowering::DAGCombinerInfo &DCI) {
23480 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
23481 "Called with something other than an x86 128-bit half shuffle!");
23483 unsigned CombineOpcode = N.getOpcode();
23485 // Walk up a single-use chain looking for a combinable shuffle.
23486 SDValue V = N.getOperand(0);
23487 for (; V.hasOneUse(); V = V.getOperand(0)) {
23488 switch (V.getOpcode()) {
23490 return false; // Nothing combined!
23493 // Skip bitcasts as we always know the type for the target specific
23497 case X86ISD::PSHUFLW:
23498 case X86ISD::PSHUFHW:
23499 if (V.getOpcode() == CombineOpcode)
23502 // Other-half shuffles are no-ops.
23505 // Break out of the loop if we break out of the switch.
23509 if (!V.hasOneUse())
23510 // We fell out of the loop without finding a viable combining instruction.
23513 // Combine away the bottom node as its shuffle will be accumulated into
23514 // a preceding shuffle.
23515 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23517 // Record the old value.
23520 // Merge this node's mask and our incoming mask (adjusted to account for all
23521 // the pshufd instructions encountered).
23522 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23523 for (int &M : Mask)
23525 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
23526 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23528 // Check that the shuffles didn't cancel each other out. If not, we need to
23529 // combine to the new one.
23531 // Replace the combinable shuffle with the combined one, updating all users
23532 // so that we re-evaluate the chain here.
23533 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
23538 /// \brief Try to combine x86 target specific shuffles.
23539 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
23540 TargetLowering::DAGCombinerInfo &DCI,
23541 const X86Subtarget *Subtarget) {
23543 MVT VT = N.getSimpleValueType();
23544 SmallVector<int, 4> Mask;
23546 switch (N.getOpcode()) {
23547 case X86ISD::PSHUFD:
23548 case X86ISD::PSHUFLW:
23549 case X86ISD::PSHUFHW:
23550 Mask = getPSHUFShuffleMask(N);
23551 assert(Mask.size() == 4);
23553 case X86ISD::UNPCKL: {
23554 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
23555 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
23556 // moves upper half elements into the lower half part. For example:
23558 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
23560 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
23562 // will be combined to:
23564 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
23566 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
23567 // happen due to advanced instructions.
23568 if (!VT.is128BitVector())
23571 auto Op0 = N.getOperand(0);
23572 auto Op1 = N.getOperand(1);
23573 if (Op0.getOpcode() == ISD::UNDEF &&
23574 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
23575 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
23577 unsigned NumElts = VT.getVectorNumElements();
23578 SmallVector<int, 8> ExpectedMask(NumElts, -1);
23579 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
23582 auto ShufOp = Op1.getOperand(0);
23583 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
23584 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
23588 case X86ISD::BLENDI: {
23589 SDValue V0 = N->getOperand(0);
23590 SDValue V1 = N->getOperand(1);
23591 assert(VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() &&
23592 "Unexpected input vector types");
23594 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23595 // operands and changing the mask to 1. This saves us a bunch of
23596 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23597 // x86InstrInfo knows how to commute this back after instruction selection
23598 // if it would help register allocation.
23600 // TODO: If optimizing for size or a processor that doesn't suffer from
23601 // partial register update stalls, this should be transformed into a MOVSD
23602 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23604 if (VT == MVT::v2f64)
23605 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23606 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23607 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
23608 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23617 // Nuke no-op shuffles that show up after combining.
23618 if (isNoopShuffleMask(Mask))
23619 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23621 // Look for simplifications involving one or two shuffle instructions.
23622 SDValue V = N.getOperand(0);
23623 switch (N.getOpcode()) {
23626 case X86ISD::PSHUFLW:
23627 case X86ISD::PSHUFHW:
23628 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
23630 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
23631 return SDValue(); // We combined away this shuffle, so we're done.
23633 // See if this reduces to a PSHUFD which is no more expensive and can
23634 // combine with more operations. Note that it has to at least flip the
23635 // dwords as otherwise it would have been removed as a no-op.
23636 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
23637 int DMask[] = {0, 1, 2, 3};
23638 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
23639 DMask[DOffset + 0] = DOffset + 1;
23640 DMask[DOffset + 1] = DOffset + 0;
23641 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
23642 V = DAG.getBitcast(DVT, V);
23643 DCI.AddToWorklist(V.getNode());
23644 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
23645 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
23646 DCI.AddToWorklist(V.getNode());
23647 return DAG.getBitcast(VT, V);
23650 // Look for shuffle patterns which can be implemented as a single unpack.
23651 // FIXME: This doesn't handle the location of the PSHUFD generically, and
23652 // only works when we have a PSHUFD followed by two half-shuffles.
23653 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
23654 (V.getOpcode() == X86ISD::PSHUFLW ||
23655 V.getOpcode() == X86ISD::PSHUFHW) &&
23656 V.getOpcode() != N.getOpcode() &&
23658 SDValue D = V.getOperand(0);
23659 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
23660 D = D.getOperand(0);
23661 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
23662 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23663 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
23664 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23665 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23667 for (int i = 0; i < 4; ++i) {
23668 WordMask[i + NOffset] = Mask[i] + NOffset;
23669 WordMask[i + VOffset] = VMask[i] + VOffset;
23671 // Map the word mask through the DWord mask.
23673 for (int i = 0; i < 8; ++i)
23674 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23675 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23676 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23677 // We can replace all three shuffles with an unpack.
23678 V = DAG.getBitcast(VT, D.getOperand(0));
23679 DCI.AddToWorklist(V.getNode());
23680 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23689 case X86ISD::PSHUFD:
23690 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23699 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23701 /// We combine this directly on the abstract vector shuffle nodes so it is
23702 /// easier to generically match. We also insert dummy vector shuffle nodes for
23703 /// the operands which explicitly discard the lanes which are unused by this
23704 /// operation to try to flow through the rest of the combiner the fact that
23705 /// they're unused.
23706 static SDValue combineShuffleToAddSub(SDNode *N, const X86Subtarget *Subtarget,
23707 SelectionDAG &DAG) {
23709 EVT VT = N->getValueType(0);
23710 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
23711 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
23714 // We only handle target-independent shuffles.
23715 // FIXME: It would be easy and harmless to use the target shuffle mask
23716 // extraction tool to support more.
23717 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23720 auto *SVN = cast<ShuffleVectorSDNode>(N);
23721 SmallVector<int, 8> Mask;
23722 for (int M : SVN->getMask())
23725 SDValue V1 = N->getOperand(0);
23726 SDValue V2 = N->getOperand(1);
23728 // We require the first shuffle operand to be the FSUB node, and the second to
23729 // be the FADD node.
23730 if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
23731 ShuffleVectorSDNode::commuteMask(Mask);
23733 } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
23736 // If there are other uses of these operations we can't fold them.
23737 if (!V1->hasOneUse() || !V2->hasOneUse())
23740 // Ensure that both operations have the same operands. Note that we can
23741 // commute the FADD operands.
23742 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23743 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23744 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23747 // We're looking for blends between FADD and FSUB nodes. We insist on these
23748 // nodes being lined up in a specific expected pattern.
23749 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23750 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23751 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23754 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23757 /// PerformShuffleCombine - Performs several different shuffle combines.
23758 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23759 TargetLowering::DAGCombinerInfo &DCI,
23760 const X86Subtarget *Subtarget) {
23762 SDValue N0 = N->getOperand(0);
23763 SDValue N1 = N->getOperand(1);
23764 EVT VT = N->getValueType(0);
23766 // Don't create instructions with illegal types after legalize types has run.
23767 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23768 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23771 // If we have legalized the vector types, look for blends of FADD and FSUB
23772 // nodes that we can fuse into an ADDSUB node.
23773 if (TLI.isTypeLegal(VT))
23774 if (SDValue AddSub = combineShuffleToAddSub(N, Subtarget, DAG))
23777 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23778 if (TLI.isTypeLegal(VT) && Subtarget->hasFp256() && VT.is256BitVector() &&
23779 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23780 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23782 // During Type Legalization, when promoting illegal vector types,
23783 // the backend might introduce new shuffle dag nodes and bitcasts.
23785 // This code performs the following transformation:
23786 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23787 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23789 // We do this only if both the bitcast and the BINOP dag nodes have
23790 // one use. Also, perform this transformation only if the new binary
23791 // operation is legal. This is to avoid introducing dag nodes that
23792 // potentially need to be further expanded (or custom lowered) into a
23793 // less optimal sequence of dag nodes.
23794 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23795 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23796 N0.getOpcode() == ISD::BITCAST) {
23797 SDValue BC0 = N0.getOperand(0);
23798 EVT SVT = BC0.getValueType();
23799 unsigned Opcode = BC0.getOpcode();
23800 unsigned NumElts = VT.getVectorNumElements();
23802 if (BC0.hasOneUse() && SVT.isVector() &&
23803 SVT.getVectorNumElements() * 2 == NumElts &&
23804 TLI.isOperationLegal(Opcode, VT)) {
23805 bool CanFold = false;
23817 unsigned SVTNumElts = SVT.getVectorNumElements();
23818 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23819 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23820 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23821 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23822 CanFold = SVOp->getMaskElt(i) < 0;
23825 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23826 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23827 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23828 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23833 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23834 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23835 // consecutive, non-overlapping, and in the right order.
23836 SmallVector<SDValue, 16> Elts;
23837 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23838 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23840 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23843 if (isTargetShuffle(N->getOpcode())) {
23845 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23846 if (Shuffle.getNode())
23849 // Try recursively combining arbitrary sequences of x86 shuffle
23850 // instructions into higher-order shuffles. We do this after combining
23851 // specific PSHUF instruction sequences into their minimal form so that we
23852 // can evaluate how many specialized shuffle instructions are involved in
23853 // a particular chain.
23854 SmallVector<int, 1> NonceMask; // Just a placeholder.
23855 NonceMask.push_back(0);
23856 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23857 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23859 return SDValue(); // This routine will use CombineTo to replace N.
23865 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23866 /// specific shuffle of a load can be folded into a single element load.
23867 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23868 /// shuffles have been custom lowered so we need to handle those here.
23869 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23870 TargetLowering::DAGCombinerInfo &DCI) {
23871 if (DCI.isBeforeLegalizeOps())
23874 SDValue InVec = N->getOperand(0);
23875 SDValue EltNo = N->getOperand(1);
23876 EVT EltVT = N->getValueType(0);
23878 if (!isa<ConstantSDNode>(EltNo))
23881 EVT OriginalVT = InVec.getValueType();
23883 if (InVec.getOpcode() == ISD::BITCAST) {
23884 // Don't duplicate a load with other uses.
23885 if (!InVec.hasOneUse())
23887 EVT BCVT = InVec.getOperand(0).getValueType();
23888 if (!BCVT.isVector() ||
23889 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23891 InVec = InVec.getOperand(0);
23894 EVT CurrentVT = InVec.getValueType();
23896 if (!isTargetShuffle(InVec.getOpcode()))
23899 // Don't duplicate a load with other uses.
23900 if (!InVec.hasOneUse())
23903 SmallVector<int, 16> ShuffleMask;
23905 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(), true,
23906 ShuffleMask, UnaryShuffle))
23909 // Select the input vector, guarding against out of range extract vector.
23910 unsigned NumElems = CurrentVT.getVectorNumElements();
23911 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23912 int Idx = (Elt > (int)NumElems) ? SM_SentinelUndef : ShuffleMask[Elt];
23914 if (Idx == SM_SentinelZero)
23915 return EltVT.isInteger() ? DAG.getConstant(0, SDLoc(N), EltVT)
23916 : DAG.getConstantFP(+0.0, SDLoc(N), EltVT);
23917 if (Idx == SM_SentinelUndef)
23918 return DAG.getUNDEF(EltVT);
23920 assert(0 <= Idx && Idx < (int)(2 * NumElems) && "Shuffle index out of range");
23921 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23922 : InVec.getOperand(1);
23924 // If inputs to shuffle are the same for both ops, then allow 2 uses
23925 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23926 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23928 if (LdNode.getOpcode() == ISD::BITCAST) {
23929 // Don't duplicate a load with other uses.
23930 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23933 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23934 LdNode = LdNode.getOperand(0);
23937 if (!ISD::isNormalLoad(LdNode.getNode()))
23940 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23942 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23945 // If there's a bitcast before the shuffle, check if the load type and
23946 // alignment is valid.
23947 unsigned Align = LN0->getAlignment();
23948 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23949 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23950 EltVT.getTypeForEVT(*DAG.getContext()));
23952 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23955 // All checks match so transform back to vector_shuffle so that DAG combiner
23956 // can finish the job
23959 // Create shuffle node taking into account the case that its a unary shuffle
23960 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23961 : InVec.getOperand(1);
23962 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23963 InVec.getOperand(0), Shuffle,
23965 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23966 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23970 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23971 const X86Subtarget *Subtarget) {
23972 SDValue N0 = N->getOperand(0);
23973 EVT VT = N->getValueType(0);
23975 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23976 // special and don't usually play with other vector types, it's better to
23977 // handle them early to be sure we emit efficient code by avoiding
23978 // store-load conversions.
23979 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23980 N0.getValueType() == MVT::v2i32 &&
23981 isNullConstant(N0.getOperand(1))) {
23982 SDValue N00 = N0->getOperand(0);
23983 if (N00.getValueType() == MVT::i32)
23984 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23987 // Convert a bitcasted integer logic operation that has one bitcasted
23988 // floating-point operand and one constant operand into a floating-point
23989 // logic operation. This may create a load of the constant, but that is
23990 // cheaper than materializing the constant in an integer register and
23991 // transferring it to an SSE register or transferring the SSE operand to
23992 // integer register and back.
23994 switch (N0.getOpcode()) {
23995 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23996 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23997 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23998 default: return SDValue();
24000 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
24001 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
24002 isa<ConstantSDNode>(N0.getOperand(1)) &&
24003 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
24004 N0.getOperand(0).getOperand(0).getValueType() == VT) {
24005 SDValue N000 = N0.getOperand(0).getOperand(0);
24006 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
24007 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
24013 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
24014 /// generation and convert it from being a bunch of shuffles and extracts
24015 /// into a somewhat faster sequence. For i686, the best sequence is apparently
24016 /// storing the value and loading scalars back, while for x64 we should
24017 /// use 64-bit extracts and shifts.
24018 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
24019 TargetLowering::DAGCombinerInfo &DCI) {
24020 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
24023 SDValue InputVector = N->getOperand(0);
24024 SDLoc dl(InputVector);
24025 // Detect mmx to i32 conversion through a v2i32 elt extract.
24026 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
24027 N->getValueType(0) == MVT::i32 &&
24028 InputVector.getValueType() == MVT::v2i32) {
24030 // The bitcast source is a direct mmx result.
24031 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
24032 if (MMXSrc.getValueType() == MVT::x86mmx)
24033 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24034 N->getValueType(0),
24035 InputVector.getNode()->getOperand(0));
24037 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
24038 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
24039 MMXSrc.getValueType() == MVT::i64) {
24040 SDValue MMXSrcOp = MMXSrc.getOperand(0);
24041 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
24042 MMXSrcOp.getValueType() == MVT::v1i64 &&
24043 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
24044 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24045 N->getValueType(0), MMXSrcOp.getOperand(0));
24049 EVT VT = N->getValueType(0);
24051 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
24052 InputVector.getOpcode() == ISD::BITCAST &&
24053 isa<ConstantSDNode>(InputVector.getOperand(0))) {
24054 uint64_t ExtractedElt =
24055 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
24056 uint64_t InputValue =
24057 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
24058 uint64_t Res = (InputValue >> ExtractedElt) & 1;
24059 return DAG.getConstant(Res, dl, MVT::i1);
24061 // Only operate on vectors of 4 elements, where the alternative shuffling
24062 // gets to be more expensive.
24063 if (InputVector.getValueType() != MVT::v4i32)
24066 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
24067 // single use which is a sign-extend or zero-extend, and all elements are
24069 SmallVector<SDNode *, 4> Uses;
24070 unsigned ExtractedElements = 0;
24071 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
24072 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
24073 if (UI.getUse().getResNo() != InputVector.getResNo())
24076 SDNode *Extract = *UI;
24077 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
24080 if (Extract->getValueType(0) != MVT::i32)
24082 if (!Extract->hasOneUse())
24084 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
24085 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
24087 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
24090 // Record which element was extracted.
24091 ExtractedElements |=
24092 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
24094 Uses.push_back(Extract);
24097 // If not all the elements were used, this may not be worthwhile.
24098 if (ExtractedElements != 15)
24101 // Ok, we've now decided to do the transformation.
24102 // If 64-bit shifts are legal, use the extract-shift sequence,
24103 // otherwise bounce the vector off the cache.
24104 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24107 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
24108 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
24109 auto &DL = DAG.getDataLayout();
24110 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
24111 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24112 DAG.getConstant(0, dl, VecIdxTy));
24113 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24114 DAG.getConstant(1, dl, VecIdxTy));
24116 SDValue ShAmt = DAG.getConstant(
24117 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
24118 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
24119 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24120 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
24121 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
24122 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24123 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
24125 // Store the value to a temporary stack slot.
24126 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
24127 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
24128 MachinePointerInfo(), false, false, 0);
24130 EVT ElementType = InputVector.getValueType().getVectorElementType();
24131 unsigned EltSize = ElementType.getSizeInBits() / 8;
24133 // Replace each use (extract) with a load of the appropriate element.
24134 for (unsigned i = 0; i < 4; ++i) {
24135 uint64_t Offset = EltSize * i;
24136 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
24137 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
24139 SDValue ScalarAddr =
24140 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
24142 // Load the scalar.
24143 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
24144 ScalarAddr, MachinePointerInfo(),
24145 false, false, false, 0);
24150 // Replace the extracts
24151 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
24152 UE = Uses.end(); UI != UE; ++UI) {
24153 SDNode *Extract = *UI;
24155 SDValue Idx = Extract->getOperand(1);
24156 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
24157 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
24160 // The replacement was made in place; don't return anything.
24165 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
24166 const X86Subtarget *Subtarget) {
24168 SDValue Cond = N->getOperand(0);
24169 SDValue LHS = N->getOperand(1);
24170 SDValue RHS = N->getOperand(2);
24172 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
24173 SDValue CondSrc = Cond->getOperand(0);
24174 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
24175 Cond = CondSrc->getOperand(0);
24178 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
24181 // A vselect where all conditions and data are constants can be optimized into
24182 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
24183 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
24184 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
24187 unsigned MaskValue = 0;
24188 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
24191 MVT VT = N->getSimpleValueType(0);
24192 unsigned NumElems = VT.getVectorNumElements();
24193 SmallVector<int, 8> ShuffleMask(NumElems, -1);
24194 for (unsigned i = 0; i < NumElems; ++i) {
24195 // Be sure we emit undef where we can.
24196 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
24197 ShuffleMask[i] = -1;
24199 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
24202 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24203 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
24205 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
24208 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
24210 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
24211 TargetLowering::DAGCombinerInfo &DCI,
24212 const X86Subtarget *Subtarget) {
24214 SDValue Cond = N->getOperand(0);
24215 // Get the LHS/RHS of the select.
24216 SDValue LHS = N->getOperand(1);
24217 SDValue RHS = N->getOperand(2);
24218 EVT VT = LHS.getValueType();
24219 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24221 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
24222 // instructions match the semantics of the common C idiom x<y?x:y but not
24223 // x<=y?x:y, because of how they handle negative zero (which can be
24224 // ignored in unsafe-math mode).
24225 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
24226 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
24227 VT != MVT::f80 && VT != MVT::f128 &&
24228 (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
24229 (Subtarget->hasSSE2() ||
24230 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
24231 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24233 unsigned Opcode = 0;
24234 // Check for x CC y ? x : y.
24235 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24236 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24240 // Converting this to a min would handle NaNs incorrectly, and swapping
24241 // the operands would cause it to handle comparisons between positive
24242 // and negative zero incorrectly.
24243 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24244 if (!DAG.getTarget().Options.UnsafeFPMath &&
24245 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24247 std::swap(LHS, RHS);
24249 Opcode = X86ISD::FMIN;
24252 // Converting this to a min would handle comparisons between positive
24253 // and negative zero incorrectly.
24254 if (!DAG.getTarget().Options.UnsafeFPMath &&
24255 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24257 Opcode = X86ISD::FMIN;
24260 // Converting this to a min would handle both negative zeros and NaNs
24261 // incorrectly, but we can swap the operands to fix both.
24262 std::swap(LHS, RHS);
24266 Opcode = X86ISD::FMIN;
24270 // Converting this to a max would handle comparisons between positive
24271 // and negative zero incorrectly.
24272 if (!DAG.getTarget().Options.UnsafeFPMath &&
24273 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24275 Opcode = X86ISD::FMAX;
24278 // Converting this to a max would handle NaNs incorrectly, and swapping
24279 // the operands would cause it to handle comparisons between positive
24280 // and negative zero incorrectly.
24281 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24282 if (!DAG.getTarget().Options.UnsafeFPMath &&
24283 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24285 std::swap(LHS, RHS);
24287 Opcode = X86ISD::FMAX;
24290 // Converting this to a max would handle both negative zeros and NaNs
24291 // incorrectly, but we can swap the operands to fix both.
24292 std::swap(LHS, RHS);
24296 Opcode = X86ISD::FMAX;
24299 // Check for x CC y ? y : x -- a min/max with reversed arms.
24300 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
24301 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
24305 // Converting this to a min would handle comparisons between positive
24306 // and negative zero incorrectly, and swapping the operands would
24307 // cause it to handle NaNs incorrectly.
24308 if (!DAG.getTarget().Options.UnsafeFPMath &&
24309 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
24310 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24312 std::swap(LHS, RHS);
24314 Opcode = X86ISD::FMIN;
24317 // Converting this to a min would handle NaNs incorrectly.
24318 if (!DAG.getTarget().Options.UnsafeFPMath &&
24319 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
24321 Opcode = X86ISD::FMIN;
24324 // Converting this to a min would handle both negative zeros and NaNs
24325 // incorrectly, but we can swap the operands to fix both.
24326 std::swap(LHS, RHS);
24330 Opcode = X86ISD::FMIN;
24334 // Converting this to a max would handle NaNs incorrectly.
24335 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24337 Opcode = X86ISD::FMAX;
24340 // Converting this to a max would handle comparisons between positive
24341 // and negative zero incorrectly, and swapping the operands would
24342 // cause it to handle NaNs incorrectly.
24343 if (!DAG.getTarget().Options.UnsafeFPMath &&
24344 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
24345 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24347 std::swap(LHS, RHS);
24349 Opcode = X86ISD::FMAX;
24352 // Converting this to a max would handle both negative zeros and NaNs
24353 // incorrectly, but we can swap the operands to fix both.
24354 std::swap(LHS, RHS);
24358 Opcode = X86ISD::FMAX;
24364 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
24367 EVT CondVT = Cond.getValueType();
24368 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
24369 CondVT.getVectorElementType() == MVT::i1) {
24370 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
24371 // lowering on KNL. In this case we convert it to
24372 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
24373 // The same situation for all 128 and 256-bit vectors of i8 and i16.
24374 // Since SKX these selects have a proper lowering.
24375 EVT OpVT = LHS.getValueType();
24376 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
24377 (OpVT.getVectorElementType() == MVT::i8 ||
24378 OpVT.getVectorElementType() == MVT::i16) &&
24379 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
24380 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
24381 DCI.AddToWorklist(Cond.getNode());
24382 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
24385 // If this is a select between two integer constants, try to do some
24387 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
24388 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
24389 // Don't do this for crazy integer types.
24390 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
24391 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
24392 // so that TrueC (the true value) is larger than FalseC.
24393 bool NeedsCondInvert = false;
24395 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
24396 // Efficiently invertible.
24397 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
24398 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
24399 isa<ConstantSDNode>(Cond.getOperand(1))))) {
24400 NeedsCondInvert = true;
24401 std::swap(TrueC, FalseC);
24404 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
24405 if (FalseC->getAPIntValue() == 0 &&
24406 TrueC->getAPIntValue().isPowerOf2()) {
24407 if (NeedsCondInvert) // Invert the condition if needed.
24408 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24409 DAG.getConstant(1, DL, Cond.getValueType()));
24411 // Zero extend the condition if needed.
24412 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
24414 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24415 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
24416 DAG.getConstant(ShAmt, DL, MVT::i8));
24419 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
24420 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24421 if (NeedsCondInvert) // Invert the condition if needed.
24422 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24423 DAG.getConstant(1, DL, Cond.getValueType()));
24425 // Zero extend the condition if needed.
24426 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24427 FalseC->getValueType(0), Cond);
24428 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24429 SDValue(FalseC, 0));
24432 // Optimize cases that will turn into an LEA instruction. This requires
24433 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24434 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24435 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24436 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24438 bool isFastMultiplier = false;
24440 switch ((unsigned char)Diff) {
24442 case 1: // result = add base, cond
24443 case 2: // result = lea base( , cond*2)
24444 case 3: // result = lea base(cond, cond*2)
24445 case 4: // result = lea base( , cond*4)
24446 case 5: // result = lea base(cond, cond*4)
24447 case 8: // result = lea base( , cond*8)
24448 case 9: // result = lea base(cond, cond*8)
24449 isFastMultiplier = true;
24454 if (isFastMultiplier) {
24455 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24456 if (NeedsCondInvert) // Invert the condition if needed.
24457 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24458 DAG.getConstant(1, DL, Cond.getValueType()));
24460 // Zero extend the condition if needed.
24461 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24463 // Scale the condition by the difference.
24465 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24466 DAG.getConstant(Diff, DL,
24467 Cond.getValueType()));
24469 // Add the base if non-zero.
24470 if (FalseC->getAPIntValue() != 0)
24471 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24472 SDValue(FalseC, 0));
24479 // Canonicalize max and min:
24480 // (x > y) ? x : y -> (x >= y) ? x : y
24481 // (x < y) ? x : y -> (x <= y) ? x : y
24482 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
24483 // the need for an extra compare
24484 // against zero. e.g.
24485 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
24487 // testl %edi, %edi
24489 // cmovgl %edi, %eax
24493 // cmovsl %eax, %edi
24494 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
24495 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24496 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24497 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24502 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
24503 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
24504 Cond.getOperand(0), Cond.getOperand(1), NewCC);
24505 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
24510 // Early exit check
24511 if (!TLI.isTypeLegal(VT))
24514 // Match VSELECTs into subs with unsigned saturation.
24515 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
24516 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
24517 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
24518 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
24519 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24521 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
24522 // left side invert the predicate to simplify logic below.
24524 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
24526 CC = ISD::getSetCCInverse(CC, true);
24527 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
24531 if (Other.getNode() && Other->getNumOperands() == 2 &&
24532 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
24533 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
24534 SDValue CondRHS = Cond->getOperand(1);
24536 // Look for a general sub with unsigned saturation first.
24537 // x >= y ? x-y : 0 --> subus x, y
24538 // x > y ? x-y : 0 --> subus x, y
24539 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
24540 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
24541 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
24543 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
24544 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
24545 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
24546 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
24547 // If the RHS is a constant we have to reverse the const
24548 // canonicalization.
24549 // x > C-1 ? x+-C : 0 --> subus x, C
24550 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
24551 CondRHSConst->getAPIntValue() ==
24552 (-OpRHSConst->getAPIntValue() - 1))
24553 return DAG.getNode(
24554 X86ISD::SUBUS, DL, VT, OpLHS,
24555 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
24557 // Another special case: If C was a sign bit, the sub has been
24558 // canonicalized into a xor.
24559 // FIXME: Would it be better to use computeKnownBits to determine
24560 // whether it's safe to decanonicalize the xor?
24561 // x s< 0 ? x^C : 0 --> subus x, C
24562 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
24563 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
24564 OpRHSConst->getAPIntValue().isSignBit())
24565 // Note that we have to rebuild the RHS constant here to ensure we
24566 // don't rely on particular values of undef lanes.
24567 return DAG.getNode(
24568 X86ISD::SUBUS, DL, VT, OpLHS,
24569 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
24574 // Simplify vector selection if condition value type matches vselect
24576 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
24577 assert(Cond.getValueType().isVector() &&
24578 "vector select expects a vector selector!");
24580 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
24581 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
24583 // Try invert the condition if true value is not all 1s and false value
24585 if (!TValIsAllOnes && !FValIsAllZeros &&
24586 // Check if the selector will be produced by CMPP*/PCMP*
24587 Cond.getOpcode() == ISD::SETCC &&
24588 // Check if SETCC has already been promoted
24589 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
24591 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
24592 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
24594 if (TValIsAllZeros || FValIsAllOnes) {
24595 SDValue CC = Cond.getOperand(2);
24596 ISD::CondCode NewCC =
24597 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
24598 Cond.getOperand(0).getValueType().isInteger());
24599 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
24600 std::swap(LHS, RHS);
24601 TValIsAllOnes = FValIsAllOnes;
24602 FValIsAllZeros = TValIsAllZeros;
24606 if (TValIsAllOnes || FValIsAllZeros) {
24609 if (TValIsAllOnes && FValIsAllZeros)
24611 else if (TValIsAllOnes)
24613 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
24614 else if (FValIsAllZeros)
24615 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24616 DAG.getBitcast(CondVT, LHS));
24618 return DAG.getBitcast(VT, Ret);
24622 // We should generate an X86ISD::BLENDI from a vselect if its argument
24623 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24624 // constants. This specific pattern gets generated when we split a
24625 // selector for a 512 bit vector in a machine without AVX512 (but with
24626 // 256-bit vectors), during legalization:
24628 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24630 // Iff we find this pattern and the build_vectors are built from
24631 // constants, we translate the vselect into a shuffle_vector that we
24632 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24633 if ((N->getOpcode() == ISD::VSELECT ||
24634 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24635 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
24636 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24637 if (Shuffle.getNode())
24641 // If this is a *dynamic* select (non-constant condition) and we can match
24642 // this node with one of the variable blend instructions, restructure the
24643 // condition so that the blends can use the high bit of each element and use
24644 // SimplifyDemandedBits to simplify the condition operand.
24645 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24646 !DCI.isBeforeLegalize() &&
24647 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24648 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
24650 // Don't optimize vector selects that map to mask-registers.
24654 // We can only handle the cases where VSELECT is directly legal on the
24655 // subtarget. We custom lower VSELECT nodes with constant conditions and
24656 // this makes it hard to see whether a dynamic VSELECT will correctly
24657 // lower, so we both check the operation's status and explicitly handle the
24658 // cases where a *dynamic* blend will fail even though a constant-condition
24659 // blend could be custom lowered.
24660 // FIXME: We should find a better way to handle this class of problems.
24661 // Potentially, we should combine constant-condition vselect nodes
24662 // pre-legalization into shuffles and not mark as many types as custom
24664 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
24666 // FIXME: We don't support i16-element blends currently. We could and
24667 // should support them by making *all* the bits in the condition be set
24668 // rather than just the high bit and using an i8-element blend.
24669 if (VT.getVectorElementType() == MVT::i16)
24671 // Dynamic blending was only available from SSE4.1 onward.
24672 if (VT.is128BitVector() && !Subtarget->hasSSE41())
24674 // Byte blends are only available in AVX2
24675 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24678 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24679 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24681 APInt KnownZero, KnownOne;
24682 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24683 DCI.isBeforeLegalizeOps());
24684 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24685 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24687 // If we changed the computation somewhere in the DAG, this change
24688 // will affect all users of Cond.
24689 // Make sure it is fine and update all the nodes so that we do not
24690 // use the generic VSELECT anymore. Otherwise, we may perform
24691 // wrong optimizations as we messed up with the actual expectation
24692 // for the vector boolean values.
24693 if (Cond != TLO.Old) {
24694 // Check all uses of that condition operand to check whether it will be
24695 // consumed by non-BLEND instructions, which may depend on all bits are
24697 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24699 if (I->getOpcode() != ISD::VSELECT)
24700 // TODO: Add other opcodes eventually lowered into BLEND.
24703 // Update all the users of the condition, before committing the change,
24704 // so that the VSELECT optimizations that expect the correct vector
24705 // boolean value will not be triggered.
24706 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24708 DAG.ReplaceAllUsesOfValueWith(
24710 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24711 Cond, I->getOperand(1), I->getOperand(2)));
24712 DCI.CommitTargetLoweringOpt(TLO);
24715 // At this point, only Cond is changed. Change the condition
24716 // just for N to keep the opportunity to optimize all other
24717 // users their own way.
24718 DAG.ReplaceAllUsesOfValueWith(
24720 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24721 TLO.New, N->getOperand(1), N->getOperand(2)));
24729 // Check whether a boolean test is testing a boolean value generated by
24730 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24733 // Simplify the following patterns:
24734 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24735 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24736 // to (Op EFLAGS Cond)
24738 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24739 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24740 // to (Op EFLAGS !Cond)
24742 // where Op could be BRCOND or CMOV.
24744 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24745 // Quit if not CMP and SUB with its value result used.
24746 if (Cmp.getOpcode() != X86ISD::CMP &&
24747 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24750 // Quit if not used as a boolean value.
24751 if (CC != X86::COND_E && CC != X86::COND_NE)
24754 // Check CMP operands. One of them should be 0 or 1 and the other should be
24755 // an SetCC or extended from it.
24756 SDValue Op1 = Cmp.getOperand(0);
24757 SDValue Op2 = Cmp.getOperand(1);
24760 const ConstantSDNode* C = nullptr;
24761 bool needOppositeCond = (CC == X86::COND_E);
24762 bool checkAgainstTrue = false; // Is it a comparison against 1?
24764 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24766 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24768 else // Quit if all operands are not constants.
24771 if (C->getZExtValue() == 1) {
24772 needOppositeCond = !needOppositeCond;
24773 checkAgainstTrue = true;
24774 } else if (C->getZExtValue() != 0)
24775 // Quit if the constant is neither 0 or 1.
24778 bool truncatedToBoolWithAnd = false;
24779 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24780 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24781 SetCC.getOpcode() == ISD::TRUNCATE ||
24782 SetCC.getOpcode() == ISD::AND) {
24783 if (SetCC.getOpcode() == ISD::AND) {
24785 if (isOneConstant(SetCC.getOperand(0)))
24787 if (isOneConstant(SetCC.getOperand(1)))
24791 SetCC = SetCC.getOperand(OpIdx);
24792 truncatedToBoolWithAnd = true;
24794 SetCC = SetCC.getOperand(0);
24797 switch (SetCC.getOpcode()) {
24798 case X86ISD::SETCC_CARRY:
24799 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24800 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24801 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24802 // truncated to i1 using 'and'.
24803 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24805 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24806 "Invalid use of SETCC_CARRY!");
24808 case X86ISD::SETCC:
24809 // Set the condition code or opposite one if necessary.
24810 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24811 if (needOppositeCond)
24812 CC = X86::GetOppositeBranchCondition(CC);
24813 return SetCC.getOperand(1);
24814 case X86ISD::CMOV: {
24815 // Check whether false/true value has canonical one, i.e. 0 or 1.
24816 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24817 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24818 // Quit if true value is not a constant.
24821 // Quit if false value is not a constant.
24823 SDValue Op = SetCC.getOperand(0);
24824 // Skip 'zext' or 'trunc' node.
24825 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24826 Op.getOpcode() == ISD::TRUNCATE)
24827 Op = Op.getOperand(0);
24828 // A special case for rdrand/rdseed, where 0 is set if false cond is
24830 if ((Op.getOpcode() != X86ISD::RDRAND &&
24831 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24834 // Quit if false value is not the constant 0 or 1.
24835 bool FValIsFalse = true;
24836 if (FVal && FVal->getZExtValue() != 0) {
24837 if (FVal->getZExtValue() != 1)
24839 // If FVal is 1, opposite cond is needed.
24840 needOppositeCond = !needOppositeCond;
24841 FValIsFalse = false;
24843 // Quit if TVal is not the constant opposite of FVal.
24844 if (FValIsFalse && TVal->getZExtValue() != 1)
24846 if (!FValIsFalse && TVal->getZExtValue() != 0)
24848 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24849 if (needOppositeCond)
24850 CC = X86::GetOppositeBranchCondition(CC);
24851 return SetCC.getOperand(3);
24858 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24860 /// (X86or (X86setcc) (X86setcc))
24861 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24862 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24863 X86::CondCode &CC1, SDValue &Flags,
24865 if (Cond->getOpcode() == X86ISD::CMP) {
24866 if (!isNullConstant(Cond->getOperand(1)))
24869 Cond = Cond->getOperand(0);
24874 SDValue SetCC0, SetCC1;
24875 switch (Cond->getOpcode()) {
24876 default: return false;
24883 SetCC0 = Cond->getOperand(0);
24884 SetCC1 = Cond->getOperand(1);
24888 // Make sure we have SETCC nodes, using the same flags value.
24889 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24890 SetCC1.getOpcode() != X86ISD::SETCC ||
24891 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24894 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24895 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24896 Flags = SetCC0->getOperand(1);
24900 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24901 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24902 TargetLowering::DAGCombinerInfo &DCI,
24903 const X86Subtarget *Subtarget) {
24906 // If the flag operand isn't dead, don't touch this CMOV.
24907 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24910 SDValue FalseOp = N->getOperand(0);
24911 SDValue TrueOp = N->getOperand(1);
24912 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24913 SDValue Cond = N->getOperand(3);
24915 if (CC == X86::COND_E || CC == X86::COND_NE) {
24916 switch (Cond.getOpcode()) {
24920 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24921 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24922 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24928 Flags = checkBoolTestSetCCCombine(Cond, CC);
24929 if (Flags.getNode() &&
24930 // Extra check as FCMOV only supports a subset of X86 cond.
24931 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24932 SDValue Ops[] = { FalseOp, TrueOp,
24933 DAG.getConstant(CC, DL, MVT::i8), Flags };
24934 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24937 // If this is a select between two integer constants, try to do some
24938 // optimizations. Note that the operands are ordered the opposite of SELECT
24940 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24941 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24942 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24943 // larger than FalseC (the false value).
24944 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24945 CC = X86::GetOppositeBranchCondition(CC);
24946 std::swap(TrueC, FalseC);
24947 std::swap(TrueOp, FalseOp);
24950 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24951 // This is efficient for any integer data type (including i8/i16) and
24953 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24954 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24955 DAG.getConstant(CC, DL, MVT::i8), Cond);
24957 // Zero extend the condition if needed.
24958 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24960 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24961 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24962 DAG.getConstant(ShAmt, DL, MVT::i8));
24963 if (N->getNumValues() == 2) // Dead flag value?
24964 return DCI.CombineTo(N, Cond, SDValue());
24968 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24969 // for any integer data type, including i8/i16.
24970 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24971 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24972 DAG.getConstant(CC, DL, MVT::i8), Cond);
24974 // Zero extend the condition if needed.
24975 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24976 FalseC->getValueType(0), Cond);
24977 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24978 SDValue(FalseC, 0));
24980 if (N->getNumValues() == 2) // Dead flag value?
24981 return DCI.CombineTo(N, Cond, SDValue());
24985 // Optimize cases that will turn into an LEA instruction. This requires
24986 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24987 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24988 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24989 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24991 bool isFastMultiplier = false;
24993 switch ((unsigned char)Diff) {
24995 case 1: // result = add base, cond
24996 case 2: // result = lea base( , cond*2)
24997 case 3: // result = lea base(cond, cond*2)
24998 case 4: // result = lea base( , cond*4)
24999 case 5: // result = lea base(cond, cond*4)
25000 case 8: // result = lea base( , cond*8)
25001 case 9: // result = lea base(cond, cond*8)
25002 isFastMultiplier = true;
25007 if (isFastMultiplier) {
25008 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
25009 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
25010 DAG.getConstant(CC, DL, MVT::i8), Cond);
25011 // Zero extend the condition if needed.
25012 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
25014 // Scale the condition by the difference.
25016 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
25017 DAG.getConstant(Diff, DL, Cond.getValueType()));
25019 // Add the base if non-zero.
25020 if (FalseC->getAPIntValue() != 0)
25021 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
25022 SDValue(FalseC, 0));
25023 if (N->getNumValues() == 2) // Dead flag value?
25024 return DCI.CombineTo(N, Cond, SDValue());
25031 // Handle these cases:
25032 // (select (x != c), e, c) -> select (x != c), e, x),
25033 // (select (x == c), c, e) -> select (x == c), x, e)
25034 // where the c is an integer constant, and the "select" is the combination
25035 // of CMOV and CMP.
25037 // The rationale for this change is that the conditional-move from a constant
25038 // needs two instructions, however, conditional-move from a register needs
25039 // only one instruction.
25041 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
25042 // some instruction-combining opportunities. This opt needs to be
25043 // postponed as late as possible.
25045 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
25046 // the DCI.xxxx conditions are provided to postpone the optimization as
25047 // late as possible.
25049 ConstantSDNode *CmpAgainst = nullptr;
25050 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
25051 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
25052 !isa<ConstantSDNode>(Cond.getOperand(0))) {
25054 if (CC == X86::COND_NE &&
25055 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
25056 CC = X86::GetOppositeBranchCondition(CC);
25057 std::swap(TrueOp, FalseOp);
25060 if (CC == X86::COND_E &&
25061 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
25062 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
25063 DAG.getConstant(CC, DL, MVT::i8), Cond };
25064 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
25069 // Fold and/or of setcc's to double CMOV:
25070 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
25071 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
25073 // This combine lets us generate:
25074 // cmovcc1 (jcc1 if we don't have CMOV)
25080 // cmovne (jne if we don't have CMOV)
25081 // When we can't use the CMOV instruction, it might increase branch
25083 // When we can use CMOV, or when there is no mispredict, this improves
25084 // throughput and reduces register pressure.
25086 if (CC == X86::COND_NE) {
25088 X86::CondCode CC0, CC1;
25090 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
25092 std::swap(FalseOp, TrueOp);
25093 CC0 = X86::GetOppositeBranchCondition(CC0);
25094 CC1 = X86::GetOppositeBranchCondition(CC1);
25097 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
25099 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
25100 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
25101 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
25102 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
25110 /// PerformMulCombine - Optimize a single multiply with constant into two
25111 /// in order to implement it with two cheaper instructions, e.g.
25112 /// LEA + SHL, LEA + LEA.
25113 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
25114 TargetLowering::DAGCombinerInfo &DCI) {
25115 // An imul is usually smaller than the alternative sequence.
25116 if (DAG.getMachineFunction().getFunction()->optForMinSize())
25119 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
25122 EVT VT = N->getValueType(0);
25123 if (VT != MVT::i64 && VT != MVT::i32)
25126 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
25129 uint64_t MulAmt = C->getZExtValue();
25130 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
25133 uint64_t MulAmt1 = 0;
25134 uint64_t MulAmt2 = 0;
25135 if ((MulAmt % 9) == 0) {
25137 MulAmt2 = MulAmt / 9;
25138 } else if ((MulAmt % 5) == 0) {
25140 MulAmt2 = MulAmt / 5;
25141 } else if ((MulAmt % 3) == 0) {
25143 MulAmt2 = MulAmt / 3;
25149 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
25151 if (isPowerOf2_64(MulAmt2) &&
25152 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
25153 // If second multiplifer is pow2, issue it first. We want the multiply by
25154 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
25156 std::swap(MulAmt1, MulAmt2);
25158 if (isPowerOf2_64(MulAmt1))
25159 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25160 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
25162 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
25163 DAG.getConstant(MulAmt1, DL, VT));
25165 if (isPowerOf2_64(MulAmt2))
25166 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
25167 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
25169 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
25170 DAG.getConstant(MulAmt2, DL, VT));
25174 assert(MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX)
25175 && "Both cases that could cause potential overflows should have "
25176 "already been handled.");
25177 if (isPowerOf2_64(MulAmt - 1))
25178 // (mul x, 2^N + 1) => (add (shl x, N), x)
25179 NewMul = DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
25180 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25181 DAG.getConstant(Log2_64(MulAmt - 1), DL,
25184 else if (isPowerOf2_64(MulAmt + 1))
25185 // (mul x, 2^N - 1) => (sub (shl x, N), x)
25186 NewMul = DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::SHL, DL, VT,
25188 DAG.getConstant(Log2_64(MulAmt + 1),
25189 DL, MVT::i8)), N->getOperand(0));
25193 // Do not add new nodes to DAG combiner worklist.
25194 DCI.CombineTo(N, NewMul, false);
25199 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
25200 SDValue N0 = N->getOperand(0);
25201 SDValue N1 = N->getOperand(1);
25202 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
25203 EVT VT = N0.getValueType();
25205 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
25206 // since the result of setcc_c is all zero's or all ones.
25207 if (VT.isInteger() && !VT.isVector() &&
25208 N1C && N0.getOpcode() == ISD::AND &&
25209 N0.getOperand(1).getOpcode() == ISD::Constant) {
25210 SDValue N00 = N0.getOperand(0);
25211 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
25212 APInt ShAmt = N1C->getAPIntValue();
25213 Mask = Mask.shl(ShAmt);
25214 bool MaskOK = false;
25215 // We can handle cases concerning bit-widening nodes containing setcc_c if
25216 // we carefully interrogate the mask to make sure we are semantics
25218 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
25219 // of the underlying setcc_c operation if the setcc_c was zero extended.
25220 // Consider the following example:
25221 // zext(setcc_c) -> i32 0x0000FFFF
25222 // c1 -> i32 0x0000FFFF
25223 // c2 -> i32 0x00000001
25224 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
25225 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
25226 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25228 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
25229 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25231 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
25232 N00.getOpcode() == ISD::ANY_EXTEND) &&
25233 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25234 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
25236 if (MaskOK && Mask != 0) {
25238 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
25242 // Hardware support for vector shifts is sparse which makes us scalarize the
25243 // vector operations in many cases. Also, on sandybridge ADD is faster than
25245 // (shl V, 1) -> add V,V
25246 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
25247 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
25248 assert(N0.getValueType().isVector() && "Invalid vector shift type");
25249 // We shift all of the values by one. In many cases we do not have
25250 // hardware support for this operation. This is better expressed as an ADD
25252 if (N1SplatC->getAPIntValue() == 1)
25253 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
25259 static SDValue PerformSRACombine(SDNode *N, SelectionDAG &DAG) {
25260 SDValue N0 = N->getOperand(0);
25261 SDValue N1 = N->getOperand(1);
25262 EVT VT = N0.getValueType();
25263 unsigned Size = VT.getSizeInBits();
25265 // fold (ashr (shl, a, [56,48,32,24,16]), SarConst)
25266 // into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or
25267 // into (lshr, (sext (a), SarConst - [56,48,32,24,16]))
25268 // depending on sign of (SarConst - [56,48,32,24,16])
25270 // sexts in X86 are MOVs. The MOVs have the same code size
25271 // as above SHIFTs (only SHIFT on 1 has lower code size).
25272 // However the MOVs have 2 advantages to a SHIFT:
25273 // 1. MOVs can write to a register that differs from source
25274 // 2. MOVs accept memory operands
25276 if (!VT.isInteger() || VT.isVector() || N1.getOpcode() != ISD::Constant ||
25277 N0.getOpcode() != ISD::SHL || !N0.hasOneUse() ||
25278 N0.getOperand(1).getOpcode() != ISD::Constant)
25281 SDValue N00 = N0.getOperand(0);
25282 SDValue N01 = N0.getOperand(1);
25283 APInt ShlConst = (cast<ConstantSDNode>(N01))->getAPIntValue();
25284 APInt SarConst = (cast<ConstantSDNode>(N1))->getAPIntValue();
25285 EVT CVT = N1.getValueType();
25287 if (SarConst.isNegative())
25290 for (MVT SVT : MVT::integer_valuetypes()) {
25291 unsigned ShiftSize = SVT.getSizeInBits();
25292 // skipping types without corresponding sext/zext and
25293 // ShlConst that is not one of [56,48,32,24,16]
25294 if (ShiftSize < 8 || ShiftSize > 64 || ShlConst != Size - ShiftSize)
25298 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, N00, DAG.getValueType(SVT));
25299 SarConst = SarConst - (Size - ShiftSize);
25302 else if (SarConst.isNegative())
25303 return DAG.getNode(ISD::SHL, DL, VT, NN,
25304 DAG.getConstant(-SarConst, DL, CVT));
25306 return DAG.getNode(ISD::SRA, DL, VT, NN,
25307 DAG.getConstant(SarConst, DL, CVT));
25312 /// \brief Returns a vector of 0s if the node in input is a vector logical
25313 /// shift by a constant amount which is known to be bigger than or equal
25314 /// to the vector element size in bits.
25315 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
25316 const X86Subtarget *Subtarget) {
25317 EVT VT = N->getValueType(0);
25319 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
25320 (!Subtarget->hasInt256() ||
25321 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
25324 SDValue Amt = N->getOperand(1);
25326 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
25327 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
25328 APInt ShiftAmt = AmtSplat->getAPIntValue();
25329 unsigned MaxAmount =
25330 VT.getSimpleVT().getVectorElementType().getSizeInBits();
25332 // SSE2/AVX2 logical shifts always return a vector of 0s
25333 // if the shift amount is bigger than or equal to
25334 // the element size. The constant shift amount will be
25335 // encoded as a 8-bit immediate.
25336 if (ShiftAmt.trunc(8).uge(MaxAmount))
25337 return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL);
25343 /// PerformShiftCombine - Combine shifts.
25344 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
25345 TargetLowering::DAGCombinerInfo &DCI,
25346 const X86Subtarget *Subtarget) {
25347 if (N->getOpcode() == ISD::SHL)
25348 if (SDValue V = PerformSHLCombine(N, DAG))
25351 if (N->getOpcode() == ISD::SRA)
25352 if (SDValue V = PerformSRACombine(N, DAG))
25355 // Try to fold this logical shift into a zero vector.
25356 if (N->getOpcode() != ISD::SRA)
25357 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
25363 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
25364 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
25365 // and friends. Likewise for OR -> CMPNEQSS.
25366 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
25367 TargetLowering::DAGCombinerInfo &DCI,
25368 const X86Subtarget *Subtarget) {
25371 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
25372 // we're requiring SSE2 for both.
25373 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
25374 SDValue N0 = N->getOperand(0);
25375 SDValue N1 = N->getOperand(1);
25376 SDValue CMP0 = N0->getOperand(1);
25377 SDValue CMP1 = N1->getOperand(1);
25380 // The SETCCs should both refer to the same CMP.
25381 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
25384 SDValue CMP00 = CMP0->getOperand(0);
25385 SDValue CMP01 = CMP0->getOperand(1);
25386 EVT VT = CMP00.getValueType();
25388 if (VT == MVT::f32 || VT == MVT::f64) {
25389 bool ExpectingFlags = false;
25390 // Check for any users that want flags:
25391 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
25392 !ExpectingFlags && UI != UE; ++UI)
25393 switch (UI->getOpcode()) {
25398 ExpectingFlags = true;
25400 case ISD::CopyToReg:
25401 case ISD::SIGN_EXTEND:
25402 case ISD::ZERO_EXTEND:
25403 case ISD::ANY_EXTEND:
25407 if (!ExpectingFlags) {
25408 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
25409 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
25411 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
25412 X86::CondCode tmp = cc0;
25417 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
25418 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
25419 // FIXME: need symbolic constants for these magic numbers.
25420 // See X86ATTInstPrinter.cpp:printSSECC().
25421 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
25422 if (Subtarget->hasAVX512()) {
25423 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
25425 DAG.getConstant(x86cc, DL, MVT::i8));
25426 if (N->getValueType(0) != MVT::i1)
25427 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
25431 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
25432 CMP00.getValueType(), CMP00, CMP01,
25433 DAG.getConstant(x86cc, DL,
25436 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
25437 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
25439 if (is64BitFP && !Subtarget->is64Bit()) {
25440 // On a 32-bit target, we cannot bitcast the 64-bit float to a
25441 // 64-bit integer, since that's not a legal type. Since
25442 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
25443 // bits, but can do this little dance to extract the lowest 32 bits
25444 // and work with those going forward.
25445 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
25447 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
25448 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
25449 Vector32, DAG.getIntPtrConstant(0, DL));
25453 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
25454 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
25455 DAG.getConstant(1, DL, IntVT));
25456 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
25458 return OneBitOfTruth;
25466 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
25467 /// so it can be folded inside ANDNP.
25468 static bool CanFoldXORWithAllOnes(const SDNode *N) {
25469 EVT VT = N->getValueType(0);
25471 // Match direct AllOnes for 128 and 256-bit vectors
25472 if (ISD::isBuildVectorAllOnes(N))
25475 // Look through a bit convert.
25476 if (N->getOpcode() == ISD::BITCAST)
25477 N = N->getOperand(0).getNode();
25479 // Sometimes the operand may come from a insert_subvector building a 256-bit
25481 if (VT.is256BitVector() &&
25482 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
25483 SDValue V1 = N->getOperand(0);
25484 SDValue V2 = N->getOperand(1);
25486 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
25487 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
25488 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
25489 ISD::isBuildVectorAllOnes(V2.getNode()))
25496 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
25497 // register. In most cases we actually compare or select YMM-sized registers
25498 // and mixing the two types creates horrible code. This method optimizes
25499 // some of the transition sequences.
25500 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
25501 TargetLowering::DAGCombinerInfo &DCI,
25502 const X86Subtarget *Subtarget) {
25503 EVT VT = N->getValueType(0);
25504 if (!VT.is256BitVector())
25507 assert((N->getOpcode() == ISD::ANY_EXTEND ||
25508 N->getOpcode() == ISD::ZERO_EXTEND ||
25509 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
25511 SDValue Narrow = N->getOperand(0);
25512 EVT NarrowVT = Narrow->getValueType(0);
25513 if (!NarrowVT.is128BitVector())
25516 if (Narrow->getOpcode() != ISD::XOR &&
25517 Narrow->getOpcode() != ISD::AND &&
25518 Narrow->getOpcode() != ISD::OR)
25521 SDValue N0 = Narrow->getOperand(0);
25522 SDValue N1 = Narrow->getOperand(1);
25525 // The Left side has to be a trunc.
25526 if (N0.getOpcode() != ISD::TRUNCATE)
25529 // The type of the truncated inputs.
25530 EVT WideVT = N0->getOperand(0)->getValueType(0);
25534 // The right side has to be a 'trunc' or a constant vector.
25535 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
25536 ConstantSDNode *RHSConstSplat = nullptr;
25537 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
25538 RHSConstSplat = RHSBV->getConstantSplatNode();
25539 if (!RHSTrunc && !RHSConstSplat)
25542 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25544 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
25547 // Set N0 and N1 to hold the inputs to the new wide operation.
25548 N0 = N0->getOperand(0);
25549 if (RHSConstSplat) {
25550 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
25551 SDValue(RHSConstSplat, 0));
25552 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
25553 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
25554 } else if (RHSTrunc) {
25555 N1 = N1->getOperand(0);
25558 // Generate the wide operation.
25559 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
25560 unsigned Opcode = N->getOpcode();
25562 case ISD::ANY_EXTEND:
25564 case ISD::ZERO_EXTEND: {
25565 unsigned InBits = NarrowVT.getScalarSizeInBits();
25566 APInt Mask = APInt::getAllOnesValue(InBits);
25567 Mask = Mask.zext(VT.getScalarSizeInBits());
25568 return DAG.getNode(ISD::AND, DL, VT,
25569 Op, DAG.getConstant(Mask, DL, VT));
25571 case ISD::SIGN_EXTEND:
25572 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
25573 Op, DAG.getValueType(NarrowVT));
25575 llvm_unreachable("Unexpected opcode");
25579 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
25580 TargetLowering::DAGCombinerInfo &DCI,
25581 const X86Subtarget *Subtarget) {
25582 SDValue N0 = N->getOperand(0);
25583 SDValue N1 = N->getOperand(1);
25586 // A vector zext_in_reg may be represented as a shuffle,
25587 // feeding into a bitcast (this represents anyext) feeding into
25588 // an and with a mask.
25589 // We'd like to try to combine that into a shuffle with zero
25590 // plus a bitcast, removing the and.
25591 if (N0.getOpcode() != ISD::BITCAST ||
25592 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
25595 // The other side of the AND should be a splat of 2^C, where C
25596 // is the number of bits in the source type.
25597 if (N1.getOpcode() == ISD::BITCAST)
25598 N1 = N1.getOperand(0);
25599 if (N1.getOpcode() != ISD::BUILD_VECTOR)
25601 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
25603 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
25604 EVT SrcType = Shuffle->getValueType(0);
25606 // We expect a single-source shuffle
25607 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
25610 unsigned SrcSize = SrcType.getScalarSizeInBits();
25612 APInt SplatValue, SplatUndef;
25613 unsigned SplatBitSize;
25615 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
25616 SplatBitSize, HasAnyUndefs))
25619 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
25620 // Make sure the splat matches the mask we expect
25621 if (SplatBitSize > ResSize ||
25622 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
25625 // Make sure the input and output size make sense
25626 if (SrcSize >= ResSize || ResSize % SrcSize)
25629 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
25630 // The number of u's between each two values depends on the ratio between
25631 // the source and dest type.
25632 unsigned ZextRatio = ResSize / SrcSize;
25633 bool IsZext = true;
25634 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
25635 if (i % ZextRatio) {
25636 if (Shuffle->getMaskElt(i) > 0) {
25642 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
25643 // Expected element number
25653 // Ok, perform the transformation - replace the shuffle with
25654 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
25655 // (instead of undef) where the k elements come from the zero vector.
25656 SmallVector<int, 8> Mask;
25657 unsigned NumElems = SrcType.getVectorNumElements();
25658 for (unsigned i = 0; i < NumElems; ++i)
25660 Mask.push_back(NumElems);
25662 Mask.push_back(i / ZextRatio);
25664 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
25665 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
25666 return DAG.getBitcast(N0.getValueType(), NewShuffle);
25669 /// If both input operands of a logic op are being cast from floating point
25670 /// types, try to convert this into a floating point logic node to avoid
25671 /// unnecessary moves from SSE to integer registers.
25672 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
25673 const X86Subtarget *Subtarget) {
25674 unsigned FPOpcode = ISD::DELETED_NODE;
25675 if (N->getOpcode() == ISD::AND)
25676 FPOpcode = X86ISD::FAND;
25677 else if (N->getOpcode() == ISD::OR)
25678 FPOpcode = X86ISD::FOR;
25679 else if (N->getOpcode() == ISD::XOR)
25680 FPOpcode = X86ISD::FXOR;
25682 assert(FPOpcode != ISD::DELETED_NODE &&
25683 "Unexpected input node for FP logic conversion");
25685 EVT VT = N->getValueType(0);
25686 SDValue N0 = N->getOperand(0);
25687 SDValue N1 = N->getOperand(1);
25689 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
25690 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
25691 (Subtarget->hasSSE2() && VT == MVT::i64))) {
25692 SDValue N00 = N0.getOperand(0);
25693 SDValue N10 = N1.getOperand(0);
25694 EVT N00Type = N00.getValueType();
25695 EVT N10Type = N10.getValueType();
25696 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
25697 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
25698 return DAG.getBitcast(VT, FPLogic);
25704 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
25705 TargetLowering::DAGCombinerInfo &DCI,
25706 const X86Subtarget *Subtarget) {
25707 if (DCI.isBeforeLegalizeOps())
25710 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
25713 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25716 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25719 EVT VT = N->getValueType(0);
25720 SDValue N0 = N->getOperand(0);
25721 SDValue N1 = N->getOperand(1);
25724 // Create BEXTR instructions
25725 // BEXTR is ((X >> imm) & (2**size-1))
25726 if (VT == MVT::i32 || VT == MVT::i64) {
25727 // Check for BEXTR.
25728 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
25729 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
25730 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
25731 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25732 if (MaskNode && ShiftNode) {
25733 uint64_t Mask = MaskNode->getZExtValue();
25734 uint64_t Shift = ShiftNode->getZExtValue();
25735 if (isMask_64(Mask)) {
25736 uint64_t MaskSize = countPopulation(Mask);
25737 if (Shift + MaskSize <= VT.getSizeInBits())
25738 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25739 DAG.getConstant(Shift | (MaskSize << 8), DL,
25748 // Want to form ANDNP nodes:
25749 // 1) In the hopes of then easily combining them with OR and AND nodes
25750 // to form PBLEND/PSIGN.
25751 // 2) To match ANDN packed intrinsics
25752 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25755 // Check LHS for vnot
25756 if (N0.getOpcode() == ISD::XOR &&
25757 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25758 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25759 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25761 // Check RHS for vnot
25762 if (N1.getOpcode() == ISD::XOR &&
25763 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25764 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25765 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25770 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25771 TargetLowering::DAGCombinerInfo &DCI,
25772 const X86Subtarget *Subtarget) {
25773 if (DCI.isBeforeLegalizeOps())
25776 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25779 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25782 SDValue N0 = N->getOperand(0);
25783 SDValue N1 = N->getOperand(1);
25784 EVT VT = N->getValueType(0);
25786 // look for psign/blend
25787 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25788 if (!Subtarget->hasSSSE3() ||
25789 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25792 // Canonicalize pandn to RHS
25793 if (N0.getOpcode() == X86ISD::ANDNP)
25795 // or (and (m, y), (pandn m, x))
25796 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25797 SDValue Mask = N1.getOperand(0);
25798 SDValue X = N1.getOperand(1);
25800 if (N0.getOperand(0) == Mask)
25801 Y = N0.getOperand(1);
25802 if (N0.getOperand(1) == Mask)
25803 Y = N0.getOperand(0);
25805 // Check to see if the mask appeared in both the AND and ANDNP and
25809 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25810 // Look through mask bitcast.
25811 if (Mask.getOpcode() == ISD::BITCAST)
25812 Mask = Mask.getOperand(0);
25813 if (X.getOpcode() == ISD::BITCAST)
25814 X = X.getOperand(0);
25815 if (Y.getOpcode() == ISD::BITCAST)
25816 Y = Y.getOperand(0);
25818 EVT MaskVT = Mask.getValueType();
25820 // Validate that the Mask operand is a vector sra node.
25821 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25822 // there is no psrai.b
25823 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25824 unsigned SraAmt = ~0;
25825 if (Mask.getOpcode() == ISD::SRA) {
25826 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25827 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25828 SraAmt = AmtConst->getZExtValue();
25829 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25830 SDValue SraC = Mask.getOperand(1);
25831 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25833 if ((SraAmt + 1) != EltBits)
25838 // Now we know we at least have a plendvb with the mask val. See if
25839 // we can form a psignb/w/d.
25840 // psign = x.type == y.type == mask.type && y = sub(0, x);
25841 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25842 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25843 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25844 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25845 "Unsupported VT for PSIGN");
25846 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25847 return DAG.getBitcast(VT, Mask);
25849 // PBLENDVB only available on SSE 4.1
25850 if (!Subtarget->hasSSE41())
25853 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25855 X = DAG.getBitcast(BlendVT, X);
25856 Y = DAG.getBitcast(BlendVT, Y);
25857 Mask = DAG.getBitcast(BlendVT, Mask);
25858 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25859 return DAG.getBitcast(VT, Mask);
25863 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25866 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25867 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25869 // SHLD/SHRD instructions have lower register pressure, but on some
25870 // platforms they have higher latency than the equivalent
25871 // series of shifts/or that would otherwise be generated.
25872 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25873 // have higher latencies and we are not optimizing for size.
25874 if (!OptForSize && Subtarget->isSHLDSlow())
25877 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25879 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25881 if (!N0.hasOneUse() || !N1.hasOneUse())
25884 SDValue ShAmt0 = N0.getOperand(1);
25885 if (ShAmt0.getValueType() != MVT::i8)
25887 SDValue ShAmt1 = N1.getOperand(1);
25888 if (ShAmt1.getValueType() != MVT::i8)
25890 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25891 ShAmt0 = ShAmt0.getOperand(0);
25892 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25893 ShAmt1 = ShAmt1.getOperand(0);
25896 unsigned Opc = X86ISD::SHLD;
25897 SDValue Op0 = N0.getOperand(0);
25898 SDValue Op1 = N1.getOperand(0);
25899 if (ShAmt0.getOpcode() == ISD::SUB) {
25900 Opc = X86ISD::SHRD;
25901 std::swap(Op0, Op1);
25902 std::swap(ShAmt0, ShAmt1);
25905 unsigned Bits = VT.getSizeInBits();
25906 if (ShAmt1.getOpcode() == ISD::SUB) {
25907 SDValue Sum = ShAmt1.getOperand(0);
25908 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25909 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25910 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25911 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25912 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25913 return DAG.getNode(Opc, DL, VT,
25915 DAG.getNode(ISD::TRUNCATE, DL,
25918 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25919 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25921 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25922 return DAG.getNode(Opc, DL, VT,
25923 N0.getOperand(0), N1.getOperand(0),
25924 DAG.getNode(ISD::TRUNCATE, DL,
25931 // Generate NEG and CMOV for integer abs.
25932 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25933 EVT VT = N->getValueType(0);
25935 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25936 // 8-bit integer abs to NEG and CMOV.
25937 if (VT.isInteger() && VT.getSizeInBits() == 8)
25940 SDValue N0 = N->getOperand(0);
25941 SDValue N1 = N->getOperand(1);
25944 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25945 // and change it to SUB and CMOV.
25946 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25947 N0.getOpcode() == ISD::ADD &&
25948 N0.getOperand(1) == N1 &&
25949 N1.getOpcode() == ISD::SRA &&
25950 N1.getOperand(0) == N0.getOperand(0))
25951 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25952 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25953 // Generate SUB & CMOV.
25954 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25955 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25957 SDValue Ops[] = { N0.getOperand(0), Neg,
25958 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25959 SDValue(Neg.getNode(), 1) };
25960 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25965 // Try to turn tests against the signbit in the form of:
25966 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25969 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25970 // This is only worth doing if the output type is i8.
25971 if (N->getValueType(0) != MVT::i8)
25974 SDValue N0 = N->getOperand(0);
25975 SDValue N1 = N->getOperand(1);
25977 // We should be performing an xor against a truncated shift.
25978 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25981 // Make sure we are performing an xor against one.
25982 if (!isOneConstant(N1))
25985 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25986 SDValue Shift = N0.getOperand(0);
25987 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25990 // Make sure we are truncating from one of i16, i32 or i64.
25991 EVT ShiftTy = Shift.getValueType();
25992 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25995 // Make sure the shift amount extracts the sign bit.
25996 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25997 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
26000 // Create a greater-than comparison against -1.
26001 // N.B. Using SETGE against 0 works but we want a canonical looking
26002 // comparison, using SETGT matches up with what TranslateX86CC.
26004 SDValue ShiftOp = Shift.getOperand(0);
26005 EVT ShiftOpTy = ShiftOp.getValueType();
26006 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
26007 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
26011 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
26012 TargetLowering::DAGCombinerInfo &DCI,
26013 const X86Subtarget *Subtarget) {
26014 if (DCI.isBeforeLegalizeOps())
26017 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
26020 if (Subtarget->hasCMov())
26021 if (SDValue RV = performIntegerAbsCombine(N, DAG))
26024 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
26030 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
26031 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
26032 /// X86ISD::AVG instruction.
26033 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
26034 const X86Subtarget *Subtarget, SDLoc DL) {
26035 if (!VT.isVector() || !VT.isSimple())
26037 EVT InVT = In.getValueType();
26038 unsigned NumElems = VT.getVectorNumElements();
26040 EVT ScalarVT = VT.getVectorElementType();
26041 if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
26042 isPowerOf2_32(NumElems)))
26045 // InScalarVT is the intermediate type in AVG pattern and it should be greater
26046 // than the original input type (i8/i16).
26047 EVT InScalarVT = InVT.getVectorElementType();
26048 if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
26051 if (Subtarget->hasAVX512()) {
26052 if (VT.getSizeInBits() > 512)
26054 } else if (Subtarget->hasAVX2()) {
26055 if (VT.getSizeInBits() > 256)
26058 if (VT.getSizeInBits() > 128)
26062 // Detect the following pattern:
26064 // %1 = zext <N x i8> %a to <N x i32>
26065 // %2 = zext <N x i8> %b to <N x i32>
26066 // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
26067 // %4 = add nuw nsw <N x i32> %3, %2
26068 // %5 = lshr <N x i32> %N, <i32 1 x N>
26069 // %6 = trunc <N x i32> %5 to <N x i8>
26071 // In AVX512, the last instruction can also be a trunc store.
26073 if (In.getOpcode() != ISD::SRL)
26076 // A lambda checking the given SDValue is a constant vector and each element
26077 // is in the range [Min, Max].
26078 auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
26079 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
26080 if (!BV || !BV->isConstant())
26082 for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
26083 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
26086 uint64_t Val = C->getZExtValue();
26087 if (Val < Min || Val > Max)
26093 // Check if each element of the vector is left-shifted by one.
26094 auto LHS = In.getOperand(0);
26095 auto RHS = In.getOperand(1);
26096 if (!IsConstVectorInRange(RHS, 1, 1))
26098 if (LHS.getOpcode() != ISD::ADD)
26101 // Detect a pattern of a + b + 1 where the order doesn't matter.
26102 SDValue Operands[3];
26103 Operands[0] = LHS.getOperand(0);
26104 Operands[1] = LHS.getOperand(1);
26106 // Take care of the case when one of the operands is a constant vector whose
26107 // element is in the range [1, 256].
26108 if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
26109 Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
26110 Operands[0].getOperand(0).getValueType() == VT) {
26111 // The pattern is detected. Subtract one from the constant vector, then
26112 // demote it and emit X86ISD::AVG instruction.
26113 SDValue One = DAG.getConstant(1, DL, InScalarVT);
26114 SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
26115 SmallVector<SDValue, 8>(NumElems, One));
26116 Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
26117 Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
26118 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26122 if (Operands[0].getOpcode() == ISD::ADD)
26123 std::swap(Operands[0], Operands[1]);
26124 else if (Operands[1].getOpcode() != ISD::ADD)
26126 Operands[2] = Operands[1].getOperand(0);
26127 Operands[1] = Operands[1].getOperand(1);
26129 // Now we have three operands of two additions. Check that one of them is a
26130 // constant vector with ones, and the other two are promoted from i8/i16.
26131 for (int i = 0; i < 3; ++i) {
26132 if (!IsConstVectorInRange(Operands[i], 1, 1))
26134 std::swap(Operands[i], Operands[2]);
26136 // Check if Operands[0] and Operands[1] are results of type promotion.
26137 for (int j = 0; j < 2; ++j)
26138 if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
26139 Operands[j].getOperand(0).getValueType() != VT)
26142 // The pattern is detected, emit X86ISD::AVG instruction.
26143 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26144 Operands[1].getOperand(0));
26150 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
26151 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
26152 TargetLowering::DAGCombinerInfo &DCI,
26153 const X86Subtarget *Subtarget) {
26154 LoadSDNode *Ld = cast<LoadSDNode>(N);
26155 EVT RegVT = Ld->getValueType(0);
26156 EVT MemVT = Ld->getMemoryVT();
26158 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26160 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
26161 // into two 16-byte operations.
26162 ISD::LoadExtType Ext = Ld->getExtensionType();
26164 unsigned AddressSpace = Ld->getAddressSpace();
26165 unsigned Alignment = Ld->getAlignment();
26166 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
26167 Ext == ISD::NON_EXTLOAD &&
26168 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
26169 AddressSpace, Alignment, &Fast) && !Fast) {
26170 unsigned NumElems = RegVT.getVectorNumElements();
26174 SDValue Ptr = Ld->getBasePtr();
26175 SDValue Increment =
26176 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26178 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
26180 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26181 Ld->getPointerInfo(), Ld->isVolatile(),
26182 Ld->isNonTemporal(), Ld->isInvariant(),
26184 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26185 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26186 Ld->getPointerInfo(), Ld->isVolatile(),
26187 Ld->isNonTemporal(), Ld->isInvariant(),
26188 std::min(16U, Alignment));
26189 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
26191 Load2.getValue(1));
26193 SDValue NewVec = DAG.getUNDEF(RegVT);
26194 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
26195 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
26196 return DCI.CombineTo(N, NewVec, TF, true);
26202 /// PerformMLOADCombine - Resolve extending loads
26203 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
26204 TargetLowering::DAGCombinerInfo &DCI,
26205 const X86Subtarget *Subtarget) {
26206 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
26207 if (Mld->getExtensionType() != ISD::SEXTLOAD)
26210 EVT VT = Mld->getValueType(0);
26211 unsigned NumElems = VT.getVectorNumElements();
26212 EVT LdVT = Mld->getMemoryVT();
26215 assert(LdVT != VT && "Cannot extend to the same type");
26216 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
26217 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
26218 // From, To sizes and ElemCount must be pow of two
26219 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26220 "Unexpected size for extending masked load");
26222 unsigned SizeRatio = ToSz / FromSz;
26223 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
26225 // Create a type on which we perform the shuffle
26226 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26227 LdVT.getScalarType(), NumElems*SizeRatio);
26228 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26230 // Convert Src0 value
26231 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
26232 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
26233 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26234 for (unsigned i = 0; i != NumElems; ++i)
26235 ShuffleVec[i] = i * SizeRatio;
26237 // Can't shuffle using an illegal type.
26238 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26239 "WideVecVT should be legal");
26240 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
26241 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
26243 // Prepare the new mask
26245 SDValue Mask = Mld->getMask();
26246 if (Mask.getValueType() == VT) {
26247 // Mask and original value have the same type
26248 NewMask = DAG.getBitcast(WideVecVT, Mask);
26249 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26250 for (unsigned i = 0; i != NumElems; ++i)
26251 ShuffleVec[i] = i * SizeRatio;
26252 for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
26253 ShuffleVec[i] = NumElems * SizeRatio;
26254 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26255 DAG.getConstant(0, dl, WideVecVT),
26259 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26260 unsigned WidenNumElts = NumElems*SizeRatio;
26261 unsigned MaskNumElts = VT.getVectorNumElements();
26262 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26265 unsigned NumConcat = WidenNumElts / MaskNumElts;
26266 SmallVector<SDValue, 16> Ops(NumConcat);
26267 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26269 for (unsigned i = 1; i != NumConcat; ++i)
26272 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26275 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
26276 Mld->getBasePtr(), NewMask, WideSrc0,
26277 Mld->getMemoryVT(), Mld->getMemOperand(),
26279 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
26280 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
26282 /// PerformMSTORECombine - Resolve truncating stores
26283 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
26284 const X86Subtarget *Subtarget) {
26285 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
26286 if (!Mst->isTruncatingStore())
26289 EVT VT = Mst->getValue().getValueType();
26290 unsigned NumElems = VT.getVectorNumElements();
26291 EVT StVT = Mst->getMemoryVT();
26294 assert(StVT != VT && "Cannot truncate to the same type");
26295 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26296 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26298 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26300 // The truncating store is legal in some cases. For example
26301 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26302 // are designated for truncate store.
26303 // In this case we don't need any further transformations.
26304 if (TLI.isTruncStoreLegal(VT, StVT))
26307 // From, To sizes and ElemCount must be pow of two
26308 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26309 "Unexpected size for truncating masked store");
26310 // We are going to use the original vector elt for storing.
26311 // Accumulated smaller vector elements must be a multiple of the store size.
26312 assert (((NumElems * FromSz) % ToSz) == 0 &&
26313 "Unexpected ratio for truncating masked store");
26315 unsigned SizeRatio = FromSz / ToSz;
26316 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26318 // Create a type on which we perform the shuffle
26319 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26320 StVT.getScalarType(), NumElems*SizeRatio);
26322 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26324 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
26325 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26326 for (unsigned i = 0; i != NumElems; ++i)
26327 ShuffleVec[i] = i * SizeRatio;
26329 // Can't shuffle using an illegal type.
26330 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26331 "WideVecVT should be legal");
26333 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26334 DAG.getUNDEF(WideVecVT),
26338 SDValue Mask = Mst->getMask();
26339 if (Mask.getValueType() == VT) {
26340 // Mask and original value have the same type
26341 NewMask = DAG.getBitcast(WideVecVT, Mask);
26342 for (unsigned i = 0; i != NumElems; ++i)
26343 ShuffleVec[i] = i * SizeRatio;
26344 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
26345 ShuffleVec[i] = NumElems*SizeRatio;
26346 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26347 DAG.getConstant(0, dl, WideVecVT),
26351 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26352 unsigned WidenNumElts = NumElems*SizeRatio;
26353 unsigned MaskNumElts = VT.getVectorNumElements();
26354 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26357 unsigned NumConcat = WidenNumElts / MaskNumElts;
26358 SmallVector<SDValue, 16> Ops(NumConcat);
26359 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26361 for (unsigned i = 1; i != NumConcat; ++i)
26364 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26367 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
26368 Mst->getBasePtr(), NewMask, StVT,
26369 Mst->getMemOperand(), false);
26371 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
26372 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
26373 const X86Subtarget *Subtarget) {
26374 StoreSDNode *St = cast<StoreSDNode>(N);
26375 EVT VT = St->getValue().getValueType();
26376 EVT StVT = St->getMemoryVT();
26378 SDValue StoredVal = St->getOperand(1);
26379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26381 // If we are saving a concatenation of two XMM registers and 32-byte stores
26382 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
26384 unsigned AddressSpace = St->getAddressSpace();
26385 unsigned Alignment = St->getAlignment();
26386 if (VT.is256BitVector() && StVT == VT &&
26387 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
26388 AddressSpace, Alignment, &Fast) && !Fast) {
26389 unsigned NumElems = VT.getVectorNumElements();
26393 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
26394 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
26397 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26398 SDValue Ptr0 = St->getBasePtr();
26399 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
26401 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
26402 St->getPointerInfo(), St->isVolatile(),
26403 St->isNonTemporal(), Alignment);
26404 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
26405 St->getPointerInfo(), St->isVolatile(),
26406 St->isNonTemporal(),
26407 std::min(16U, Alignment));
26408 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
26411 // Optimize trunc store (of multiple scalars) to shuffle and store.
26412 // First, pack all of the elements in one place. Next, store to memory
26413 // in fewer chunks.
26414 if (St->isTruncatingStore() && VT.isVector()) {
26415 // Check if we can detect an AVG pattern from the truncation. If yes,
26416 // replace the trunc store by a normal store with the result of X86ISD::AVG
26419 detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
26421 return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
26422 St->getPointerInfo(), St->isVolatile(),
26423 St->isNonTemporal(), St->getAlignment());
26425 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26426 unsigned NumElems = VT.getVectorNumElements();
26427 assert(StVT != VT && "Cannot truncate to the same type");
26428 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26429 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26431 // The truncating store is legal in some cases. For example
26432 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26433 // are designated for truncate store.
26434 // In this case we don't need any further transformations.
26435 if (TLI.isTruncStoreLegal(VT, StVT))
26438 // From, To sizes and ElemCount must be pow of two
26439 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
26440 // We are going to use the original vector elt for storing.
26441 // Accumulated smaller vector elements must be a multiple of the store size.
26442 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
26444 unsigned SizeRatio = FromSz / ToSz;
26446 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26448 // Create a type on which we perform the shuffle
26449 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26450 StVT.getScalarType(), NumElems*SizeRatio);
26452 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26454 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
26455 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
26456 for (unsigned i = 0; i != NumElems; ++i)
26457 ShuffleVec[i] = i * SizeRatio;
26459 // Can't shuffle using an illegal type.
26460 if (!TLI.isTypeLegal(WideVecVT))
26463 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26464 DAG.getUNDEF(WideVecVT),
26466 // At this point all of the data is stored at the bottom of the
26467 // register. We now need to save it to mem.
26469 // Find the largest store unit
26470 MVT StoreType = MVT::i8;
26471 for (MVT Tp : MVT::integer_valuetypes()) {
26472 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
26476 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
26477 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
26478 (64 <= NumElems * ToSz))
26479 StoreType = MVT::f64;
26481 // Bitcast the original vector into a vector of store-size units
26482 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
26483 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
26484 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
26485 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
26486 SmallVector<SDValue, 8> Chains;
26487 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
26488 TLI.getPointerTy(DAG.getDataLayout()));
26489 SDValue Ptr = St->getBasePtr();
26491 // Perform one or more big stores into memory.
26492 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
26493 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
26494 StoreType, ShuffWide,
26495 DAG.getIntPtrConstant(i, dl));
26496 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
26497 St->getPointerInfo(), St->isVolatile(),
26498 St->isNonTemporal(), St->getAlignment());
26499 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26500 Chains.push_back(Ch);
26503 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
26506 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
26507 // the FP state in cases where an emms may be missing.
26508 // A preferable solution to the general problem is to figure out the right
26509 // places to insert EMMS. This qualifies as a quick hack.
26511 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
26512 if (VT.getSizeInBits() != 64)
26515 const Function *F = DAG.getMachineFunction().getFunction();
26516 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
26518 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
26519 if ((VT.isVector() ||
26520 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
26521 isa<LoadSDNode>(St->getValue()) &&
26522 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
26523 St->getChain().hasOneUse() && !St->isVolatile()) {
26524 SDNode* LdVal = St->getValue().getNode();
26525 LoadSDNode *Ld = nullptr;
26526 int TokenFactorIndex = -1;
26527 SmallVector<SDValue, 8> Ops;
26528 SDNode* ChainVal = St->getChain().getNode();
26529 // Must be a store of a load. We currently handle two cases: the load
26530 // is a direct child, and it's under an intervening TokenFactor. It is
26531 // possible to dig deeper under nested TokenFactors.
26532 if (ChainVal == LdVal)
26533 Ld = cast<LoadSDNode>(St->getChain());
26534 else if (St->getValue().hasOneUse() &&
26535 ChainVal->getOpcode() == ISD::TokenFactor) {
26536 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
26537 if (ChainVal->getOperand(i).getNode() == LdVal) {
26538 TokenFactorIndex = i;
26539 Ld = cast<LoadSDNode>(St->getValue());
26541 Ops.push_back(ChainVal->getOperand(i));
26545 if (!Ld || !ISD::isNormalLoad(Ld))
26548 // If this is not the MMX case, i.e. we are just turning i64 load/store
26549 // into f64 load/store, avoid the transformation if there are multiple
26550 // uses of the loaded value.
26551 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
26556 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
26557 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
26559 if (Subtarget->is64Bit() || F64IsLegal) {
26560 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
26561 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
26562 Ld->getPointerInfo(), Ld->isVolatile(),
26563 Ld->isNonTemporal(), Ld->isInvariant(),
26564 Ld->getAlignment());
26565 SDValue NewChain = NewLd.getValue(1);
26566 if (TokenFactorIndex != -1) {
26567 Ops.push_back(NewChain);
26568 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26570 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
26571 St->getPointerInfo(),
26572 St->isVolatile(), St->isNonTemporal(),
26573 St->getAlignment());
26576 // Otherwise, lower to two pairs of 32-bit loads / stores.
26577 SDValue LoAddr = Ld->getBasePtr();
26578 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
26579 DAG.getConstant(4, LdDL, MVT::i32));
26581 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
26582 Ld->getPointerInfo(),
26583 Ld->isVolatile(), Ld->isNonTemporal(),
26584 Ld->isInvariant(), Ld->getAlignment());
26585 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
26586 Ld->getPointerInfo().getWithOffset(4),
26587 Ld->isVolatile(), Ld->isNonTemporal(),
26589 MinAlign(Ld->getAlignment(), 4));
26591 SDValue NewChain = LoLd.getValue(1);
26592 if (TokenFactorIndex != -1) {
26593 Ops.push_back(LoLd);
26594 Ops.push_back(HiLd);
26595 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26598 LoAddr = St->getBasePtr();
26599 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
26600 DAG.getConstant(4, StDL, MVT::i32));
26602 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
26603 St->getPointerInfo(),
26604 St->isVolatile(), St->isNonTemporal(),
26605 St->getAlignment());
26606 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
26607 St->getPointerInfo().getWithOffset(4),
26609 St->isNonTemporal(),
26610 MinAlign(St->getAlignment(), 4));
26611 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
26614 // This is similar to the above case, but here we handle a scalar 64-bit
26615 // integer store that is extracted from a vector on a 32-bit target.
26616 // If we have SSE2, then we can treat it like a floating-point double
26617 // to get past legalization. The execution dependencies fixup pass will
26618 // choose the optimal machine instruction for the store if this really is
26619 // an integer or v2f32 rather than an f64.
26620 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
26621 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
26622 SDValue OldExtract = St->getOperand(1);
26623 SDValue ExtOp0 = OldExtract.getOperand(0);
26624 unsigned VecSize = ExtOp0.getValueSizeInBits();
26625 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
26626 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
26627 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
26628 BitCast, OldExtract.getOperand(1));
26629 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
26630 St->getPointerInfo(), St->isVolatile(),
26631 St->isNonTemporal(), St->getAlignment());
26637 /// Return 'true' if this vector operation is "horizontal"
26638 /// and return the operands for the horizontal operation in LHS and RHS. A
26639 /// horizontal operation performs the binary operation on successive elements
26640 /// of its first operand, then on successive elements of its second operand,
26641 /// returning the resulting values in a vector. For example, if
26642 /// A = < float a0, float a1, float a2, float a3 >
26644 /// B = < float b0, float b1, float b2, float b3 >
26645 /// then the result of doing a horizontal operation on A and B is
26646 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
26647 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
26648 /// A horizontal-op B, for some already available A and B, and if so then LHS is
26649 /// set to A, RHS to B, and the routine returns 'true'.
26650 /// Note that the binary operation should have the property that if one of the
26651 /// operands is UNDEF then the result is UNDEF.
26652 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
26653 // Look for the following pattern: if
26654 // A = < float a0, float a1, float a2, float a3 >
26655 // B = < float b0, float b1, float b2, float b3 >
26657 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
26658 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
26659 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
26660 // which is A horizontal-op B.
26662 // At least one of the operands should be a vector shuffle.
26663 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
26664 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
26667 MVT VT = LHS.getSimpleValueType();
26669 assert((VT.is128BitVector() || VT.is256BitVector()) &&
26670 "Unsupported vector type for horizontal add/sub");
26672 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
26673 // operate independently on 128-bit lanes.
26674 unsigned NumElts = VT.getVectorNumElements();
26675 unsigned NumLanes = VT.getSizeInBits()/128;
26676 unsigned NumLaneElts = NumElts / NumLanes;
26677 assert((NumLaneElts % 2 == 0) &&
26678 "Vector type should have an even number of elements in each lane");
26679 unsigned HalfLaneElts = NumLaneElts/2;
26681 // View LHS in the form
26682 // LHS = VECTOR_SHUFFLE A, B, LMask
26683 // If LHS is not a shuffle then pretend it is the shuffle
26684 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
26685 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
26688 SmallVector<int, 16> LMask(NumElts);
26689 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26690 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
26691 A = LHS.getOperand(0);
26692 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
26693 B = LHS.getOperand(1);
26694 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
26695 std::copy(Mask.begin(), Mask.end(), LMask.begin());
26697 if (LHS.getOpcode() != ISD::UNDEF)
26699 for (unsigned i = 0; i != NumElts; ++i)
26703 // Likewise, view RHS in the form
26704 // RHS = VECTOR_SHUFFLE C, D, RMask
26706 SmallVector<int, 16> RMask(NumElts);
26707 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26708 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
26709 C = RHS.getOperand(0);
26710 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
26711 D = RHS.getOperand(1);
26712 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
26713 std::copy(Mask.begin(), Mask.end(), RMask.begin());
26715 if (RHS.getOpcode() != ISD::UNDEF)
26717 for (unsigned i = 0; i != NumElts; ++i)
26721 // Check that the shuffles are both shuffling the same vectors.
26722 if (!(A == C && B == D) && !(A == D && B == C))
26725 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
26726 if (!A.getNode() && !B.getNode())
26729 // If A and B occur in reverse order in RHS, then "swap" them (which means
26730 // rewriting the mask).
26732 ShuffleVectorSDNode::commuteMask(RMask);
26734 // At this point LHS and RHS are equivalent to
26735 // LHS = VECTOR_SHUFFLE A, B, LMask
26736 // RHS = VECTOR_SHUFFLE A, B, RMask
26737 // Check that the masks correspond to performing a horizontal operation.
26738 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
26739 for (unsigned i = 0; i != NumLaneElts; ++i) {
26740 int LIdx = LMask[i+l], RIdx = RMask[i+l];
26742 // Ignore any UNDEF components.
26743 if (LIdx < 0 || RIdx < 0 ||
26744 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
26745 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
26748 // Check that successive elements are being operated on. If not, this is
26749 // not a horizontal operation.
26750 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
26751 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
26752 if (!(LIdx == Index && RIdx == Index + 1) &&
26753 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
26758 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
26759 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
26763 /// Do target-specific dag combines on floating point adds.
26764 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
26765 const X86Subtarget *Subtarget) {
26766 EVT VT = N->getValueType(0);
26767 SDValue LHS = N->getOperand(0);
26768 SDValue RHS = N->getOperand(1);
26770 // Try to synthesize horizontal adds from adds of shuffles.
26771 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26772 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26773 isHorizontalBinOp(LHS, RHS, true))
26774 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
26778 /// Do target-specific dag combines on floating point subs.
26779 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
26780 const X86Subtarget *Subtarget) {
26781 EVT VT = N->getValueType(0);
26782 SDValue LHS = N->getOperand(0);
26783 SDValue RHS = N->getOperand(1);
26785 // Try to synthesize horizontal subs from subs of shuffles.
26786 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26787 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26788 isHorizontalBinOp(LHS, RHS, false))
26789 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
26793 /// Truncate a group of v4i32 into v16i8/v8i16 using X86ISD::PACKUS.
26795 combineVectorTruncationWithPACKUS(SDNode *N, SelectionDAG &DAG,
26796 SmallVector<SDValue, 8> &Regs) {
26797 assert(Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 ||
26798 Regs[0].getValueType() == MVT::v2i64));
26799 EVT OutVT = N->getValueType(0);
26800 EVT OutSVT = OutVT.getVectorElementType();
26801 EVT InVT = Regs[0].getValueType();
26802 EVT InSVT = InVT.getVectorElementType();
26805 // First, use mask to unset all bits that won't appear in the result.
26806 assert((OutSVT == MVT::i8 || OutSVT == MVT::i16) &&
26807 "OutSVT can only be either i8 or i16.");
26809 DAG.getConstant(OutSVT == MVT::i8 ? 0xFF : 0xFFFF, DL, InSVT);
26810 SDValue MaskVec = DAG.getNode(
26811 ISD::BUILD_VECTOR, DL, InVT,
26812 SmallVector<SDValue, 8>(InVT.getVectorNumElements(), MaskVal));
26813 for (auto &Reg : Regs)
26814 Reg = DAG.getNode(ISD::AND, DL, InVT, MaskVec, Reg);
26816 MVT UnpackedVT, PackedVT;
26817 if (OutSVT == MVT::i8) {
26818 UnpackedVT = MVT::v8i16;
26819 PackedVT = MVT::v16i8;
26821 UnpackedVT = MVT::v4i32;
26822 PackedVT = MVT::v8i16;
26825 // In each iteration, truncate the type by a half size.
26826 auto RegNum = Regs.size();
26827 for (unsigned j = 1, e = InSVT.getSizeInBits() / OutSVT.getSizeInBits();
26828 j < e; j *= 2, RegNum /= 2) {
26829 for (unsigned i = 0; i < RegNum; i++)
26830 Regs[i] = DAG.getNode(ISD::BITCAST, DL, UnpackedVT, Regs[i]);
26831 for (unsigned i = 0; i < RegNum / 2; i++)
26832 Regs[i] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[i * 2],
26836 // If the type of the result is v8i8, we need do one more X86ISD::PACKUS, and
26837 // then extract a subvector as the result since v8i8 is not a legal type.
26838 if (OutVT == MVT::v8i8) {
26839 Regs[0] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[0], Regs[0]);
26840 Regs[0] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OutVT, Regs[0],
26841 DAG.getIntPtrConstant(0, DL));
26843 } else if (RegNum > 1) {
26844 Regs.resize(RegNum);
26845 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26850 /// Truncate a group of v4i32 into v8i16 using X86ISD::PACKSS.
26852 combineVectorTruncationWithPACKSS(SDNode *N, SelectionDAG &DAG,
26853 SmallVector<SDValue, 8> &Regs) {
26854 assert(Regs.size() > 0 && Regs[0].getValueType() == MVT::v4i32);
26855 EVT OutVT = N->getValueType(0);
26858 // Shift left by 16 bits, then arithmetic-shift right by 16 bits.
26859 SDValue ShAmt = DAG.getConstant(16, DL, MVT::i32);
26860 for (auto &Reg : Regs) {
26861 Reg = getTargetVShiftNode(X86ISD::VSHLI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26862 Reg = getTargetVShiftNode(X86ISD::VSRAI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26865 for (unsigned i = 0, e = Regs.size() / 2; i < e; i++)
26866 Regs[i] = DAG.getNode(X86ISD::PACKSS, DL, MVT::v8i16, Regs[i * 2],
26869 if (Regs.size() > 2) {
26870 Regs.resize(Regs.size() / 2);
26871 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26876 /// This function transforms truncation from vXi32/vXi64 to vXi8/vXi16 into
26877 /// X86ISD::PACKUS/X86ISD::PACKSS operations. We do it here because after type
26878 /// legalization the truncation will be translated into a BUILD_VECTOR with each
26879 /// element that is extracted from a vector and then truncated, and it is
26880 /// diffcult to do this optimization based on them.
26881 static SDValue combineVectorTruncation(SDNode *N, SelectionDAG &DAG,
26882 const X86Subtarget *Subtarget) {
26883 EVT OutVT = N->getValueType(0);
26884 if (!OutVT.isVector())
26887 SDValue In = N->getOperand(0);
26888 if (!In.getValueType().isSimple())
26891 EVT InVT = In.getValueType();
26892 unsigned NumElems = OutVT.getVectorNumElements();
26894 // TODO: On AVX2, the behavior of X86ISD::PACKUS is different from that on
26895 // SSE2, and we need to take care of it specially.
26896 // AVX512 provides vpmovdb.
26897 if (!Subtarget->hasSSE2() || Subtarget->hasAVX2())
26900 EVT OutSVT = OutVT.getVectorElementType();
26901 EVT InSVT = InVT.getVectorElementType();
26902 if (!((InSVT == MVT::i32 || InSVT == MVT::i64) &&
26903 (OutSVT == MVT::i8 || OutSVT == MVT::i16) && isPowerOf2_32(NumElems) &&
26907 // SSSE3's pshufb results in less instructions in the cases below.
26908 if (Subtarget->hasSSSE3() && NumElems == 8 &&
26909 ((OutSVT == MVT::i8 && InSVT != MVT::i64) ||
26910 (InSVT == MVT::i32 && OutSVT == MVT::i16)))
26915 // Split a long vector into vectors of legal type.
26916 unsigned RegNum = InVT.getSizeInBits() / 128;
26917 SmallVector<SDValue, 8> SubVec(RegNum);
26918 if (InSVT == MVT::i32) {
26919 for (unsigned i = 0; i < RegNum; i++)
26920 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
26921 DAG.getIntPtrConstant(i * 4, DL));
26923 for (unsigned i = 0; i < RegNum; i++)
26924 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
26925 DAG.getIntPtrConstant(i * 2, DL));
26928 // SSE2 provides PACKUS for only 2 x v8i16 -> v16i8 and SSE4.1 provides PAKCUS
26929 // for 2 x v4i32 -> v8i16. For SSSE3 and below, we need to use PACKSS to
26930 // truncate 2 x v4i32 to v8i16.
26931 if (Subtarget->hasSSE41() || OutSVT == MVT::i8)
26932 return combineVectorTruncationWithPACKUS(N, DAG, SubVec);
26933 else if (InSVT == MVT::i32)
26934 return combineVectorTruncationWithPACKSS(N, DAG, SubVec);
26939 static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
26940 const X86Subtarget *Subtarget) {
26941 // Try to detect AVG pattern first.
26942 SDValue Avg = detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG,
26943 Subtarget, SDLoc(N));
26947 return combineVectorTruncation(N, DAG, Subtarget);
26950 /// Do target-specific dag combines on floating point negations.
26951 static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
26952 const X86Subtarget *Subtarget) {
26953 EVT VT = N->getValueType(0);
26954 EVT SVT = VT.getScalarType();
26955 SDValue Arg = N->getOperand(0);
26958 // Let legalize expand this if it isn't a legal type yet.
26959 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26962 // If we're negating a FMUL node on a target with FMA, then we can avoid the
26963 // use of a constant by performing (-0 - A*B) instead.
26964 // FIXME: Check rounding control flags as well once it becomes available.
26965 if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) &&
26966 Arg->getFlags()->hasNoSignedZeros() && Subtarget->hasAnyFMA()) {
26967 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
26968 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26969 Arg.getOperand(1), Zero);
26972 // If we're negating a FMA node, then we can adjust the
26973 // instruction to include the extra negation.
26974 if (Arg.hasOneUse()) {
26975 switch (Arg.getOpcode()) {
26976 case X86ISD::FMADD:
26977 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26978 Arg.getOperand(1), Arg.getOperand(2));
26979 case X86ISD::FMSUB:
26980 return DAG.getNode(X86ISD::FNMADD, DL, VT, Arg.getOperand(0),
26981 Arg.getOperand(1), Arg.getOperand(2));
26982 case X86ISD::FNMADD:
26983 return DAG.getNode(X86ISD::FMSUB, DL, VT, Arg.getOperand(0),
26984 Arg.getOperand(1), Arg.getOperand(2));
26985 case X86ISD::FNMSUB:
26986 return DAG.getNode(X86ISD::FMADD, DL, VT, Arg.getOperand(0),
26987 Arg.getOperand(1), Arg.getOperand(2));
26993 static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
26994 const X86Subtarget *Subtarget) {
26995 EVT VT = N->getValueType(0);
26996 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
26997 // VXORPS, VORPS, VANDPS, VANDNPS are supported only under DQ extention.
26998 // These logic operations may be executed in the integer domain.
27000 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
27001 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
27003 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
27004 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
27005 unsigned IntOpcode = 0;
27006 switch (N->getOpcode()) {
27007 default: llvm_unreachable("Unexpected FP logic op");
27008 case X86ISD::FOR: IntOpcode = ISD::OR; break;
27009 case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
27010 case X86ISD::FAND: IntOpcode = ISD::AND; break;
27011 case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
27013 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
27014 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
27018 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
27019 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
27020 const X86Subtarget *Subtarget) {
27021 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
27023 // F[X]OR(0.0, x) -> x
27024 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27025 if (C->getValueAPF().isPosZero())
27026 return N->getOperand(1);
27028 // F[X]OR(x, 0.0) -> x
27029 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27030 if (C->getValueAPF().isPosZero())
27031 return N->getOperand(0);
27033 return lowerX86FPLogicOp(N, DAG, Subtarget);
27036 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
27037 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
27038 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
27040 // Only perform optimizations if UnsafeMath is used.
27041 if (!DAG.getTarget().Options.UnsafeFPMath)
27044 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
27045 // into FMINC and FMAXC, which are Commutative operations.
27046 unsigned NewOp = 0;
27047 switch (N->getOpcode()) {
27048 default: llvm_unreachable("unknown opcode");
27049 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
27050 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
27053 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
27054 N->getOperand(0), N->getOperand(1));
27057 static SDValue performFMinNumFMaxNumCombine(SDNode *N, SelectionDAG &DAG,
27058 const X86Subtarget *Subtarget) {
27059 if (Subtarget->useSoftFloat())
27062 // TODO: Check for global or instruction-level "nnan". In that case, we
27063 // should be able to lower to FMAX/FMIN alone.
27064 // TODO: If an operand is already known to be a NaN or not a NaN, this
27065 // should be an optional swap and FMAX/FMIN.
27067 EVT VT = N->getValueType(0);
27068 if (!((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
27069 (Subtarget->hasSSE2() && (VT == MVT::f64 || VT == MVT::v2f64)) ||
27070 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))))
27073 // This takes at least 3 instructions, so favor a library call when operating
27074 // on a scalar and minimizing code size.
27075 if (!VT.isVector() && DAG.getMachineFunction().getFunction()->optForMinSize())
27078 SDValue Op0 = N->getOperand(0);
27079 SDValue Op1 = N->getOperand(1);
27081 EVT SetCCType = DAG.getTargetLoweringInfo().getSetCCResultType(
27082 DAG.getDataLayout(), *DAG.getContext(), VT);
27084 // There are 4 possibilities involving NaN inputs, and these are the required
27088 // ----------------
27089 // Num | Max | Op0 |
27090 // Op0 ----------------
27091 // NaN | Op1 | NaN |
27092 // ----------------
27094 // The SSE FP max/min instructions were not designed for this case, but rather
27096 // Min = Op1 < Op0 ? Op1 : Op0
27097 // Max = Op1 > Op0 ? Op1 : Op0
27099 // So they always return Op0 if either input is a NaN. However, we can still
27100 // use those instructions for fmaxnum by selecting away a NaN input.
27102 // If either operand is NaN, the 2nd source operand (Op0) is passed through.
27103 auto MinMaxOp = N->getOpcode() == ISD::FMAXNUM ? X86ISD::FMAX : X86ISD::FMIN;
27104 SDValue MinOrMax = DAG.getNode(MinMaxOp, DL, VT, Op1, Op0);
27105 SDValue IsOp0Nan = DAG.getSetCC(DL, SetCCType , Op0, Op0, ISD::SETUO);
27107 // If Op0 is a NaN, select Op1. Otherwise, select the max. If both operands
27108 // are NaN, the NaN value of Op1 is the result.
27109 auto SelectOpcode = VT.isVector() ? ISD::VSELECT : ISD::SELECT;
27110 return DAG.getNode(SelectOpcode, DL, VT, IsOp0Nan, Op1, MinOrMax);
27113 /// Do target-specific dag combines on X86ISD::FAND nodes.
27114 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG,
27115 const X86Subtarget *Subtarget) {
27116 // FAND(0.0, x) -> 0.0
27117 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27118 if (C->getValueAPF().isPosZero())
27119 return N->getOperand(0);
27121 // FAND(x, 0.0) -> 0.0
27122 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27123 if (C->getValueAPF().isPosZero())
27124 return N->getOperand(1);
27126 return lowerX86FPLogicOp(N, DAG, Subtarget);
27129 /// Do target-specific dag combines on X86ISD::FANDN nodes
27130 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG,
27131 const X86Subtarget *Subtarget) {
27132 // FANDN(0.0, x) -> x
27133 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27134 if (C->getValueAPF().isPosZero())
27135 return N->getOperand(1);
27137 // FANDN(x, 0.0) -> 0.0
27138 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27139 if (C->getValueAPF().isPosZero())
27140 return N->getOperand(1);
27142 return lowerX86FPLogicOp(N, DAG, Subtarget);
27145 static SDValue PerformBTCombine(SDNode *N,
27147 TargetLowering::DAGCombinerInfo &DCI) {
27148 // BT ignores high bits in the bit index operand.
27149 SDValue Op1 = N->getOperand(1);
27150 if (Op1.hasOneUse()) {
27151 unsigned BitWidth = Op1.getValueSizeInBits();
27152 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
27153 APInt KnownZero, KnownOne;
27154 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
27155 !DCI.isBeforeLegalizeOps());
27156 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27157 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
27158 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
27159 DCI.CommitTargetLoweringOpt(TLO);
27164 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
27165 SDValue Op = N->getOperand(0);
27166 if (Op.getOpcode() == ISD::BITCAST)
27167 Op = Op.getOperand(0);
27168 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
27169 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
27170 VT.getVectorElementType().getSizeInBits() ==
27171 OpVT.getVectorElementType().getSizeInBits()) {
27172 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
27177 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
27178 const X86Subtarget *Subtarget) {
27179 EVT VT = N->getValueType(0);
27180 if (!VT.isVector())
27183 SDValue N0 = N->getOperand(0);
27184 SDValue N1 = N->getOperand(1);
27185 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
27188 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
27189 // both SSE and AVX2 since there is no sign-extended shift right
27190 // operation on a vector with 64-bit elements.
27191 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
27192 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
27193 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
27194 N0.getOpcode() == ISD::SIGN_EXTEND)) {
27195 SDValue N00 = N0.getOperand(0);
27197 // EXTLOAD has a better solution on AVX2,
27198 // it may be replaced with X86ISD::VSEXT node.
27199 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
27200 if (!ISD::isNormalLoad(N00.getNode()))
27203 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
27204 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
27206 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
27212 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
27213 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
27214 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
27215 /// eliminate extend, add, and shift instructions.
27216 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
27217 const X86Subtarget *Subtarget) {
27218 // TODO: This should be valid for other integer types.
27219 EVT VT = Sext->getValueType(0);
27220 if (VT != MVT::i64)
27223 // We need an 'add nsw' feeding into the 'sext'.
27224 SDValue Add = Sext->getOperand(0);
27225 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
27228 // Having a constant operand to the 'add' ensures that we are not increasing
27229 // the instruction count because the constant is extended for free below.
27230 // A constant operand can also become the displacement field of an LEA.
27231 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
27235 // Don't make the 'add' bigger if there's no hope of combining it with some
27236 // other 'add' or 'shl' instruction.
27237 // TODO: It may be profitable to generate simpler LEA instructions in place
27238 // of single 'add' instructions, but the cost model for selecting an LEA
27239 // currently has a high threshold.
27240 bool HasLEAPotential = false;
27241 for (auto *User : Sext->uses()) {
27242 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
27243 HasLEAPotential = true;
27247 if (!HasLEAPotential)
27250 // Everything looks good, so pull the 'sext' ahead of the 'add'.
27251 int64_t AddConstant = AddOp1->getSExtValue();
27252 SDValue AddOp0 = Add.getOperand(0);
27253 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
27254 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
27256 // The wider add is guaranteed to not wrap because both operands are
27259 Flags.setNoSignedWrap(true);
27260 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
27263 /// (i8,i32 {s/z}ext ({s/u}divrem (i8 x, i8 y)) ->
27264 /// (i8,i32 ({s/u}divrem_sext_hreg (i8 x, i8 y)
27265 /// This exposes the {s/z}ext to the sdivrem lowering, so that it directly
27266 /// extends from AH (which we otherwise need to do contortions to access).
27267 static SDValue getDivRem8(SDNode *N, SelectionDAG &DAG) {
27268 SDValue N0 = N->getOperand(0);
27269 auto OpcodeN = N->getOpcode();
27270 auto OpcodeN0 = N0.getOpcode();
27271 if (!((OpcodeN == ISD::SIGN_EXTEND && OpcodeN0 == ISD::SDIVREM) ||
27272 (OpcodeN == ISD::ZERO_EXTEND && OpcodeN0 == ISD::UDIVREM)))
27275 EVT VT = N->getValueType(0);
27276 EVT InVT = N0.getValueType();
27277 if (N0.getResNo() != 1 || InVT != MVT::i8 || VT != MVT::i32)
27280 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
27281 auto DivRemOpcode = OpcodeN0 == ISD::SDIVREM ? X86ISD::SDIVREM8_SEXT_HREG
27282 : X86ISD::UDIVREM8_ZEXT_HREG;
27283 SDValue R = DAG.getNode(DivRemOpcode, SDLoc(N), NodeTys, N0.getOperand(0),
27285 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
27286 return R.getValue(1);
27289 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
27290 TargetLowering::DAGCombinerInfo &DCI,
27291 const X86Subtarget *Subtarget) {
27292 SDValue N0 = N->getOperand(0);
27293 EVT VT = N->getValueType(0);
27294 EVT SVT = VT.getScalarType();
27295 EVT InVT = N0.getValueType();
27296 EVT InSVT = InVT.getScalarType();
27299 if (SDValue DivRem8 = getDivRem8(N, DAG))
27302 if (!DCI.isBeforeLegalizeOps()) {
27303 if (InVT == MVT::i1) {
27304 SDValue Zero = DAG.getConstant(0, DL, VT);
27306 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
27307 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
27312 if (VT.isVector() && Subtarget->hasSSE2()) {
27313 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
27314 EVT InVT = N.getValueType();
27315 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
27316 Size / InVT.getScalarSizeInBits());
27317 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
27318 DAG.getUNDEF(InVT));
27320 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
27323 // If target-size is less than 128-bits, extend to a type that would extend
27324 // to 128 bits, extend that and extract the original target vector.
27325 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
27326 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27327 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27328 unsigned Scale = 128 / VT.getSizeInBits();
27330 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
27331 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
27332 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
27333 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
27334 DAG.getIntPtrConstant(0, DL));
27337 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
27338 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
27339 if (VT.getSizeInBits() == 128 &&
27340 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27341 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27342 SDValue ExOp = ExtendVecSize(DL, N0, 128);
27343 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
27346 // On pre-AVX2 targets, split into 128-bit nodes of
27347 // ISD::SIGN_EXTEND_VECTOR_INREG.
27348 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
27349 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27350 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27351 unsigned NumVecs = VT.getSizeInBits() / 128;
27352 unsigned NumSubElts = 128 / SVT.getSizeInBits();
27353 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
27354 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
27356 SmallVector<SDValue, 8> Opnds;
27357 for (unsigned i = 0, Offset = 0; i != NumVecs;
27358 ++i, Offset += NumSubElts) {
27359 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
27360 DAG.getIntPtrConstant(Offset, DL));
27361 SrcVec = ExtendVecSize(DL, SrcVec, 128);
27362 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
27363 Opnds.push_back(SrcVec);
27365 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
27369 if (Subtarget->hasAVX() && VT.is256BitVector())
27370 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27373 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
27379 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
27380 const X86Subtarget* Subtarget) {
27382 EVT VT = N->getValueType(0);
27384 // Let legalize expand this if it isn't a legal type yet.
27385 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
27388 EVT ScalarVT = VT.getScalarType();
27389 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
27392 SDValue A = N->getOperand(0);
27393 SDValue B = N->getOperand(1);
27394 SDValue C = N->getOperand(2);
27396 bool NegA = (A.getOpcode() == ISD::FNEG);
27397 bool NegB = (B.getOpcode() == ISD::FNEG);
27398 bool NegC = (C.getOpcode() == ISD::FNEG);
27400 // Negative multiplication when NegA xor NegB
27401 bool NegMul = (NegA != NegB);
27403 A = A.getOperand(0);
27405 B = B.getOperand(0);
27407 C = C.getOperand(0);
27411 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
27413 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
27415 return DAG.getNode(Opcode, dl, VT, A, B, C);
27418 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
27419 TargetLowering::DAGCombinerInfo &DCI,
27420 const X86Subtarget *Subtarget) {
27421 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
27422 // (and (i32 x86isd::setcc_carry), 1)
27423 // This eliminates the zext. This transformation is necessary because
27424 // ISD::SETCC is always legalized to i8.
27426 SDValue N0 = N->getOperand(0);
27427 EVT VT = N->getValueType(0);
27429 if (N0.getOpcode() == ISD::AND &&
27431 N0.getOperand(0).hasOneUse()) {
27432 SDValue N00 = N0.getOperand(0);
27433 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27434 if (!isOneConstant(N0.getOperand(1)))
27436 return DAG.getNode(ISD::AND, dl, VT,
27437 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27438 N00.getOperand(0), N00.getOperand(1)),
27439 DAG.getConstant(1, dl, VT));
27443 if (N0.getOpcode() == ISD::TRUNCATE &&
27445 N0.getOperand(0).hasOneUse()) {
27446 SDValue N00 = N0.getOperand(0);
27447 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27448 return DAG.getNode(ISD::AND, dl, VT,
27449 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27450 N00.getOperand(0), N00.getOperand(1)),
27451 DAG.getConstant(1, dl, VT));
27455 if (VT.is256BitVector())
27456 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27459 if (SDValue DivRem8 = getDivRem8(N, DAG))
27465 // Optimize x == -y --> x+y == 0
27466 // x != -y --> x+y != 0
27467 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
27468 const X86Subtarget* Subtarget) {
27469 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
27470 SDValue LHS = N->getOperand(0);
27471 SDValue RHS = N->getOperand(1);
27472 EVT VT = N->getValueType(0);
27475 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
27476 if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
27477 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
27478 LHS.getOperand(1));
27479 return DAG.getSetCC(DL, N->getValueType(0), addV,
27480 DAG.getConstant(0, DL, addV.getValueType()), CC);
27482 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
27483 if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
27484 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
27485 RHS.getOperand(1));
27486 return DAG.getSetCC(DL, N->getValueType(0), addV,
27487 DAG.getConstant(0, DL, addV.getValueType()), CC);
27490 if (VT.getScalarType() == MVT::i1 &&
27491 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
27493 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27494 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27495 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27497 if (!IsSEXT0 || !IsVZero1) {
27498 // Swap the operands and update the condition code.
27499 std::swap(LHS, RHS);
27500 CC = ISD::getSetCCSwappedOperands(CC);
27502 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27503 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27504 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27507 if (IsSEXT0 && IsVZero1) {
27508 assert(VT == LHS.getOperand(0).getValueType() &&
27509 "Uexpected operand type");
27510 if (CC == ISD::SETGT)
27511 return DAG.getConstant(0, DL, VT);
27512 if (CC == ISD::SETLE)
27513 return DAG.getConstant(1, DL, VT);
27514 if (CC == ISD::SETEQ || CC == ISD::SETGE)
27515 return DAG.getNOT(DL, LHS.getOperand(0), VT);
27517 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
27518 "Unexpected condition code!");
27519 return LHS.getOperand(0);
27526 static SDValue PerformGatherScatterCombine(SDNode *N, SelectionDAG &DAG) {
27528 // Gather and Scatter instructions use k-registers for masks. The type of
27529 // the masks is v*i1. So the mask will be truncated anyway.
27530 // The SIGN_EXTEND_INREG my be dropped.
27531 SDValue Mask = N->getOperand(2);
27532 if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) {
27533 SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end());
27534 NewOps[2] = Mask.getOperand(0);
27535 DAG.UpdateNodeOperands(N, NewOps);
27540 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
27541 // as "sbb reg,reg", since it can be extended without zext and produces
27542 // an all-ones bit which is more useful than 0/1 in some cases.
27543 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
27546 return DAG.getNode(ISD::AND, DL, VT,
27547 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27548 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27550 DAG.getConstant(1, DL, VT));
27551 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
27552 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
27553 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27554 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27558 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
27559 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
27560 TargetLowering::DAGCombinerInfo &DCI,
27561 const X86Subtarget *Subtarget) {
27563 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
27564 SDValue EFLAGS = N->getOperand(1);
27566 if (CC == X86::COND_A) {
27567 // Try to convert COND_A into COND_B in an attempt to facilitate
27568 // materializing "setb reg".
27570 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
27571 // cannot take an immediate as its first operand.
27573 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
27574 EFLAGS.getValueType().isInteger() &&
27575 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
27576 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
27577 EFLAGS.getNode()->getVTList(),
27578 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
27579 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
27580 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
27584 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
27585 // a zext and produces an all-ones bit which is more useful than 0/1 in some
27587 if (CC == X86::COND_B)
27588 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
27590 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27591 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27592 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
27598 // Optimize branch condition evaluation.
27600 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
27601 TargetLowering::DAGCombinerInfo &DCI,
27602 const X86Subtarget *Subtarget) {
27604 SDValue Chain = N->getOperand(0);
27605 SDValue Dest = N->getOperand(1);
27606 SDValue EFLAGS = N->getOperand(3);
27607 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
27609 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27610 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27611 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
27618 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
27619 SelectionDAG &DAG) {
27620 // Take advantage of vector comparisons producing 0 or -1 in each lane to
27621 // optimize away operation when it's from a constant.
27623 // The general transformation is:
27624 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
27625 // AND(VECTOR_CMP(x,y), constant2)
27626 // constant2 = UNARYOP(constant)
27628 // Early exit if this isn't a vector operation, the operand of the
27629 // unary operation isn't a bitwise AND, or if the sizes of the operations
27630 // aren't the same.
27631 EVT VT = N->getValueType(0);
27632 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
27633 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
27634 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
27637 // Now check that the other operand of the AND is a constant. We could
27638 // make the transformation for non-constant splats as well, but it's unclear
27639 // that would be a benefit as it would not eliminate any operations, just
27640 // perform one more step in scalar code before moving to the vector unit.
27641 if (BuildVectorSDNode *BV =
27642 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
27643 // Bail out if the vector isn't a constant.
27644 if (!BV->isConstant())
27647 // Everything checks out. Build up the new and improved node.
27649 EVT IntVT = BV->getValueType(0);
27650 // Create a new constant of the appropriate type for the transformed
27652 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
27653 // The AND node needs bitcasts to/from an integer vector type around it.
27654 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
27655 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
27656 N->getOperand(0)->getOperand(0), MaskConst);
27657 SDValue Res = DAG.getBitcast(VT, NewAnd);
27664 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27665 const X86Subtarget *Subtarget) {
27666 SDValue Op0 = N->getOperand(0);
27667 EVT VT = N->getValueType(0);
27668 EVT InVT = Op0.getValueType();
27669 EVT InSVT = InVT.getScalarType();
27670 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27672 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
27673 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
27674 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27676 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27677 InVT.getVectorNumElements());
27678 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
27680 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
27681 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
27683 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27689 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27690 const X86Subtarget *Subtarget) {
27691 // First try to optimize away the conversion entirely when it's
27692 // conditionally from a constant. Vectors only.
27693 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
27696 // Now move on to more general possibilities.
27697 SDValue Op0 = N->getOperand(0);
27698 EVT VT = N->getValueType(0);
27699 EVT InVT = Op0.getValueType();
27700 EVT InSVT = InVT.getScalarType();
27702 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
27703 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
27704 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27706 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27707 InVT.getVectorNumElements());
27708 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
27709 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27712 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
27713 // a 32-bit target where SSE doesn't support i64->FP operations.
27714 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
27715 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
27716 EVT LdVT = Ld->getValueType(0);
27718 // This transformation is not supported if the result type is f16
27719 if (VT == MVT::f16)
27722 if (!Ld->isVolatile() && !VT.isVector() &&
27723 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
27724 !Subtarget->is64Bit() && LdVT == MVT::i64) {
27725 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
27726 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
27727 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
27734 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
27735 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
27736 X86TargetLowering::DAGCombinerInfo &DCI) {
27737 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
27738 // the result is either zero or one (depending on the input carry bit).
27739 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
27740 if (X86::isZeroNode(N->getOperand(0)) &&
27741 X86::isZeroNode(N->getOperand(1)) &&
27742 // We don't have a good way to replace an EFLAGS use, so only do this when
27744 SDValue(N, 1).use_empty()) {
27746 EVT VT = N->getValueType(0);
27747 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
27748 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
27749 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
27750 DAG.getConstant(X86::COND_B, DL,
27753 DAG.getConstant(1, DL, VT));
27754 return DCI.CombineTo(N, Res1, CarryOut);
27760 // fold (add Y, (sete X, 0)) -> adc 0, Y
27761 // (add Y, (setne X, 0)) -> sbb -1, Y
27762 // (sub (sete X, 0), Y) -> sbb 0, Y
27763 // (sub (setne X, 0), Y) -> adc -1, Y
27764 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
27767 // Look through ZExts.
27768 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
27769 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
27772 SDValue SetCC = Ext.getOperand(0);
27773 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
27776 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
27777 if (CC != X86::COND_E && CC != X86::COND_NE)
27780 SDValue Cmp = SetCC.getOperand(1);
27781 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
27782 !X86::isZeroNode(Cmp.getOperand(1)) ||
27783 !Cmp.getOperand(0).getValueType().isInteger())
27786 SDValue CmpOp0 = Cmp.getOperand(0);
27787 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
27788 DAG.getConstant(1, DL, CmpOp0.getValueType()));
27790 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
27791 if (CC == X86::COND_NE)
27792 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
27793 DL, OtherVal.getValueType(), OtherVal,
27794 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
27796 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
27797 DL, OtherVal.getValueType(), OtherVal,
27798 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
27801 /// PerformADDCombine - Do target-specific dag combines on integer adds.
27802 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
27803 const X86Subtarget *Subtarget) {
27804 EVT VT = N->getValueType(0);
27805 SDValue Op0 = N->getOperand(0);
27806 SDValue Op1 = N->getOperand(1);
27808 // Try to synthesize horizontal adds from adds of shuffles.
27809 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27810 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27811 isHorizontalBinOp(Op0, Op1, true))
27812 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
27814 return OptimizeConditionalInDecrement(N, DAG);
27817 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
27818 const X86Subtarget *Subtarget) {
27819 SDValue Op0 = N->getOperand(0);
27820 SDValue Op1 = N->getOperand(1);
27822 // X86 can't encode an immediate LHS of a sub. See if we can push the
27823 // negation into a preceding instruction.
27824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
27825 // If the RHS of the sub is a XOR with one use and a constant, invert the
27826 // immediate. Then add one to the LHS of the sub so we can turn
27827 // X-Y -> X+~Y+1, saving one register.
27828 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
27829 isa<ConstantSDNode>(Op1.getOperand(1))) {
27830 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
27831 EVT VT = Op0.getValueType();
27832 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
27834 DAG.getConstant(~XorC, SDLoc(Op1), VT));
27835 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
27836 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
27840 // Try to synthesize horizontal adds from adds of shuffles.
27841 EVT VT = N->getValueType(0);
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::HSUB, SDLoc(N), VT, Op0, Op1);
27847 return OptimizeConditionalInDecrement(N, DAG);
27850 /// performVZEXTCombine - Performs build vector combines
27851 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
27852 TargetLowering::DAGCombinerInfo &DCI,
27853 const X86Subtarget *Subtarget) {
27855 MVT VT = N->getSimpleValueType(0);
27856 SDValue Op = N->getOperand(0);
27857 MVT OpVT = Op.getSimpleValueType();
27858 MVT OpEltVT = OpVT.getVectorElementType();
27859 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
27861 // (vzext (bitcast (vzext (x)) -> (vzext x)
27863 while (V.getOpcode() == ISD::BITCAST)
27864 V = V.getOperand(0);
27866 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
27867 MVT InnerVT = V.getSimpleValueType();
27868 MVT InnerEltVT = InnerVT.getVectorElementType();
27870 // If the element sizes match exactly, we can just do one larger vzext. This
27871 // is always an exact type match as vzext operates on integer types.
27872 if (OpEltVT == InnerEltVT) {
27873 assert(OpVT == InnerVT && "Types must match for vzext!");
27874 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
27877 // The only other way we can combine them is if only a single element of the
27878 // inner vzext is used in the input to the outer vzext.
27879 if (InnerEltVT.getSizeInBits() < InputBits)
27882 // In this case, the inner vzext is completely dead because we're going to
27883 // only look at bits inside of the low element. Just do the outer vzext on
27884 // a bitcast of the input to the inner.
27885 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
27888 // Check if we can bypass extracting and re-inserting an element of an input
27889 // vector. Essentially:
27890 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
27891 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
27892 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
27893 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
27894 SDValue ExtractedV = V.getOperand(0);
27895 SDValue OrigV = ExtractedV.getOperand(0);
27896 if (isNullConstant(ExtractedV.getOperand(1))) {
27897 MVT OrigVT = OrigV.getSimpleValueType();
27898 // Extract a subvector if necessary...
27899 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
27900 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
27901 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
27902 OrigVT.getVectorNumElements() / Ratio);
27903 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
27904 DAG.getIntPtrConstant(0, DL));
27906 Op = DAG.getBitcast(OpVT, OrigV);
27907 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
27914 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
27915 DAGCombinerInfo &DCI) const {
27916 SelectionDAG &DAG = DCI.DAG;
27917 switch (N->getOpcode()) {
27919 case ISD::EXTRACT_VECTOR_ELT:
27920 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
27923 case X86ISD::SHRUNKBLEND:
27924 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
27925 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
27926 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
27927 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
27928 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
27929 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
27930 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
27933 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
27934 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
27935 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
27936 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
27937 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
27938 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
27939 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
27940 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
27941 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
27942 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
27943 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
27944 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
27945 case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
27946 case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
27948 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
27950 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
27952 case ISD::FMAXNUM: return performFMinNumFMaxNumCombine(N, DAG,
27954 case X86ISD::FAND: return PerformFANDCombine(N, DAG, Subtarget);
27955 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG, Subtarget);
27956 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
27957 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
27958 case ISD::ANY_EXTEND:
27959 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
27960 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
27961 case ISD::SIGN_EXTEND_INREG:
27962 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
27963 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
27964 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
27965 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
27966 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
27967 case X86ISD::SHUFP: // Handle all target specific shuffles
27968 case X86ISD::PALIGNR:
27969 case X86ISD::BLENDI:
27970 case X86ISD::UNPCKH:
27971 case X86ISD::UNPCKL:
27972 case X86ISD::MOVHLPS:
27973 case X86ISD::MOVLHPS:
27974 case X86ISD::PSHUFB:
27975 case X86ISD::PSHUFD:
27976 case X86ISD::PSHUFHW:
27977 case X86ISD::PSHUFLW:
27978 case X86ISD::MOVSS:
27979 case X86ISD::MOVSD:
27980 case X86ISD::VPERMILPI:
27981 case X86ISD::VPERM2X128:
27982 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
27983 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
27985 case ISD::MSCATTER: return PerformGatherScatterCombine(N, DAG);
27991 /// isTypeDesirableForOp - Return true if the target has native support for
27992 /// the specified value type and it is 'desirable' to use the type for the
27993 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
27994 /// instruction encodings are longer and some i16 instructions are slow.
27995 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
27996 if (!isTypeLegal(VT))
27998 if (VT != MVT::i16)
28005 case ISD::SIGN_EXTEND:
28006 case ISD::ZERO_EXTEND:
28007 case ISD::ANY_EXTEND:
28020 /// This function checks if any of the users of EFLAGS copies the EFLAGS. We
28021 /// know that the code that lowers COPY of EFLAGS has to use the stack, and if
28022 /// we don't adjust the stack we clobber the first frame index.
28023 /// See X86InstrInfo::copyPhysReg.
28024 bool X86TargetLowering::hasCopyImplyingStackAdjustment(
28025 MachineFunction *MF) const {
28026 const MachineRegisterInfo &MRI = MF->getRegInfo();
28028 return any_of(MRI.reg_instructions(X86::EFLAGS),
28029 [](const MachineInstr &RI) { return RI.isCopy(); });
28032 /// IsDesirableToPromoteOp - This method query the target whether it is
28033 /// beneficial for dag combiner to promote the specified node. If true, it
28034 /// should return the desired promotion type by reference.
28035 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
28036 EVT VT = Op.getValueType();
28037 if (VT != MVT::i16)
28040 bool Promote = false;
28041 bool Commute = false;
28042 switch (Op.getOpcode()) {
28045 LoadSDNode *LD = cast<LoadSDNode>(Op);
28046 // If the non-extending load has a single use and it's not live out, then it
28047 // might be folded.
28048 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
28049 Op.hasOneUse()*/) {
28050 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
28051 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
28052 // The only case where we'd want to promote LOAD (rather then it being
28053 // promoted as an operand is when it's only use is liveout.
28054 if (UI->getOpcode() != ISD::CopyToReg)
28061 case ISD::SIGN_EXTEND:
28062 case ISD::ZERO_EXTEND:
28063 case ISD::ANY_EXTEND:
28068 SDValue N0 = Op.getOperand(0);
28069 // Look out for (store (shl (load), x)).
28070 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
28083 SDValue N0 = Op.getOperand(0);
28084 SDValue N1 = Op.getOperand(1);
28085 if (!Commute && MayFoldLoad(N1))
28087 // Avoid disabling potential load folding opportunities.
28088 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
28090 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
28100 //===----------------------------------------------------------------------===//
28101 // X86 Inline Assembly Support
28102 //===----------------------------------------------------------------------===//
28104 // Helper to match a string separated by whitespace.
28105 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
28106 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
28108 for (StringRef Piece : Pieces) {
28109 if (!S.startswith(Piece)) // Check if the piece matches.
28112 S = S.substr(Piece.size());
28113 StringRef::size_type Pos = S.find_first_not_of(" \t");
28114 if (Pos == 0) // We matched a prefix.
28123 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
28125 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
28126 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
28127 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
28128 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
28130 if (AsmPieces.size() == 3)
28132 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
28139 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
28140 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
28142 std::string AsmStr = IA->getAsmString();
28144 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
28145 if (!Ty || Ty->getBitWidth() % 16 != 0)
28148 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
28149 SmallVector<StringRef, 4> AsmPieces;
28150 SplitString(AsmStr, AsmPieces, ";\n");
28152 switch (AsmPieces.size()) {
28153 default: return false;
28155 // FIXME: this should verify that we are targeting a 486 or better. If not,
28156 // we will turn this bswap into something that will be lowered to logical
28157 // ops instead of emitting the bswap asm. For now, we don't support 486 or
28158 // lower so don't worry about this.
28160 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
28161 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
28162 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
28163 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
28164 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
28165 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
28166 // No need to check constraints, nothing other than the equivalent of
28167 // "=r,0" would be valid here.
28168 return IntrinsicLowering::LowerToByteSwap(CI);
28171 // rorw $$8, ${0:w} --> llvm.bswap.i16
28172 if (CI->getType()->isIntegerTy(16) &&
28173 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28174 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
28175 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
28177 StringRef ConstraintsStr = IA->getConstraintString();
28178 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28179 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28180 if (clobbersFlagRegisters(AsmPieces))
28181 return IntrinsicLowering::LowerToByteSwap(CI);
28185 if (CI->getType()->isIntegerTy(32) &&
28186 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28187 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
28188 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
28189 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
28191 StringRef ConstraintsStr = IA->getConstraintString();
28192 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28193 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28194 if (clobbersFlagRegisters(AsmPieces))
28195 return IntrinsicLowering::LowerToByteSwap(CI);
28198 if (CI->getType()->isIntegerTy(64)) {
28199 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
28200 if (Constraints.size() >= 2 &&
28201 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
28202 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
28203 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
28204 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
28205 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
28206 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
28207 return IntrinsicLowering::LowerToByteSwap(CI);
28215 /// getConstraintType - Given a constraint letter, return the type of
28216 /// constraint it is for this target.
28217 X86TargetLowering::ConstraintType
28218 X86TargetLowering::getConstraintType(StringRef Constraint) const {
28219 if (Constraint.size() == 1) {
28220 switch (Constraint[0]) {
28231 return C_RegisterClass;
28255 return TargetLowering::getConstraintType(Constraint);
28258 /// Examine constraint type and operand type and determine a weight value.
28259 /// This object must already have been set up with the operand type
28260 /// and the current alternative constraint selected.
28261 TargetLowering::ConstraintWeight
28262 X86TargetLowering::getSingleConstraintMatchWeight(
28263 AsmOperandInfo &info, const char *constraint) const {
28264 ConstraintWeight weight = CW_Invalid;
28265 Value *CallOperandVal = info.CallOperandVal;
28266 // If we don't have a value, we can't do a match,
28267 // but allow it at the lowest weight.
28268 if (!CallOperandVal)
28270 Type *type = CallOperandVal->getType();
28271 // Look at the constraint type.
28272 switch (*constraint) {
28274 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
28285 if (CallOperandVal->getType()->isIntegerTy())
28286 weight = CW_SpecificReg;
28291 if (type->isFloatingPointTy())
28292 weight = CW_SpecificReg;
28295 if (type->isX86_MMXTy() && Subtarget->hasMMX())
28296 weight = CW_SpecificReg;
28300 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
28301 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
28302 weight = CW_Register;
28305 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
28306 if (C->getZExtValue() <= 31)
28307 weight = CW_Constant;
28311 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28312 if (C->getZExtValue() <= 63)
28313 weight = CW_Constant;
28317 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28318 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
28319 weight = CW_Constant;
28323 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28324 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
28325 weight = CW_Constant;
28329 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28330 if (C->getZExtValue() <= 3)
28331 weight = CW_Constant;
28335 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28336 if (C->getZExtValue() <= 0xff)
28337 weight = CW_Constant;
28342 if (isa<ConstantFP>(CallOperandVal)) {
28343 weight = CW_Constant;
28347 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28348 if ((C->getSExtValue() >= -0x80000000LL) &&
28349 (C->getSExtValue() <= 0x7fffffffLL))
28350 weight = CW_Constant;
28354 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28355 if (C->getZExtValue() <= 0xffffffff)
28356 weight = CW_Constant;
28363 /// LowerXConstraint - try to replace an X constraint, which matches anything,
28364 /// with another that has more specific requirements based on the type of the
28365 /// corresponding operand.
28366 const char *X86TargetLowering::
28367 LowerXConstraint(EVT ConstraintVT) const {
28368 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
28369 // 'f' like normal targets.
28370 if (ConstraintVT.isFloatingPoint()) {
28371 if (Subtarget->hasSSE2())
28373 if (Subtarget->hasSSE1())
28377 return TargetLowering::LowerXConstraint(ConstraintVT);
28380 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
28381 /// vector. If it is invalid, don't add anything to Ops.
28382 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
28383 std::string &Constraint,
28384 std::vector<SDValue>&Ops,
28385 SelectionDAG &DAG) const {
28388 // Only support length 1 constraints for now.
28389 if (Constraint.length() > 1) return;
28391 char ConstraintLetter = Constraint[0];
28392 switch (ConstraintLetter) {
28395 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28396 if (C->getZExtValue() <= 31) {
28397 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28398 Op.getValueType());
28404 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28405 if (C->getZExtValue() <= 63) {
28406 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28407 Op.getValueType());
28413 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28414 if (isInt<8>(C->getSExtValue())) {
28415 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28416 Op.getValueType());
28422 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28423 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
28424 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
28425 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
28426 Op.getValueType());
28432 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28433 if (C->getZExtValue() <= 3) {
28434 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28435 Op.getValueType());
28441 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28442 if (C->getZExtValue() <= 255) {
28443 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28444 Op.getValueType());
28450 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28451 if (C->getZExtValue() <= 127) {
28452 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28453 Op.getValueType());
28459 // 32-bit signed value
28460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28461 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28462 C->getSExtValue())) {
28463 // Widen to 64 bits here to get it sign extended.
28464 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
28467 // FIXME gcc accepts some relocatable values here too, but only in certain
28468 // memory models; it's complicated.
28473 // 32-bit unsigned value
28474 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28475 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28476 C->getZExtValue())) {
28477 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28478 Op.getValueType());
28482 // FIXME gcc accepts some relocatable values here too, but only in certain
28483 // memory models; it's complicated.
28487 // Literal immediates are always ok.
28488 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
28489 // Widen to 64 bits here to get it sign extended.
28490 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
28494 // In any sort of PIC mode addresses need to be computed at runtime by
28495 // adding in a register or some sort of table lookup. These can't
28496 // be used as immediates.
28497 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
28500 // If we are in non-pic codegen mode, we allow the address of a global (with
28501 // an optional displacement) to be used with 'i'.
28502 GlobalAddressSDNode *GA = nullptr;
28503 int64_t Offset = 0;
28505 // Match either (GA), (GA+C), (GA+C1+C2), etc.
28507 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
28508 Offset += GA->getOffset();
28510 } else if (Op.getOpcode() == ISD::ADD) {
28511 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28512 Offset += C->getZExtValue();
28513 Op = Op.getOperand(0);
28516 } else if (Op.getOpcode() == ISD::SUB) {
28517 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28518 Offset += -C->getZExtValue();
28519 Op = Op.getOperand(0);
28524 // Otherwise, this isn't something we can handle, reject it.
28528 const GlobalValue *GV = GA->getGlobal();
28529 // If we require an extra load to get this address, as in PIC mode, we
28530 // can't accept it.
28531 if (isGlobalStubReference(
28532 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
28535 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
28536 GA->getValueType(0), Offset);
28541 if (Result.getNode()) {
28542 Ops.push_back(Result);
28545 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
28548 std::pair<unsigned, const TargetRegisterClass *>
28549 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
28550 StringRef Constraint,
28552 // First, see if this is a constraint that directly corresponds to an LLVM
28554 if (Constraint.size() == 1) {
28555 // GCC Constraint Letters
28556 switch (Constraint[0]) {
28558 // TODO: Slight differences here in allocation order and leaving
28559 // RIP in the class. Do they matter any more here than they do
28560 // in the normal allocation?
28561 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
28562 if (Subtarget->is64Bit()) {
28563 if (VT == MVT::i32 || VT == MVT::f32)
28564 return std::make_pair(0U, &X86::GR32RegClass);
28565 if (VT == MVT::i16)
28566 return std::make_pair(0U, &X86::GR16RegClass);
28567 if (VT == MVT::i8 || VT == MVT::i1)
28568 return std::make_pair(0U, &X86::GR8RegClass);
28569 if (VT == MVT::i64 || VT == MVT::f64)
28570 return std::make_pair(0U, &X86::GR64RegClass);
28573 // 32-bit fallthrough
28574 case 'Q': // Q_REGS
28575 if (VT == MVT::i32 || VT == MVT::f32)
28576 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
28577 if (VT == MVT::i16)
28578 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
28579 if (VT == MVT::i8 || VT == MVT::i1)
28580 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
28581 if (VT == MVT::i64)
28582 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
28584 case 'r': // GENERAL_REGS
28585 case 'l': // INDEX_REGS
28586 if (VT == MVT::i8 || VT == MVT::i1)
28587 return std::make_pair(0U, &X86::GR8RegClass);
28588 if (VT == MVT::i16)
28589 return std::make_pair(0U, &X86::GR16RegClass);
28590 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
28591 return std::make_pair(0U, &X86::GR32RegClass);
28592 return std::make_pair(0U, &X86::GR64RegClass);
28593 case 'R': // LEGACY_REGS
28594 if (VT == MVT::i8 || VT == MVT::i1)
28595 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
28596 if (VT == MVT::i16)
28597 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
28598 if (VT == MVT::i32 || !Subtarget->is64Bit())
28599 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
28600 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
28601 case 'f': // FP Stack registers.
28602 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
28603 // value to the correct fpstack register class.
28604 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
28605 return std::make_pair(0U, &X86::RFP32RegClass);
28606 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
28607 return std::make_pair(0U, &X86::RFP64RegClass);
28608 return std::make_pair(0U, &X86::RFP80RegClass);
28609 case 'y': // MMX_REGS if MMX allowed.
28610 if (!Subtarget->hasMMX()) break;
28611 return std::make_pair(0U, &X86::VR64RegClass);
28612 case 'Y': // SSE_REGS if SSE2 allowed
28613 if (!Subtarget->hasSSE2()) break;
28615 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
28616 if (!Subtarget->hasSSE1()) break;
28618 switch (VT.SimpleTy) {
28620 // Scalar SSE types.
28623 return std::make_pair(0U, &X86::FR32RegClass);
28626 return std::make_pair(0U, &X86::FR64RegClass);
28627 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28635 return std::make_pair(0U, &X86::VR128RegClass);
28643 return std::make_pair(0U, &X86::VR256RegClass);
28648 return std::make_pair(0U, &X86::VR512RegClass);
28654 // Use the default implementation in TargetLowering to convert the register
28655 // constraint into a member of a register class.
28656 std::pair<unsigned, const TargetRegisterClass*> Res;
28657 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
28659 // Not found as a standard register?
28661 // Map st(0) -> st(7) -> ST0
28662 if (Constraint.size() == 7 && Constraint[0] == '{' &&
28663 tolower(Constraint[1]) == 's' &&
28664 tolower(Constraint[2]) == 't' &&
28665 Constraint[3] == '(' &&
28666 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
28667 Constraint[5] == ')' &&
28668 Constraint[6] == '}') {
28670 Res.first = X86::FP0+Constraint[4]-'0';
28671 Res.second = &X86::RFP80RegClass;
28675 // GCC allows "st(0)" to be called just plain "st".
28676 if (StringRef("{st}").equals_lower(Constraint)) {
28677 Res.first = X86::FP0;
28678 Res.second = &X86::RFP80RegClass;
28683 if (StringRef("{flags}").equals_lower(Constraint)) {
28684 Res.first = X86::EFLAGS;
28685 Res.second = &X86::CCRRegClass;
28689 // 'A' means EAX + EDX.
28690 if (Constraint == "A") {
28691 Res.first = X86::EAX;
28692 Res.second = &X86::GR32_ADRegClass;
28698 // Otherwise, check to see if this is a register class of the wrong value
28699 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
28700 // turn into {ax},{dx}.
28701 // MVT::Other is used to specify clobber names.
28702 if (Res.second->hasType(VT) || VT == MVT::Other)
28703 return Res; // Correct type already, nothing to do.
28705 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
28706 // return "eax". This should even work for things like getting 64bit integer
28707 // registers when given an f64 type.
28708 const TargetRegisterClass *Class = Res.second;
28709 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
28710 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
28711 unsigned Size = VT.getSizeInBits();
28712 if (Size == 1) Size = 8;
28713 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, Size);
28715 Res.first = DestReg;
28716 Res.second = Size == 8 ? &X86::GR8RegClass
28717 : Size == 16 ? &X86::GR16RegClass
28718 : Size == 32 ? &X86::GR32RegClass
28719 : &X86::GR64RegClass;
28720 assert(Res.second->contains(Res.first) && "Register in register class");
28722 // No register found/type mismatch.
28724 Res.second = nullptr;
28726 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
28727 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
28728 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
28729 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
28730 Class == &X86::VR512RegClass) {
28731 // Handle references to XMM physical registers that got mapped into the
28732 // wrong class. This can happen with constraints like {xmm0} where the
28733 // target independent register mapper will just pick the first match it can
28734 // find, ignoring the required type.
28736 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28737 if (VT == MVT::f32 || VT == MVT::i32)
28738 Res.second = &X86::FR32RegClass;
28739 else if (VT == MVT::f64 || VT == MVT::i64)
28740 Res.second = &X86::FR64RegClass;
28741 else if (X86::VR128RegClass.hasType(VT))
28742 Res.second = &X86::VR128RegClass;
28743 else if (X86::VR256RegClass.hasType(VT))
28744 Res.second = &X86::VR256RegClass;
28745 else if (X86::VR512RegClass.hasType(VT))
28746 Res.second = &X86::VR512RegClass;
28748 // Type mismatch and not a clobber: Return an error;
28750 Res.second = nullptr;
28757 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
28758 const AddrMode &AM, Type *Ty,
28759 unsigned AS) const {
28760 // Scaling factors are not free at all.
28761 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
28762 // will take 2 allocations in the out of order engine instead of 1
28763 // for plain addressing mode, i.e. inst (reg1).
28765 // vaddps (%rsi,%drx), %ymm0, %ymm1
28766 // Requires two allocations (one for the load, one for the computation)
28768 // vaddps (%rsi), %ymm0, %ymm1
28769 // Requires just 1 allocation, i.e., freeing allocations for other operations
28770 // and having less micro operations to execute.
28772 // For some X86 architectures, this is even worse because for instance for
28773 // stores, the complex addressing mode forces the instruction to use the
28774 // "load" ports instead of the dedicated "store" port.
28775 // E.g., on Haswell:
28776 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
28777 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
28778 if (isLegalAddressingMode(DL, AM, Ty, AS))
28779 // Scale represents reg2 * scale, thus account for 1
28780 // as soon as we use a second register.
28781 return AM.Scale != 0;
28785 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
28786 // Integer division on x86 is expensive. However, when aggressively optimizing
28787 // for code size, we prefer to use a div instruction, as it is usually smaller
28788 // than the alternative sequence.
28789 // The exception to this is vector division. Since x86 doesn't have vector
28790 // integer division, leaving the division as-is is a loss even in terms of
28791 // size, because it will have to be scalarized, while the alternative code
28792 // sequence can be performed in vector form.
28793 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
28794 Attribute::MinSize);
28795 return OptSize && !VT.isVector();