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 "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/Analysis/EHPersonalities.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/WinEHFuncInfo.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalAlias.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/MC/MCAsmInfo.h"
47 #include "llvm/MC/MCContext.h"
48 #include "llvm/MC/MCExpr.h"
49 #include "llvm/MC/MCSymbol.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetOptions.h"
55 #include "X86IntrinsicsInfo.h"
61 #define DEBUG_TYPE "x86-isel"
63 STATISTIC(NumTailCalls, "Number of tail calls");
65 static cl::opt<bool> ExperimentalVectorWideningLegalization(
66 "x86-experimental-vector-widening-legalization", cl::init(false),
67 cl::desc("Enable an experimental vector type legalization through widening "
68 "rather than promotion."),
71 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
72 const X86Subtarget &STI)
73 : TargetLowering(TM), Subtarget(&STI) {
74 X86ScalarSSEf64 = Subtarget->hasSSE2();
75 X86ScalarSSEf32 = Subtarget->hasSSE1();
76 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
78 // Set up the TargetLowering object.
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
119 if (Subtarget->isTargetDarwin()) {
120 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(false);
122 setUseUnderscoreLongJmp(false);
123 } else if (Subtarget->isTargetWindowsGNU()) {
124 // MS runtime is weird: it exports _setjmp, but longjmp!
125 setUseUnderscoreSetJmp(true);
126 setUseUnderscoreLongJmp(false);
128 setUseUnderscoreSetJmp(true);
129 setUseUnderscoreLongJmp(true);
132 // Set up the register classes.
133 addRegisterClass(MVT::i8, &X86::GR8RegClass);
134 addRegisterClass(MVT::i16, &X86::GR16RegClass);
135 addRegisterClass(MVT::i32, &X86::GR32RegClass);
136 if (Subtarget->is64Bit())
137 addRegisterClass(MVT::i64, &X86::GR64RegClass);
139 for (MVT VT : MVT::integer_valuetypes())
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
142 // We don't accept any truncstore of integer registers.
143 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
146 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
147 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
148 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
150 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
152 // SETOEQ and SETUNE require checking two conditions.
153 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
160 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
162 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
166 if (Subtarget->is64Bit()) {
167 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
168 // f32/f64 are legal, f80 is custom.
169 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 } else if (!Subtarget->useSoftFloat()) {
174 // We have an algorithm for SSE2->double, and we turn this into a
175 // 64-bit FILD followed by conditional FADD for other targets.
176 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
177 // We have an algorithm for SSE2, and we turn this into a 64-bit
178 // FILD or VCVTUSI2SS/SD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
182 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
184 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
187 if (!Subtarget->useSoftFloat()) {
188 // SSE has no i16 to fp conversion, only i32
189 if (X86ScalarSSEf32) {
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
191 // f32 and f64 cases are Legal, f80 case is not
192 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
202 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
204 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
205 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
207 if (!Subtarget->useSoftFloat()) {
208 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
209 // are Legal, f80 is custom lowered.
210 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
211 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
213 if (X86ScalarSSEf32) {
214 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
215 // f32 and f64 cases are Legal, f80 case is not
216 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
218 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
219 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
222 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand);
224 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand);
227 // Handle FP_TO_UINT by promoting the destination to a larger signed
229 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
233 if (Subtarget->is64Bit()) {
234 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
235 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
236 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
237 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
239 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
240 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
242 } else if (!Subtarget->useSoftFloat()) {
243 // Since AVX is a superset of SSE3, only check for SSE here.
244 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
245 // Expand FP_TO_UINT into a select.
246 // FIXME: We would like to use a Custom expander here eventually to do
247 // the optimal thing for SSE vs. the default expansion in the legalizer.
248 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
250 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
251 // With SSE3 we can use fisttpll to convert to a signed i64; without
252 // SSE, we're stuck with a fistpll.
253 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
255 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
258 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
259 if (!X86ScalarSSEf64) {
260 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
261 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
262 if (Subtarget->is64Bit()) {
263 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
264 // Without SSE, i64->f64 goes through memory.
265 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
269 // Scalar integer divide and remainder are lowered to use operations that
270 // produce two results, to match the available instructions. This exposes
271 // the two-result form to trivial CSE, which is able to combine x/y and x%y
272 // into a single instruction.
274 // Scalar integer multiply-high is also lowered to use two-result
275 // operations, to match the available instructions. However, plain multiply
276 // (low) operations are left as Legal, as there are single-result
277 // instructions for this in x86. Using the two-result multiply instructions
278 // when both high and low results are needed must be arranged by dagcombine.
279 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
280 setOperationAction(ISD::MULHS, VT, Expand);
281 setOperationAction(ISD::MULHU, VT, Expand);
282 setOperationAction(ISD::SDIV, VT, Expand);
283 setOperationAction(ISD::UDIV, VT, Expand);
284 setOperationAction(ISD::SREM, VT, Expand);
285 setOperationAction(ISD::UREM, VT, Expand);
287 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
288 setOperationAction(ISD::ADDC, VT, Custom);
289 setOperationAction(ISD::ADDE, VT, Custom);
290 setOperationAction(ISD::SUBC, VT, Custom);
291 setOperationAction(ISD::SUBE, VT, Custom);
294 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
295 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
296 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
298 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
299 setOperationAction(ISD::BR_CC , MVT::f128, Expand);
300 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
303 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::f128, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
312 if (Subtarget->is64Bit())
313 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
317 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
319 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
320 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
321 // is. We should promote the value to 64-bits to solve this.
322 // This is what the CRT headers do - `fmodf` is an inline header
323 // function casting to f64 and calling `fmod`.
324 setOperationAction(ISD::FREM , MVT::f32 , Promote);
326 setOperationAction(ISD::FREM , MVT::f32 , Expand);
329 setOperationAction(ISD::FREM , MVT::f64 , Expand);
330 setOperationAction(ISD::FREM , MVT::f80 , Expand);
331 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
333 // Promote the i8 variants and force them on up to i32 which has a shorter
335 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
336 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
337 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
338 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
339 if (Subtarget->hasBMI()) {
340 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
342 if (Subtarget->is64Bit())
343 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
345 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
346 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
347 if (Subtarget->is64Bit())
348 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
351 if (Subtarget->hasLZCNT()) {
352 // When promoting the i8 variants, force them to i32 for a shorter
354 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
355 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
357 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
358 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
360 if (Subtarget->is64Bit())
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
363 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
364 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
365 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
366 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
367 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
368 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
369 if (Subtarget->is64Bit()) {
370 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
371 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
375 // Special handling for half-precision floating point conversions.
376 // If we don't have F16C support, then lower half float conversions
377 // into library calls.
378 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
379 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
380 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
383 // There's never any support for operations beyond MVT::f32.
384 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
385 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
386 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
387 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
389 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
390 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
391 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
392 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
393 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
394 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
396 if (Subtarget->hasPOPCNT()) {
397 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
399 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
400 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
401 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
402 if (Subtarget->is64Bit())
403 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
406 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
408 if (!Subtarget->hasMOVBE())
409 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
411 // These should be promoted to a larger select which is supported.
412 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
413 // X86 wants to expand cmov itself.
414 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
415 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
416 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
417 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
418 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
419 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
420 setOperationAction(ISD::SELECT , MVT::f128 , Custom);
421 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
422 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
423 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
424 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
425 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
426 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
427 setOperationAction(ISD::SETCC , MVT::f128 , Custom);
428 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
429 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
431 if (Subtarget->is64Bit()) {
432 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
433 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
434 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
436 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
437 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
438 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
439 // support continuation, user-level threading, and etc.. As a result, no
440 // other SjLj exception interfaces are implemented and please don't build
441 // your own exception handling based on them.
442 // LLVM/Clang supports zero-cost DWARF exception handling.
443 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
444 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
447 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
448 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
449 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
450 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
451 if (Subtarget->is64Bit())
452 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
453 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
454 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
455 if (Subtarget->is64Bit()) {
456 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
457 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
458 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
459 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
460 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
462 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
463 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
464 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
465 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
466 if (Subtarget->is64Bit()) {
467 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
468 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
469 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
472 if (Subtarget->hasSSE1())
473 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
475 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
477 // Expand certain atomics
478 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
479 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
480 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
481 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
484 if (Subtarget->hasCmpxchg16b()) {
485 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
488 // FIXME - use subtarget debug flags
489 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
490 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
491 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
494 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
495 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
497 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
498 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
500 setOperationAction(ISD::TRAP, MVT::Other, Legal);
501 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
503 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
504 setOperationAction(ISD::VASTART , MVT::Other, Custom);
505 setOperationAction(ISD::VAEND , MVT::Other, Expand);
506 if (Subtarget->is64Bit()) {
507 setOperationAction(ISD::VAARG , MVT::Other, Custom);
508 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
510 // TargetInfo::CharPtrBuiltinVaList
511 setOperationAction(ISD::VAARG , MVT::Other, Expand);
512 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
515 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
516 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
518 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
520 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
521 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
522 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
524 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
525 // f32 and f64 use SSE.
526 // Set up the FP register classes.
527 addRegisterClass(MVT::f32, &X86::FR32RegClass);
528 addRegisterClass(MVT::f64, &X86::FR64RegClass);
530 // Use ANDPD to simulate FABS.
531 setOperationAction(ISD::FABS , MVT::f64, Custom);
532 setOperationAction(ISD::FABS , MVT::f32, Custom);
534 // Use XORP to simulate FNEG.
535 setOperationAction(ISD::FNEG , MVT::f64, Custom);
536 setOperationAction(ISD::FNEG , MVT::f32, Custom);
538 // Use ANDPD and ORPD to simulate FCOPYSIGN.
539 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
540 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
542 // Lower this to FGETSIGNx86 plus an AND.
543 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
544 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
546 // We don't support sin/cos/fmod
547 setOperationAction(ISD::FSIN , MVT::f64, Expand);
548 setOperationAction(ISD::FCOS , MVT::f64, Expand);
549 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
550 setOperationAction(ISD::FSIN , MVT::f32, Expand);
551 setOperationAction(ISD::FCOS , MVT::f32, Expand);
552 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
554 // Expand FP immediates into loads from the stack, except for the special
556 addLegalFPImmediate(APFloat(+0.0)); // xorpd
557 addLegalFPImmediate(APFloat(+0.0f)); // xorps
558 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
559 // Use SSE for f32, x87 for f64.
560 // Set up the FP register classes.
561 addRegisterClass(MVT::f32, &X86::FR32RegClass);
562 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
564 // Use ANDPS to simulate FABS.
565 setOperationAction(ISD::FABS , MVT::f32, Custom);
567 // Use XORP to simulate FNEG.
568 setOperationAction(ISD::FNEG , MVT::f32, Custom);
570 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
572 // Use ANDPS and ORPS to simulate FCOPYSIGN.
573 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
574 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
576 // We don't support sin/cos/fmod
577 setOperationAction(ISD::FSIN , MVT::f32, Expand);
578 setOperationAction(ISD::FCOS , MVT::f32, Expand);
579 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
581 // Special cases we handle for FP constants.
582 addLegalFPImmediate(APFloat(+0.0f)); // xorps
583 addLegalFPImmediate(APFloat(+0.0)); // FLD0
584 addLegalFPImmediate(APFloat(+1.0)); // FLD1
585 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
586 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
588 if (!TM.Options.UnsafeFPMath) {
589 setOperationAction(ISD::FSIN , MVT::f64, Expand);
590 setOperationAction(ISD::FCOS , MVT::f64, Expand);
591 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
593 } else if (!Subtarget->useSoftFloat()) {
594 // f32 and f64 in x87.
595 // Set up the FP register classes.
596 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
597 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
599 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
600 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
601 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
604 if (!TM.Options.UnsafeFPMath) {
605 setOperationAction(ISD::FSIN , MVT::f64, Expand);
606 setOperationAction(ISD::FSIN , MVT::f32, Expand);
607 setOperationAction(ISD::FCOS , MVT::f64, Expand);
608 setOperationAction(ISD::FCOS , MVT::f32, Expand);
609 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
610 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
612 addLegalFPImmediate(APFloat(+0.0)); // FLD0
613 addLegalFPImmediate(APFloat(+1.0)); // FLD1
614 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
615 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
616 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
617 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
618 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
619 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
622 // We don't support FMA.
623 setOperationAction(ISD::FMA, MVT::f64, Expand);
624 setOperationAction(ISD::FMA, MVT::f32, Expand);
626 // Long double always uses X87, except f128 in MMX.
627 if (!Subtarget->useSoftFloat()) {
628 if (Subtarget->is64Bit() && Subtarget->hasMMX()) {
629 addRegisterClass(MVT::f128, &X86::FR128RegClass);
630 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
631 setOperationAction(ISD::FABS , MVT::f128, Custom);
632 setOperationAction(ISD::FNEG , MVT::f128, Custom);
633 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
636 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
637 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
638 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
640 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
641 addLegalFPImmediate(TmpFlt); // FLD0
643 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
646 APFloat TmpFlt2(+1.0);
647 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
649 addLegalFPImmediate(TmpFlt2); // FLD1
650 TmpFlt2.changeSign();
651 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
654 if (!TM.Options.UnsafeFPMath) {
655 setOperationAction(ISD::FSIN , MVT::f80, Expand);
656 setOperationAction(ISD::FCOS , MVT::f80, Expand);
657 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
660 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
661 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
662 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
663 setOperationAction(ISD::FRINT, MVT::f80, Expand);
664 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
665 setOperationAction(ISD::FMA, MVT::f80, Expand);
668 // Always use a library call for pow.
669 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
670 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
671 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
673 setOperationAction(ISD::FLOG, MVT::f80, Expand);
674 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
675 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
676 setOperationAction(ISD::FEXP, MVT::f80, Expand);
677 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
678 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
679 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
681 // First set operation action for all vector types to either promote
682 // (for widening) or expand (for scalarization). Then we will selectively
683 // turn on ones that can be effectively codegen'd.
684 for (MVT VT : MVT::vector_valuetypes()) {
685 setOperationAction(ISD::ADD , VT, Expand);
686 setOperationAction(ISD::SUB , VT, Expand);
687 setOperationAction(ISD::FADD, VT, Expand);
688 setOperationAction(ISD::FNEG, VT, Expand);
689 setOperationAction(ISD::FSUB, VT, Expand);
690 setOperationAction(ISD::MUL , VT, Expand);
691 setOperationAction(ISD::FMUL, VT, Expand);
692 setOperationAction(ISD::SDIV, VT, Expand);
693 setOperationAction(ISD::UDIV, VT, Expand);
694 setOperationAction(ISD::FDIV, VT, Expand);
695 setOperationAction(ISD::SREM, VT, Expand);
696 setOperationAction(ISD::UREM, VT, Expand);
697 setOperationAction(ISD::LOAD, VT, Expand);
698 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
699 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
700 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
701 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
702 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
703 setOperationAction(ISD::FABS, VT, Expand);
704 setOperationAction(ISD::FSIN, VT, Expand);
705 setOperationAction(ISD::FSINCOS, VT, Expand);
706 setOperationAction(ISD::FCOS, VT, Expand);
707 setOperationAction(ISD::FSINCOS, VT, Expand);
708 setOperationAction(ISD::FREM, VT, Expand);
709 setOperationAction(ISD::FMA, VT, Expand);
710 setOperationAction(ISD::FPOWI, VT, Expand);
711 setOperationAction(ISD::FSQRT, VT, Expand);
712 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
713 setOperationAction(ISD::FFLOOR, VT, Expand);
714 setOperationAction(ISD::FCEIL, VT, Expand);
715 setOperationAction(ISD::FTRUNC, VT, Expand);
716 setOperationAction(ISD::FRINT, VT, Expand);
717 setOperationAction(ISD::FNEARBYINT, VT, Expand);
718 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
719 setOperationAction(ISD::MULHS, VT, Expand);
720 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
721 setOperationAction(ISD::MULHU, VT, Expand);
722 setOperationAction(ISD::SDIVREM, VT, Expand);
723 setOperationAction(ISD::UDIVREM, VT, Expand);
724 setOperationAction(ISD::FPOW, VT, Expand);
725 setOperationAction(ISD::CTPOP, VT, Expand);
726 setOperationAction(ISD::CTTZ, VT, Expand);
727 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
728 setOperationAction(ISD::CTLZ, VT, Expand);
729 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
730 setOperationAction(ISD::SHL, VT, Expand);
731 setOperationAction(ISD::SRA, VT, Expand);
732 setOperationAction(ISD::SRL, VT, Expand);
733 setOperationAction(ISD::ROTL, VT, Expand);
734 setOperationAction(ISD::ROTR, VT, Expand);
735 setOperationAction(ISD::BSWAP, VT, Expand);
736 setOperationAction(ISD::SETCC, VT, Expand);
737 setOperationAction(ISD::FLOG, VT, Expand);
738 setOperationAction(ISD::FLOG2, VT, Expand);
739 setOperationAction(ISD::FLOG10, VT, Expand);
740 setOperationAction(ISD::FEXP, VT, Expand);
741 setOperationAction(ISD::FEXP2, VT, Expand);
742 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
743 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
744 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
745 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
746 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
747 setOperationAction(ISD::TRUNCATE, VT, Expand);
748 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
749 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
750 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
751 setOperationAction(ISD::VSELECT, VT, Expand);
752 setOperationAction(ISD::SELECT_CC, VT, Expand);
753 for (MVT InnerVT : MVT::vector_valuetypes()) {
754 setTruncStoreAction(InnerVT, VT, Expand);
756 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
757 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
759 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
760 // types, we have to deal with them whether we ask for Expansion or not.
761 // Setting Expand causes its own optimisation problems though, so leave
763 if (VT.getVectorElementType() == MVT::i1)
764 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
766 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
767 // split/scalarized right now.
768 if (VT.getVectorElementType() == MVT::f16)
769 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
773 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
774 // with -msoft-float, disable use of MMX as well.
775 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
776 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
777 // No operations on x86mmx supported, everything uses intrinsics.
780 // MMX-sized vectors (other than x86mmx) are expected to be expanded
781 // into smaller operations.
782 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
783 setOperationAction(ISD::MULHS, MMXTy, Expand);
784 setOperationAction(ISD::AND, MMXTy, Expand);
785 setOperationAction(ISD::OR, MMXTy, Expand);
786 setOperationAction(ISD::XOR, MMXTy, Expand);
787 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
788 setOperationAction(ISD::SELECT, MMXTy, Expand);
789 setOperationAction(ISD::BITCAST, MMXTy, Expand);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
793 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
794 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
796 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
797 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
798 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
799 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
800 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
801 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
802 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
803 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
804 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
805 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
806 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
807 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
808 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
809 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
812 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
813 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
815 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
816 // registers cannot be used even for integer operations.
817 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
818 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
819 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
820 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
822 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
823 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
824 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
825 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
826 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
827 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
828 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
829 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
830 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
831 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
832 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
833 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
834 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
835 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
836 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
837 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
838 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
839 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
840 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
841 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
842 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
843 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
844 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
846 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
847 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
848 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
849 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
851 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
852 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
853 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
854 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
856 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
857 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
859 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
862 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
863 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
864 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
865 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
867 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
868 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
869 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
870 // ISD::CTTZ v2i64 - scalarization is faster.
871 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
872 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
873 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
874 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
876 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
877 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
878 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
880 setOperationAction(ISD::VSELECT, VT, Custom);
881 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
884 // We support custom legalizing of sext and anyext loads for specific
885 // memory vector types which we can load as a scalar (or sequence of
886 // scalars) and extend in-register to a legal 128-bit vector type. For sext
887 // loads these must work with a single scalar load.
888 for (MVT VT : MVT::integer_vector_valuetypes()) {
889 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
890 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
891 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
892 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
893 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
894 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
895 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
896 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
897 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
900 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
901 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
902 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
903 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
904 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
905 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
907 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
909 if (Subtarget->is64Bit()) {
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
911 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
914 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
915 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
916 setOperationAction(ISD::AND, VT, Promote);
917 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
918 setOperationAction(ISD::OR, VT, Promote);
919 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
920 setOperationAction(ISD::XOR, VT, Promote);
921 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
922 setOperationAction(ISD::LOAD, VT, Promote);
923 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
924 setOperationAction(ISD::SELECT, VT, Promote);
925 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
928 // Custom lower v2i64 and v2f64 selects.
929 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
930 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
931 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
932 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
934 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
935 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
937 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
939 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
940 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
941 // As there is no 64-bit GPR available, we need build a special custom
942 // sequence to convert from v2i32 to v2f32.
943 if (!Subtarget->is64Bit())
944 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
946 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
947 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
949 for (MVT VT : MVT::fp_vector_valuetypes())
950 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
952 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
953 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
954 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
957 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
958 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
959 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
960 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
961 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
962 setOperationAction(ISD::FRINT, RoundedTy, Legal);
963 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
966 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
967 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
968 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
969 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
970 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
971 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
972 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
973 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
975 // FIXME: Do we need to handle scalar-to-vector here?
976 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
978 // We directly match byte blends in the backend as they match the VSELECT
980 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
982 // SSE41 brings specific instructions for doing vector sign extend even in
983 // cases where we don't have SRA.
984 for (MVT VT : MVT::integer_vector_valuetypes()) {
985 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
986 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
990 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
991 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
998 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1005 // i8 and i16 vectors are custom because the source register and source
1006 // source memory operand types are not the same width. f32 vectors are
1007 // custom since the immediate controlling the insert encodes additional
1009 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1014 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1015 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1019 // FIXME: these should be Legal, but that's only for the case where
1020 // the index is constant. For now custom expand to deal with that.
1021 if (Subtarget->is64Bit()) {
1022 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1023 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1027 if (Subtarget->hasSSE2()) {
1028 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1032 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1035 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1036 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1038 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1039 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1041 // In the customized shift lowering, the legal cases in AVX2 will be
1043 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1044 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1046 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1047 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1049 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1050 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1053 if (Subtarget->hasXOP()) {
1054 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1055 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1064 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1065 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1066 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1072 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1073 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1074 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1076 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1077 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1087 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1089 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1090 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1100 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1102 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1103 // even though v8i16 is a legal type.
1104 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1105 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1106 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1108 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1109 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1110 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1112 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1113 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1115 for (MVT VT : MVT::fp_vector_valuetypes())
1116 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1118 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1119 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1121 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1122 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1124 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1125 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1127 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1128 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1132 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1133 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1136 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1137 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1139 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1140 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1142 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1143 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1145 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1146 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1149 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1150 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1154 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1155 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1158 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1163 if (Subtarget->hasAnyFMA()) {
1164 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1165 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1166 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1167 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1168 setOperationAction(ISD::FMA, MVT::f32, Legal);
1169 setOperationAction(ISD::FMA, MVT::f64, Legal);
1172 if (Subtarget->hasInt256()) {
1173 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1174 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1175 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1176 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1178 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1179 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1180 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1181 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1183 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1184 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1185 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1186 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1188 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1189 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1190 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1191 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1193 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1194 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1196 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1197 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1199 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1200 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1202 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1203 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1206 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1207 // when we have a 256bit-wide blend with immediate.
1208 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1210 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1211 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1218 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1225 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1226 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1227 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1228 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1230 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1232 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1235 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1236 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1237 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1238 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1240 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1241 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1243 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1244 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1246 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1247 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1249 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1250 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1254 // In the customized shift lowering, the legal cases in AVX2 will be
1256 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1257 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1259 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1260 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1262 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1263 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1265 // Custom lower several nodes for 256-bit types.
1266 for (MVT VT : MVT::vector_valuetypes()) {
1267 if (VT.getScalarSizeInBits() >= 32) {
1268 setOperationAction(ISD::MLOAD, VT, Legal);
1269 setOperationAction(ISD::MSTORE, VT, Legal);
1271 // Extract subvector is special because the value type
1272 // (result) is 128-bit but the source is 256-bit wide.
1273 if (VT.is128BitVector()) {
1274 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1276 // Do not attempt to custom lower other non-256-bit vectors
1277 if (!VT.is256BitVector())
1280 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1281 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1282 setOperationAction(ISD::VSELECT, VT, Custom);
1283 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1284 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1285 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1286 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1287 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1290 if (Subtarget->hasInt256())
1291 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1293 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1294 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1295 setOperationAction(ISD::AND, VT, Promote);
1296 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1297 setOperationAction(ISD::OR, VT, Promote);
1298 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1299 setOperationAction(ISD::XOR, VT, Promote);
1300 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1301 setOperationAction(ISD::LOAD, VT, Promote);
1302 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1303 setOperationAction(ISD::SELECT, VT, Promote);
1304 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1308 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1309 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1310 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1311 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1312 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1314 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1315 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1316 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1318 for (MVT VT : MVT::fp_vector_valuetypes())
1319 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1321 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1322 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1323 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1324 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1325 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1326 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1327 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1328 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1329 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1330 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1331 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1332 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1334 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1335 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1336 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1337 setOperationAction(ISD::XOR, MVT::i1, Legal);
1338 setOperationAction(ISD::OR, MVT::i1, Legal);
1339 setOperationAction(ISD::AND, MVT::i1, Legal);
1340 setOperationAction(ISD::SUB, MVT::i1, Custom);
1341 setOperationAction(ISD::ADD, MVT::i1, Custom);
1342 setOperationAction(ISD::MUL, MVT::i1, Custom);
1343 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1344 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1345 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1346 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1349 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1350 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1351 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1355 setOperationAction(ISD::FABS, MVT::v16f32, Custom);
1357 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1358 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1359 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1363 setOperationAction(ISD::FABS, MVT::v8f64, Custom);
1364 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1365 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1367 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1368 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1369 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1371 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1376 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1377 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1381 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1382 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1384 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1385 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1387 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1388 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1389 if (Subtarget->hasVLX()){
1390 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1391 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1393 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1394 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1396 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1397 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1399 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1400 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1402 setOperationAction(ISD::MLOAD, MVT::v8i32, Custom);
1403 setOperationAction(ISD::MLOAD, MVT::v8f32, Custom);
1404 setOperationAction(ISD::MSTORE, MVT::v8i32, Custom);
1405 setOperationAction(ISD::MSTORE, MVT::v8f32, Custom);
1407 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1409 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1410 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1411 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1412 if (Subtarget->hasDQI()) {
1413 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1414 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1416 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1417 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1418 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1419 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1420 if (Subtarget->hasVLX()) {
1421 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1422 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1423 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1424 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1425 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1426 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1427 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1428 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1431 if (Subtarget->hasVLX()) {
1432 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1433 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1434 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1435 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1437 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1438 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1439 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1441 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1442 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1443 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1444 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1445 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1446 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1447 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1448 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1449 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1452 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1453 if (Subtarget->hasDQI()) {
1454 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1455 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1457 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1458 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1459 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1460 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1461 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1462 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1463 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1464 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1465 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1466 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1468 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
1474 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1475 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1477 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1479 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1480 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1481 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom);
1482 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1484 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1485 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1486 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1490 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1492 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1493 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1494 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1495 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1496 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1497 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1498 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1499 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1501 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1502 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1504 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1505 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1507 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1509 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1510 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1512 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1513 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1515 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1516 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1518 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1519 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1520 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1521 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1522 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1523 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1525 if (Subtarget->hasCDI()) {
1526 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1527 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1528 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1531 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1532 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1533 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1535 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1536 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1537 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1540 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1541 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1543 if (Subtarget->hasVLX()) {
1544 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1545 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1546 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1547 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1548 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1549 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1550 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1551 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1553 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1554 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1555 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1556 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1558 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1559 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1560 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1561 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1562 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1563 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1564 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1565 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1567 } // Subtarget->hasCDI()
1569 if (Subtarget->hasDQI()) {
1570 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1571 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1572 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1574 // Custom lower several nodes.
1575 for (MVT VT : MVT::vector_valuetypes()) {
1576 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1578 setOperationAction(ISD::AND, VT, Legal);
1579 setOperationAction(ISD::OR, VT, Legal);
1580 setOperationAction(ISD::XOR, VT, Legal);
1582 if ((VT.is128BitVector() || VT.is256BitVector()) && EltSize >= 32) {
1583 setOperationAction(ISD::MGATHER, VT, Custom);
1584 setOperationAction(ISD::MSCATTER, VT, Custom);
1586 // Extract subvector is special because the value type
1587 // (result) is 256/128-bit but the source is 512-bit wide.
1588 if (VT.is128BitVector() || VT.is256BitVector()) {
1589 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1591 if (VT.getVectorElementType() == MVT::i1)
1592 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1594 // Do not attempt to custom lower other non-512-bit vectors
1595 if (!VT.is512BitVector())
1598 if (EltSize >= 32) {
1599 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1600 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1601 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1602 setOperationAction(ISD::VSELECT, VT, Legal);
1603 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1604 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1605 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1606 setOperationAction(ISD::MLOAD, VT, Legal);
1607 setOperationAction(ISD::MSTORE, VT, Legal);
1608 setOperationAction(ISD::MGATHER, VT, Legal);
1609 setOperationAction(ISD::MSCATTER, VT, Custom);
1612 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1613 setOperationAction(ISD::SELECT, VT, Promote);
1614 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1618 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1619 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1620 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1622 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1623 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1625 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1626 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1627 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1628 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1629 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1630 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1631 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1632 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1633 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1634 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1635 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1636 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1637 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1640 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1641 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1645 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1646 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1647 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1648 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1649 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1650 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1651 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1654 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1655 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1656 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1660 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1661 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1662 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1663 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1664 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1668 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1669 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1670 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1671 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1672 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1673 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1674 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1675 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1677 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1678 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1679 if (Subtarget->hasVLX())
1680 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1682 if (Subtarget->hasCDI()) {
1683 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1684 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1685 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1686 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1689 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1690 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1691 setOperationAction(ISD::VSELECT, VT, Legal);
1695 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1696 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1697 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1699 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1700 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1701 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1702 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1703 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1704 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1705 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1706 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1707 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1708 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1712 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1713 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1714 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1715 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1716 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1717 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1718 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1719 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1721 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1722 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1723 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1724 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1725 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1726 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1727 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1728 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1731 // We want to custom lower some of our intrinsics.
1732 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1733 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1734 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1735 if (!Subtarget->is64Bit()) {
1736 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1737 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1740 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1741 // handle type legalization for these operations here.
1743 // FIXME: We really should do custom legalization for addition and
1744 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1745 // than generic legalization for 64-bit multiplication-with-overflow, though.
1746 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1747 if (VT == MVT::i64 && !Subtarget->is64Bit())
1749 // Add/Sub/Mul with overflow operations are custom lowered.
1750 setOperationAction(ISD::SADDO, VT, Custom);
1751 setOperationAction(ISD::UADDO, VT, Custom);
1752 setOperationAction(ISD::SSUBO, VT, Custom);
1753 setOperationAction(ISD::USUBO, VT, Custom);
1754 setOperationAction(ISD::SMULO, VT, Custom);
1755 setOperationAction(ISD::UMULO, VT, Custom);
1758 if (!Subtarget->is64Bit()) {
1759 // These libcalls are not available in 32-bit.
1760 setLibcallName(RTLIB::SHL_I128, nullptr);
1761 setLibcallName(RTLIB::SRL_I128, nullptr);
1762 setLibcallName(RTLIB::SRA_I128, nullptr);
1765 // Combine sin / cos into one node or libcall if possible.
1766 if (Subtarget->hasSinCos()) {
1767 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1768 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1769 if (Subtarget->isTargetDarwin()) {
1770 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1771 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1772 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1773 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1777 if (Subtarget->isTargetWin64()) {
1778 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1779 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1780 setOperationAction(ISD::SREM, MVT::i128, Custom);
1781 setOperationAction(ISD::UREM, MVT::i128, Custom);
1782 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1783 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1786 // We have target-specific dag combine patterns for the following nodes:
1787 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1788 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1789 setTargetDAGCombine(ISD::BITCAST);
1790 setTargetDAGCombine(ISD::VSELECT);
1791 setTargetDAGCombine(ISD::SELECT);
1792 setTargetDAGCombine(ISD::SHL);
1793 setTargetDAGCombine(ISD::SRA);
1794 setTargetDAGCombine(ISD::SRL);
1795 setTargetDAGCombine(ISD::OR);
1796 setTargetDAGCombine(ISD::AND);
1797 setTargetDAGCombine(ISD::ADD);
1798 setTargetDAGCombine(ISD::FADD);
1799 setTargetDAGCombine(ISD::FSUB);
1800 setTargetDAGCombine(ISD::FNEG);
1801 setTargetDAGCombine(ISD::FMA);
1802 setTargetDAGCombine(ISD::FMAXNUM);
1803 setTargetDAGCombine(ISD::SUB);
1804 setTargetDAGCombine(ISD::LOAD);
1805 setTargetDAGCombine(ISD::MLOAD);
1806 setTargetDAGCombine(ISD::STORE);
1807 setTargetDAGCombine(ISD::MSTORE);
1808 setTargetDAGCombine(ISD::TRUNCATE);
1809 setTargetDAGCombine(ISD::ZERO_EXTEND);
1810 setTargetDAGCombine(ISD::ANY_EXTEND);
1811 setTargetDAGCombine(ISD::SIGN_EXTEND);
1812 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1813 setTargetDAGCombine(ISD::SINT_TO_FP);
1814 setTargetDAGCombine(ISD::UINT_TO_FP);
1815 setTargetDAGCombine(ISD::SETCC);
1816 setTargetDAGCombine(ISD::BUILD_VECTOR);
1817 setTargetDAGCombine(ISD::MUL);
1818 setTargetDAGCombine(ISD::XOR);
1819 setTargetDAGCombine(ISD::MSCATTER);
1820 setTargetDAGCombine(ISD::MGATHER);
1822 computeRegisterProperties(Subtarget->getRegisterInfo());
1824 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1825 MaxStoresPerMemsetOptSize = 8;
1826 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1827 MaxStoresPerMemcpyOptSize = 4;
1828 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1829 MaxStoresPerMemmoveOptSize = 4;
1830 setPrefLoopAlignment(4); // 2^4 bytes.
1832 // A predictable cmov does not hurt on an in-order CPU.
1833 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1834 PredictableSelectIsExpensive = !Subtarget->isAtom();
1835 EnableExtLdPromotion = true;
1836 setPrefFunctionAlignment(4); // 2^4 bytes.
1838 verifyIntrinsicTables();
1841 // This has so far only been implemented for 64-bit MachO.
1842 bool X86TargetLowering::useLoadStackGuardNode() const {
1843 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1846 TargetLoweringBase::LegalizeTypeAction
1847 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1848 if (ExperimentalVectorWideningLegalization &&
1849 VT.getVectorNumElements() != 1 &&
1850 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1851 return TypeWidenVector;
1853 return TargetLoweringBase::getPreferredVectorAction(VT);
1856 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1859 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1861 if (VT.isSimple()) {
1862 MVT VVT = VT.getSimpleVT();
1863 const unsigned NumElts = VVT.getVectorNumElements();
1864 const MVT EltVT = VVT.getVectorElementType();
1865 if (VVT.is512BitVector()) {
1866 if (Subtarget->hasAVX512())
1867 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1868 EltVT == MVT::f32 || EltVT == MVT::f64)
1870 case 8: return MVT::v8i1;
1871 case 16: return MVT::v16i1;
1873 if (Subtarget->hasBWI())
1874 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1876 case 32: return MVT::v32i1;
1877 case 64: return MVT::v64i1;
1881 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1882 if (Subtarget->hasVLX())
1883 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1884 EltVT == MVT::f32 || EltVT == MVT::f64)
1886 case 2: return MVT::v2i1;
1887 case 4: return MVT::v4i1;
1888 case 8: return MVT::v8i1;
1890 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1891 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1893 case 8: return MVT::v8i1;
1894 case 16: return MVT::v16i1;
1895 case 32: return MVT::v32i1;
1900 return VT.changeVectorElementTypeToInteger();
1903 /// Helper for getByValTypeAlignment to determine
1904 /// the desired ByVal argument alignment.
1905 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1908 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1909 if (VTy->getBitWidth() == 128)
1911 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1912 unsigned EltAlign = 0;
1913 getMaxByValAlign(ATy->getElementType(), EltAlign);
1914 if (EltAlign > MaxAlign)
1915 MaxAlign = EltAlign;
1916 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1917 for (auto *EltTy : STy->elements()) {
1918 unsigned EltAlign = 0;
1919 getMaxByValAlign(EltTy, EltAlign);
1920 if (EltAlign > MaxAlign)
1921 MaxAlign = EltAlign;
1928 /// Return the desired alignment for ByVal aggregate
1929 /// function arguments in the caller parameter area. For X86, aggregates
1930 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1931 /// are at 4-byte boundaries.
1932 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1933 const DataLayout &DL) const {
1934 if (Subtarget->is64Bit()) {
1935 // Max of 8 and alignment of type.
1936 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1943 if (Subtarget->hasSSE1())
1944 getMaxByValAlign(Ty, Align);
1948 /// Returns the target specific optimal type for load
1949 /// and store operations as a result of memset, memcpy, and memmove
1950 /// lowering. If DstAlign is zero that means it's safe to destination
1951 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1952 /// means there isn't a need to check it against alignment requirement,
1953 /// probably because the source does not need to be loaded. If 'IsMemset' is
1954 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1955 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1956 /// source is constant so it does not need to be loaded.
1957 /// It returns EVT::Other if the type should be determined using generic
1958 /// target-independent logic.
1960 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1961 unsigned DstAlign, unsigned SrcAlign,
1962 bool IsMemset, bool ZeroMemset,
1964 MachineFunction &MF) const {
1965 const Function *F = MF.getFunction();
1966 if ((!IsMemset || ZeroMemset) &&
1967 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1969 (!Subtarget->isUnalignedMem16Slow() ||
1970 ((DstAlign == 0 || DstAlign >= 16) &&
1971 (SrcAlign == 0 || SrcAlign >= 16)))) {
1973 // FIXME: Check if unaligned 32-byte accesses are slow.
1974 if (Subtarget->hasInt256())
1976 if (Subtarget->hasFp256())
1979 if (Subtarget->hasSSE2())
1981 if (Subtarget->hasSSE1())
1983 } else if (!MemcpyStrSrc && Size >= 8 &&
1984 !Subtarget->is64Bit() &&
1985 Subtarget->hasSSE2()) {
1986 // Do not use f64 to lower memcpy if source is string constant. It's
1987 // better to use i32 to avoid the loads.
1991 // This is a compromise. If we reach here, unaligned accesses may be slow on
1992 // this target. However, creating smaller, aligned accesses could be even
1993 // slower and would certainly be a lot more code.
1994 if (Subtarget->is64Bit() && Size >= 8)
1999 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2001 return X86ScalarSSEf32;
2002 else if (VT == MVT::f64)
2003 return X86ScalarSSEf64;
2008 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2013 switch (VT.getSizeInBits()) {
2015 // 8-byte and under are always assumed to be fast.
2019 *Fast = !Subtarget->isUnalignedMem16Slow();
2022 *Fast = !Subtarget->isUnalignedMem32Slow();
2024 // TODO: What about AVX-512 (512-bit) accesses?
2027 // Misaligned accesses of any size are always allowed.
2031 /// Return the entry encoding for a jump table in the
2032 /// current function. The returned value is a member of the
2033 /// MachineJumpTableInfo::JTEntryKind enum.
2034 unsigned X86TargetLowering::getJumpTableEncoding() const {
2035 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2037 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2038 Subtarget->isPICStyleGOT())
2039 return MachineJumpTableInfo::EK_Custom32;
2041 // Otherwise, use the normal jump table encoding heuristics.
2042 return TargetLowering::getJumpTableEncoding();
2045 bool X86TargetLowering::useSoftFloat() const {
2046 return Subtarget->useSoftFloat();
2050 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2051 const MachineBasicBlock *MBB,
2052 unsigned uid,MCContext &Ctx) const{
2053 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2054 Subtarget->isPICStyleGOT());
2055 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2057 return MCSymbolRefExpr::create(MBB->getSymbol(),
2058 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2061 /// Returns relocation base for the given PIC jumptable.
2062 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2063 SelectionDAG &DAG) const {
2064 if (!Subtarget->is64Bit())
2065 // This doesn't have SDLoc associated with it, but is not really the
2066 // same as a Register.
2067 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2068 getPointerTy(DAG.getDataLayout()));
2072 /// This returns the relocation base for the given PIC jumptable,
2073 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2074 const MCExpr *X86TargetLowering::
2075 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2076 MCContext &Ctx) const {
2077 // X86-64 uses RIP relative addressing based on the jump table label.
2078 if (Subtarget->isPICStyleRIPRel())
2079 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2081 // Otherwise, the reference is relative to the PIC base.
2082 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2085 std::pair<const TargetRegisterClass *, uint8_t>
2086 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2088 const TargetRegisterClass *RRC = nullptr;
2090 switch (VT.SimpleTy) {
2092 return TargetLowering::findRepresentativeClass(TRI, VT);
2093 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2094 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2097 RRC = &X86::VR64RegClass;
2099 case MVT::f32: case MVT::f64:
2100 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2101 case MVT::v4f32: case MVT::v2f64:
2102 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2104 RRC = &X86::VR128RegClass;
2107 return std::make_pair(RRC, Cost);
2110 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2111 unsigned &Offset) const {
2112 if (!Subtarget->isTargetLinux())
2115 if (Subtarget->is64Bit()) {
2116 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2118 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2130 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2131 if (!Subtarget->isTargetAndroid())
2132 return TargetLowering::getSafeStackPointerLocation(IRB);
2134 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2135 // definition of TLS_SLOT_SAFESTACK in
2136 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2137 unsigned AddressSpace, Offset;
2138 if (Subtarget->is64Bit()) {
2139 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2141 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2151 return ConstantExpr::getIntToPtr(
2152 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2153 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2156 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2157 unsigned DestAS) const {
2158 assert(SrcAS != DestAS && "Expected different address spaces!");
2160 return SrcAS < 256 && DestAS < 256;
2163 //===----------------------------------------------------------------------===//
2164 // Return Value Calling Convention Implementation
2165 //===----------------------------------------------------------------------===//
2167 #include "X86GenCallingConv.inc"
2169 bool X86TargetLowering::CanLowerReturn(
2170 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2171 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2172 SmallVector<CCValAssign, 16> RVLocs;
2173 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2174 return CCInfo.CheckReturn(Outs, RetCC_X86);
2177 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2178 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2183 X86TargetLowering::LowerReturn(SDValue Chain,
2184 CallingConv::ID CallConv, bool isVarArg,
2185 const SmallVectorImpl<ISD::OutputArg> &Outs,
2186 const SmallVectorImpl<SDValue> &OutVals,
2187 SDLoc dl, SelectionDAG &DAG) const {
2188 MachineFunction &MF = DAG.getMachineFunction();
2189 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2191 SmallVector<CCValAssign, 16> RVLocs;
2192 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2193 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2196 SmallVector<SDValue, 6> RetOps;
2197 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2198 // Operand #1 = Bytes To Pop
2199 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2202 // Copy the result values into the output registers.
2203 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2204 CCValAssign &VA = RVLocs[i];
2205 assert(VA.isRegLoc() && "Can only return in registers!");
2206 SDValue ValToCopy = OutVals[i];
2207 EVT ValVT = ValToCopy.getValueType();
2209 // Promote values to the appropriate types.
2210 if (VA.getLocInfo() == CCValAssign::SExt)
2211 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2212 else if (VA.getLocInfo() == CCValAssign::ZExt)
2213 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2214 else if (VA.getLocInfo() == CCValAssign::AExt) {
2215 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2216 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2218 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2220 else if (VA.getLocInfo() == CCValAssign::BCvt)
2221 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2223 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2224 "Unexpected FP-extend for return value.");
2226 // If this is x86-64, and we disabled SSE, we can't return FP values,
2227 // or SSE or MMX vectors.
2228 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2229 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2230 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2231 report_fatal_error("SSE register return with SSE disabled");
2233 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2234 // llvm-gcc has never done it right and no one has noticed, so this
2235 // should be OK for now.
2236 if (ValVT == MVT::f64 &&
2237 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2238 report_fatal_error("SSE2 register return with SSE2 disabled");
2240 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2241 // the RET instruction and handled by the FP Stackifier.
2242 if (VA.getLocReg() == X86::FP0 ||
2243 VA.getLocReg() == X86::FP1) {
2244 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2245 // change the value to the FP stack register class.
2246 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2247 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2248 RetOps.push_back(ValToCopy);
2249 // Don't emit a copytoreg.
2253 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2254 // which is returned in RAX / RDX.
2255 if (Subtarget->is64Bit()) {
2256 if (ValVT == MVT::x86mmx) {
2257 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2258 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2259 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2261 // If we don't have SSE2 available, convert to v4f32 so the generated
2262 // register is legal.
2263 if (!Subtarget->hasSSE2())
2264 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2269 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2270 Flag = Chain.getValue(1);
2271 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2274 // All x86 ABIs require that for returning structs by value we copy
2275 // the sret argument into %rax/%eax (depending on ABI) for the return.
2276 // We saved the argument into a virtual register in the entry block,
2277 // so now we copy the value out and into %rax/%eax.
2279 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2280 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2281 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2282 // either case FuncInfo->setSRetReturnReg() will have been called.
2283 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2284 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2285 getPointerTy(MF.getDataLayout()));
2288 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2289 X86::RAX : X86::EAX;
2290 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2291 Flag = Chain.getValue(1);
2293 // RAX/EAX now acts like a return value.
2295 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2298 RetOps[0] = Chain; // Update chain.
2300 // Add the flag if we have it.
2302 RetOps.push_back(Flag);
2304 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2307 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2308 if (N->getNumValues() != 1)
2310 if (!N->hasNUsesOfValue(1, 0))
2313 SDValue TCChain = Chain;
2314 SDNode *Copy = *N->use_begin();
2315 if (Copy->getOpcode() == ISD::CopyToReg) {
2316 // If the copy has a glue operand, we conservatively assume it isn't safe to
2317 // perform a tail call.
2318 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2320 TCChain = Copy->getOperand(0);
2321 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2324 bool HasRet = false;
2325 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2327 if (UI->getOpcode() != X86ISD::RET_FLAG)
2329 // If we are returning more than one value, we can definitely
2330 // not make a tail call see PR19530
2331 if (UI->getNumOperands() > 4)
2333 if (UI->getNumOperands() == 4 &&
2334 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2347 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2348 ISD::NodeType ExtendKind) const {
2350 // TODO: Is this also valid on 32-bit?
2351 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2352 ReturnMVT = MVT::i8;
2354 ReturnMVT = MVT::i32;
2356 EVT MinVT = getRegisterType(Context, ReturnMVT);
2357 return VT.bitsLT(MinVT) ? MinVT : VT;
2360 /// Lower the result values of a call into the
2361 /// appropriate copies out of appropriate physical registers.
2364 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2365 CallingConv::ID CallConv, bool isVarArg,
2366 const SmallVectorImpl<ISD::InputArg> &Ins,
2367 SDLoc dl, SelectionDAG &DAG,
2368 SmallVectorImpl<SDValue> &InVals) const {
2370 // Assign locations to each value returned by this call.
2371 SmallVector<CCValAssign, 16> RVLocs;
2372 bool Is64Bit = Subtarget->is64Bit();
2373 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2375 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2377 // Copy all of the result registers out of their specified physreg.
2378 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2379 CCValAssign &VA = RVLocs[i];
2380 EVT CopyVT = VA.getLocVT();
2382 // If this is x86-64, and we disabled SSE, we can't return FP values
2383 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) &&
2384 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2385 report_fatal_error("SSE register return with SSE disabled");
2388 // If we prefer to use the value in xmm registers, copy it out as f80 and
2389 // use a truncate to move it from fp stack reg to xmm reg.
2390 bool RoundAfterCopy = false;
2391 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2392 isScalarFPTypeInSSEReg(VA.getValVT())) {
2394 RoundAfterCopy = (CopyVT != VA.getLocVT());
2397 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2398 CopyVT, InFlag).getValue(1);
2399 SDValue Val = Chain.getValue(0);
2402 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2403 // This truncation won't change the value.
2404 DAG.getIntPtrConstant(1, dl));
2406 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2407 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2409 InFlag = Chain.getValue(2);
2410 InVals.push_back(Val);
2416 //===----------------------------------------------------------------------===//
2417 // C & StdCall & Fast Calling Convention implementation
2418 //===----------------------------------------------------------------------===//
2419 // StdCall calling convention seems to be standard for many Windows' API
2420 // routines and around. It differs from C calling convention just a little:
2421 // callee should clean up the stack, not caller. Symbols should be also
2422 // decorated in some fancy way :) It doesn't support any vector arguments.
2423 // For info on fast calling convention see Fast Calling Convention (tail call)
2424 // implementation LowerX86_32FastCCCallTo.
2426 /// CallIsStructReturn - Determines whether a call uses struct return
2428 enum StructReturnType {
2433 static StructReturnType
2434 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2436 return NotStructReturn;
2438 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2439 if (!Flags.isSRet())
2440 return NotStructReturn;
2441 if (Flags.isInReg())
2442 return RegStructReturn;
2443 return StackStructReturn;
2446 /// Determines whether a function uses struct return semantics.
2447 static StructReturnType
2448 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2450 return NotStructReturn;
2452 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2453 if (!Flags.isSRet())
2454 return NotStructReturn;
2455 if (Flags.isInReg())
2456 return RegStructReturn;
2457 return StackStructReturn;
2460 /// Make a copy of an aggregate at address specified by "Src" to address
2461 /// "Dst" with size and alignment information specified by the specific
2462 /// parameter attribute. The copy will be passed as a byval function parameter.
2464 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2465 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2467 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2469 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2470 /*isVolatile*/false, /*AlwaysInline=*/true,
2471 /*isTailCall*/false,
2472 MachinePointerInfo(), MachinePointerInfo());
2475 /// Return true if the calling convention is one that we can guarantee TCO for.
2476 static bool canGuaranteeTCO(CallingConv::ID CC) {
2477 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2478 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2481 /// Return true if we might ever do TCO for calls with this calling convention.
2482 static bool mayTailCallThisCC(CallingConv::ID CC) {
2484 // C calling conventions:
2485 case CallingConv::C:
2486 case CallingConv::X86_64_Win64:
2487 case CallingConv::X86_64_SysV:
2488 // Callee pop conventions:
2489 case CallingConv::X86_ThisCall:
2490 case CallingConv::X86_StdCall:
2491 case CallingConv::X86_VectorCall:
2492 case CallingConv::X86_FastCall:
2495 return canGuaranteeTCO(CC);
2499 /// Return true if the function is being made into a tailcall target by
2500 /// changing its ABI.
2501 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2502 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2505 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2507 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2508 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2512 CallingConv::ID CalleeCC = CS.getCallingConv();
2513 if (!mayTailCallThisCC(CalleeCC))
2520 X86TargetLowering::LowerMemArgument(SDValue Chain,
2521 CallingConv::ID CallConv,
2522 const SmallVectorImpl<ISD::InputArg> &Ins,
2523 SDLoc dl, SelectionDAG &DAG,
2524 const CCValAssign &VA,
2525 MachineFrameInfo *MFI,
2527 // Create the nodes corresponding to a load from this parameter slot.
2528 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2529 bool AlwaysUseMutable = shouldGuaranteeTCO(
2530 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2531 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2534 // If value is passed by pointer we have address passed instead of the value
2536 bool ExtendedInMem = VA.isExtInLoc() &&
2537 VA.getValVT().getScalarType() == MVT::i1;
2539 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2540 ValVT = VA.getLocVT();
2542 ValVT = VA.getValVT();
2544 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2545 // changed with more analysis.
2546 // In case of tail call optimization mark all arguments mutable. Since they
2547 // could be overwritten by lowering of arguments in case of a tail call.
2548 if (Flags.isByVal()) {
2549 unsigned Bytes = Flags.getByValSize();
2550 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2551 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2552 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2554 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2555 VA.getLocMemOffset(), isImmutable);
2556 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2557 SDValue Val = DAG.getLoad(
2558 ValVT, dl, Chain, FIN,
2559 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2561 return ExtendedInMem ?
2562 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2566 // FIXME: Get this from tablegen.
2567 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2568 const X86Subtarget *Subtarget) {
2569 assert(Subtarget->is64Bit());
2571 if (Subtarget->isCallingConvWin64(CallConv)) {
2572 static const MCPhysReg GPR64ArgRegsWin64[] = {
2573 X86::RCX, X86::RDX, X86::R8, X86::R9
2575 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2578 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2579 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2581 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2584 // FIXME: Get this from tablegen.
2585 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2586 CallingConv::ID CallConv,
2587 const X86Subtarget *Subtarget) {
2588 assert(Subtarget->is64Bit());
2589 if (Subtarget->isCallingConvWin64(CallConv)) {
2590 // The XMM registers which might contain var arg parameters are shadowed
2591 // in their paired GPR. So we only need to save the GPR to their home
2593 // TODO: __vectorcall will change this.
2597 const Function *Fn = MF.getFunction();
2598 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2599 bool isSoftFloat = Subtarget->useSoftFloat();
2600 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2601 "SSE register cannot be used when SSE is disabled!");
2602 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2603 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2607 static const MCPhysReg XMMArgRegs64Bit[] = {
2608 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2609 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2611 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2614 SDValue X86TargetLowering::LowerFormalArguments(
2615 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2616 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2617 SmallVectorImpl<SDValue> &InVals) const {
2618 MachineFunction &MF = DAG.getMachineFunction();
2619 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2620 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2622 const Function* Fn = MF.getFunction();
2623 if (Fn->hasExternalLinkage() &&
2624 Subtarget->isTargetCygMing() &&
2625 Fn->getName() == "main")
2626 FuncInfo->setForceFramePointer(true);
2628 MachineFrameInfo *MFI = MF.getFrameInfo();
2629 bool Is64Bit = Subtarget->is64Bit();
2630 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2632 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2633 "Var args not supported with calling convention fastcc, ghc or hipe");
2635 // Assign locations to all of the incoming arguments.
2636 SmallVector<CCValAssign, 16> ArgLocs;
2637 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2639 // Allocate shadow area for Win64
2641 CCInfo.AllocateStack(32, 8);
2643 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2645 unsigned LastVal = ~0U;
2647 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2648 CCValAssign &VA = ArgLocs[i];
2649 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2651 assert(VA.getValNo() != LastVal &&
2652 "Don't support value assigned to multiple locs yet");
2654 LastVal = VA.getValNo();
2656 if (VA.isRegLoc()) {
2657 EVT RegVT = VA.getLocVT();
2658 const TargetRegisterClass *RC;
2659 if (RegVT == MVT::i32)
2660 RC = &X86::GR32RegClass;
2661 else if (Is64Bit && RegVT == MVT::i64)
2662 RC = &X86::GR64RegClass;
2663 else if (RegVT == MVT::f32)
2664 RC = &X86::FR32RegClass;
2665 else if (RegVT == MVT::f64)
2666 RC = &X86::FR64RegClass;
2667 else if (RegVT == MVT::f128)
2668 RC = &X86::FR128RegClass;
2669 else if (RegVT.is512BitVector())
2670 RC = &X86::VR512RegClass;
2671 else if (RegVT.is256BitVector())
2672 RC = &X86::VR256RegClass;
2673 else if (RegVT.is128BitVector())
2674 RC = &X86::VR128RegClass;
2675 else if (RegVT == MVT::x86mmx)
2676 RC = &X86::VR64RegClass;
2677 else if (RegVT == MVT::i1)
2678 RC = &X86::VK1RegClass;
2679 else if (RegVT == MVT::v8i1)
2680 RC = &X86::VK8RegClass;
2681 else if (RegVT == MVT::v16i1)
2682 RC = &X86::VK16RegClass;
2683 else if (RegVT == MVT::v32i1)
2684 RC = &X86::VK32RegClass;
2685 else if (RegVT == MVT::v64i1)
2686 RC = &X86::VK64RegClass;
2688 llvm_unreachable("Unknown argument type!");
2690 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2691 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2693 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2694 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2696 if (VA.getLocInfo() == CCValAssign::SExt)
2697 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2698 DAG.getValueType(VA.getValVT()));
2699 else if (VA.getLocInfo() == CCValAssign::ZExt)
2700 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2701 DAG.getValueType(VA.getValVT()));
2702 else if (VA.getLocInfo() == CCValAssign::BCvt)
2703 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2705 if (VA.isExtInLoc()) {
2706 // Handle MMX values passed in XMM regs.
2707 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2708 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2710 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2713 assert(VA.isMemLoc());
2714 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2717 // If value is passed via pointer - do a load.
2718 if (VA.getLocInfo() == CCValAssign::Indirect)
2719 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2720 MachinePointerInfo(), false, false, false, 0);
2722 InVals.push_back(ArgValue);
2725 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2726 // All x86 ABIs require that for returning structs by value we copy the
2727 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2728 // the argument into a virtual register so that we can access it from the
2730 if (Ins[i].Flags.isSRet()) {
2731 unsigned Reg = FuncInfo->getSRetReturnReg();
2733 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2734 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2735 FuncInfo->setSRetReturnReg(Reg);
2737 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2738 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2743 unsigned StackSize = CCInfo.getNextStackOffset();
2744 // Align stack specially for tail calls.
2745 if (shouldGuaranteeTCO(CallConv,
2746 MF.getTarget().Options.GuaranteedTailCallOpt))
2747 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2749 // If the function takes variable number of arguments, make a frame index for
2750 // the start of the first vararg value... for expansion of llvm.va_start. We
2751 // can skip this if there are no va_start calls.
2752 if (MFI->hasVAStart() &&
2753 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2754 CallConv != CallingConv::X86_ThisCall))) {
2755 FuncInfo->setVarArgsFrameIndex(
2756 MFI->CreateFixedObject(1, StackSize, true));
2759 // Figure out if XMM registers are in use.
2760 assert(!(Subtarget->useSoftFloat() &&
2761 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2762 "SSE register cannot be used when SSE is disabled!");
2764 // 64-bit calling conventions support varargs and register parameters, so we
2765 // have to do extra work to spill them in the prologue.
2766 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2767 // Find the first unallocated argument registers.
2768 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2769 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2770 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2771 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2772 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2773 "SSE register cannot be used when SSE is disabled!");
2775 // Gather all the live in physical registers.
2776 SmallVector<SDValue, 6> LiveGPRs;
2777 SmallVector<SDValue, 8> LiveXMMRegs;
2779 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2780 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2782 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2784 if (!ArgXMMs.empty()) {
2785 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2786 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2787 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2788 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2789 LiveXMMRegs.push_back(
2790 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2795 // Get to the caller-allocated home save location. Add 8 to account
2796 // for the return address.
2797 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2798 FuncInfo->setRegSaveFrameIndex(
2799 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2800 // Fixup to set vararg frame on shadow area (4 x i64).
2802 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2804 // For X86-64, if there are vararg parameters that are passed via
2805 // registers, then we must store them to their spots on the stack so
2806 // they may be loaded by deferencing the result of va_next.
2807 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2808 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2809 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2810 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2813 // Store the integer parameter registers.
2814 SmallVector<SDValue, 8> MemOps;
2815 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2816 getPointerTy(DAG.getDataLayout()));
2817 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2818 for (SDValue Val : LiveGPRs) {
2819 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2820 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2822 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2823 MachinePointerInfo::getFixedStack(
2824 DAG.getMachineFunction(),
2825 FuncInfo->getRegSaveFrameIndex(), Offset),
2827 MemOps.push_back(Store);
2831 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2832 // Now store the XMM (fp + vector) parameter registers.
2833 SmallVector<SDValue, 12> SaveXMMOps;
2834 SaveXMMOps.push_back(Chain);
2835 SaveXMMOps.push_back(ALVal);
2836 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2837 FuncInfo->getRegSaveFrameIndex(), dl));
2838 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2839 FuncInfo->getVarArgsFPOffset(), dl));
2840 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2842 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2843 MVT::Other, SaveXMMOps));
2846 if (!MemOps.empty())
2847 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2850 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2851 // Find the largest legal vector type.
2852 MVT VecVT = MVT::Other;
2853 // FIXME: Only some x86_32 calling conventions support AVX512.
2854 if (Subtarget->hasAVX512() &&
2855 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2856 CallConv == CallingConv::Intel_OCL_BI)))
2857 VecVT = MVT::v16f32;
2858 else if (Subtarget->hasAVX())
2860 else if (Subtarget->hasSSE2())
2863 // We forward some GPRs and some vector types.
2864 SmallVector<MVT, 2> RegParmTypes;
2865 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2866 RegParmTypes.push_back(IntVT);
2867 if (VecVT != MVT::Other)
2868 RegParmTypes.push_back(VecVT);
2870 // Compute the set of forwarded registers. The rest are scratch.
2871 SmallVectorImpl<ForwardedRegister> &Forwards =
2872 FuncInfo->getForwardedMustTailRegParms();
2873 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2875 // Conservatively forward AL on x86_64, since it might be used for varargs.
2876 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2877 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2878 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2881 // Copy all forwards from physical to virtual registers.
2882 for (ForwardedRegister &F : Forwards) {
2883 // FIXME: Can we use a less constrained schedule?
2884 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2885 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2886 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2890 // Some CCs need callee pop.
2891 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2892 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2893 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2895 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2896 // If this is an sret function, the return should pop the hidden pointer.
2897 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2898 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2899 argsAreStructReturn(Ins) == StackStructReturn)
2900 FuncInfo->setBytesToPopOnReturn(4);
2904 // RegSaveFrameIndex is X86-64 only.
2905 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2906 if (CallConv == CallingConv::X86_FastCall ||
2907 CallConv == CallingConv::X86_ThisCall)
2908 // fastcc functions can't have varargs.
2909 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2912 FuncInfo->setArgumentStackSize(StackSize);
2914 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2915 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2916 if (Personality == EHPersonality::CoreCLR) {
2918 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2919 // that we'd prefer this slot be allocated towards the bottom of the frame
2920 // (i.e. near the stack pointer after allocating the frame). Every
2921 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2922 // offset from the bottom of this and each funclet's frame must be the
2923 // same, so the size of funclets' (mostly empty) frames is dictated by
2924 // how far this slot is from the bottom (since they allocate just enough
2925 // space to accomodate holding this slot at the correct offset).
2926 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2927 EHInfo->PSPSymFrameIdx = PSPSymFI;
2935 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2936 SDValue StackPtr, SDValue Arg,
2937 SDLoc dl, SelectionDAG &DAG,
2938 const CCValAssign &VA,
2939 ISD::ArgFlagsTy Flags) const {
2940 unsigned LocMemOffset = VA.getLocMemOffset();
2941 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2942 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2944 if (Flags.isByVal())
2945 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2947 return DAG.getStore(
2948 Chain, dl, Arg, PtrOff,
2949 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2953 /// Emit a load of return address if tail call
2954 /// optimization is performed and it is required.
2956 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2957 SDValue &OutRetAddr, SDValue Chain,
2958 bool IsTailCall, bool Is64Bit,
2959 int FPDiff, SDLoc dl) const {
2960 // Adjust the Return address stack slot.
2961 EVT VT = getPointerTy(DAG.getDataLayout());
2962 OutRetAddr = getReturnAddressFrameIndex(DAG);
2964 // Load the "old" Return address.
2965 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2966 false, false, false, 0);
2967 return SDValue(OutRetAddr.getNode(), 1);
2970 /// Emit a store of the return address if tail call
2971 /// optimization is performed and it is required (FPDiff!=0).
2972 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2973 SDValue Chain, SDValue RetAddrFrIdx,
2974 EVT PtrVT, unsigned SlotSize,
2975 int FPDiff, SDLoc dl) {
2976 // Store the return address to the appropriate stack slot.
2977 if (!FPDiff) return Chain;
2978 // Calculate the new stack slot for the return address.
2979 int NewReturnAddrFI =
2980 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2982 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2983 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2984 MachinePointerInfo::getFixedStack(
2985 DAG.getMachineFunction(), NewReturnAddrFI),
2990 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2991 /// operation of specified width.
2992 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
2994 unsigned NumElems = VT.getVectorNumElements();
2995 SmallVector<int, 8> Mask;
2996 Mask.push_back(NumElems);
2997 for (unsigned i = 1; i != NumElems; ++i)
2999 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3003 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3004 SmallVectorImpl<SDValue> &InVals) const {
3005 SelectionDAG &DAG = CLI.DAG;
3007 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3008 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3009 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3010 SDValue Chain = CLI.Chain;
3011 SDValue Callee = CLI.Callee;
3012 CallingConv::ID CallConv = CLI.CallConv;
3013 bool &isTailCall = CLI.IsTailCall;
3014 bool isVarArg = CLI.IsVarArg;
3016 MachineFunction &MF = DAG.getMachineFunction();
3017 bool Is64Bit = Subtarget->is64Bit();
3018 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3019 StructReturnType SR = callIsStructReturn(Outs);
3020 bool IsSibcall = false;
3021 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3022 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3024 if (Attr.getValueAsString() == "true")
3027 if (Subtarget->isPICStyleGOT() &&
3028 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3029 // If we are using a GOT, disable tail calls to external symbols with
3030 // default visibility. Tail calling such a symbol requires using a GOT
3031 // relocation, which forces early binding of the symbol. This breaks code
3032 // that require lazy function symbol resolution. Using musttail or
3033 // GuaranteedTailCallOpt will override this.
3034 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3035 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3036 G->getGlobal()->hasDefaultVisibility()))
3040 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3042 // Force this to be a tail call. The verifier rules are enough to ensure
3043 // that we can lower this successfully without moving the return address
3046 } else if (isTailCall) {
3047 // Check if it's really possible to do a tail call.
3048 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3049 isVarArg, SR != NotStructReturn,
3050 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3051 Outs, OutVals, Ins, DAG);
3053 // Sibcalls are automatically detected tailcalls which do not require
3055 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3062 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3063 "Var args not supported with calling convention fastcc, ghc or hipe");
3065 // Analyze operands of the call, assigning locations to each operand.
3066 SmallVector<CCValAssign, 16> ArgLocs;
3067 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3069 // Allocate shadow area for Win64
3071 CCInfo.AllocateStack(32, 8);
3073 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3075 // Get a count of how many bytes are to be pushed on the stack.
3076 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3078 // This is a sibcall. The memory operands are available in caller's
3079 // own caller's stack.
3081 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3082 canGuaranteeTCO(CallConv))
3083 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3086 if (isTailCall && !IsSibcall && !IsMustTail) {
3087 // Lower arguments at fp - stackoffset + fpdiff.
3088 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3090 FPDiff = NumBytesCallerPushed - NumBytes;
3092 // Set the delta of movement of the returnaddr stackslot.
3093 // But only set if delta is greater than previous delta.
3094 if (FPDiff < X86Info->getTCReturnAddrDelta())
3095 X86Info->setTCReturnAddrDelta(FPDiff);
3098 unsigned NumBytesToPush = NumBytes;
3099 unsigned NumBytesToPop = NumBytes;
3101 // If we have an inalloca argument, all stack space has already been allocated
3102 // for us and be right at the top of the stack. We don't support multiple
3103 // arguments passed in memory when using inalloca.
3104 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3106 if (!ArgLocs.back().isMemLoc())
3107 report_fatal_error("cannot use inalloca attribute on a register "
3109 if (ArgLocs.back().getLocMemOffset() != 0)
3110 report_fatal_error("any parameter with the inalloca attribute must be "
3111 "the only memory argument");
3115 Chain = DAG.getCALLSEQ_START(
3116 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3118 SDValue RetAddrFrIdx;
3119 // Load return address for tail calls.
3120 if (isTailCall && FPDiff)
3121 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3122 Is64Bit, FPDiff, dl);
3124 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3125 SmallVector<SDValue, 8> MemOpChains;
3128 // Walk the register/memloc assignments, inserting copies/loads. In the case
3129 // of tail call optimization arguments are handle later.
3130 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3131 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3132 // Skip inalloca arguments, they have already been written.
3133 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3134 if (Flags.isInAlloca())
3137 CCValAssign &VA = ArgLocs[i];
3138 EVT RegVT = VA.getLocVT();
3139 SDValue Arg = OutVals[i];
3140 bool isByVal = Flags.isByVal();
3142 // Promote the value if needed.
3143 switch (VA.getLocInfo()) {
3144 default: llvm_unreachable("Unknown loc info!");
3145 case CCValAssign::Full: break;
3146 case CCValAssign::SExt:
3147 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3149 case CCValAssign::ZExt:
3150 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3152 case CCValAssign::AExt:
3153 if (Arg.getValueType().isVector() &&
3154 Arg.getValueType().getVectorElementType() == MVT::i1)
3155 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3156 else if (RegVT.is128BitVector()) {
3157 // Special case: passing MMX values in XMM registers.
3158 Arg = DAG.getBitcast(MVT::i64, Arg);
3159 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3160 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3162 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3164 case CCValAssign::BCvt:
3165 Arg = DAG.getBitcast(RegVT, Arg);
3167 case CCValAssign::Indirect: {
3168 // Store the argument.
3169 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3170 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3171 Chain = DAG.getStore(
3172 Chain, dl, Arg, SpillSlot,
3173 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3180 if (VA.isRegLoc()) {
3181 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3182 if (isVarArg && IsWin64) {
3183 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3184 // shadow reg if callee is a varargs function.
3185 unsigned ShadowReg = 0;
3186 switch (VA.getLocReg()) {
3187 case X86::XMM0: ShadowReg = X86::RCX; break;
3188 case X86::XMM1: ShadowReg = X86::RDX; break;
3189 case X86::XMM2: ShadowReg = X86::R8; break;
3190 case X86::XMM3: ShadowReg = X86::R9; break;
3193 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3195 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3196 assert(VA.isMemLoc());
3197 if (!StackPtr.getNode())
3198 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3199 getPointerTy(DAG.getDataLayout()));
3200 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3201 dl, DAG, VA, Flags));
3205 if (!MemOpChains.empty())
3206 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3208 if (Subtarget->isPICStyleGOT()) {
3209 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3212 RegsToPass.push_back(std::make_pair(
3213 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3214 getPointerTy(DAG.getDataLayout()))));
3216 // If we are tail calling and generating PIC/GOT style code load the
3217 // address of the callee into ECX. The value in ecx is used as target of
3218 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3219 // for tail calls on PIC/GOT architectures. Normally we would just put the
3220 // address of GOT into ebx and then call target@PLT. But for tail calls
3221 // ebx would be restored (since ebx is callee saved) before jumping to the
3224 // Note: The actual moving to ECX is done further down.
3225 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3226 if (G && !G->getGlobal()->hasLocalLinkage() &&
3227 G->getGlobal()->hasDefaultVisibility())
3228 Callee = LowerGlobalAddress(Callee, DAG);
3229 else if (isa<ExternalSymbolSDNode>(Callee))
3230 Callee = LowerExternalSymbol(Callee, DAG);
3234 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3235 // From AMD64 ABI document:
3236 // For calls that may call functions that use varargs or stdargs
3237 // (prototype-less calls or calls to functions containing ellipsis (...) in
3238 // the declaration) %al is used as hidden argument to specify the number
3239 // of SSE registers used. The contents of %al do not need to match exactly
3240 // the number of registers, but must be an ubound on the number of SSE
3241 // registers used and is in the range 0 - 8 inclusive.
3243 // Count the number of XMM registers allocated.
3244 static const MCPhysReg XMMArgRegs[] = {
3245 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3246 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3248 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3249 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3250 && "SSE registers cannot be used when SSE is disabled");
3252 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3253 DAG.getConstant(NumXMMRegs, dl,
3257 if (isVarArg && IsMustTail) {
3258 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3259 for (const auto &F : Forwards) {
3260 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3261 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3265 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3266 // don't need this because the eligibility check rejects calls that require
3267 // shuffling arguments passed in memory.
3268 if (!IsSibcall && isTailCall) {
3269 // Force all the incoming stack arguments to be loaded from the stack
3270 // before any new outgoing arguments are stored to the stack, because the
3271 // outgoing stack slots may alias the incoming argument stack slots, and
3272 // the alias isn't otherwise explicit. This is slightly more conservative
3273 // than necessary, because it means that each store effectively depends
3274 // on every argument instead of just those arguments it would clobber.
3275 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3277 SmallVector<SDValue, 8> MemOpChains2;
3280 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3281 CCValAssign &VA = ArgLocs[i];
3284 assert(VA.isMemLoc());
3285 SDValue Arg = OutVals[i];
3286 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3287 // Skip inalloca arguments. They don't require any work.
3288 if (Flags.isInAlloca())
3290 // Create frame index.
3291 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3292 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3293 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3294 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3296 if (Flags.isByVal()) {
3297 // Copy relative to framepointer.
3298 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3299 if (!StackPtr.getNode())
3300 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3301 getPointerTy(DAG.getDataLayout()));
3302 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3305 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3309 // Store relative to framepointer.
3310 MemOpChains2.push_back(DAG.getStore(
3311 ArgChain, dl, Arg, FIN,
3312 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3317 if (!MemOpChains2.empty())
3318 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3320 // Store the return address to the appropriate stack slot.
3321 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3322 getPointerTy(DAG.getDataLayout()),
3323 RegInfo->getSlotSize(), FPDiff, dl);
3326 // Build a sequence of copy-to-reg nodes chained together with token chain
3327 // and flag operands which copy the outgoing args into registers.
3329 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3330 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3331 RegsToPass[i].second, InFlag);
3332 InFlag = Chain.getValue(1);
3335 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3336 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3337 // In the 64-bit large code model, we have to make all calls
3338 // through a register, since the call instruction's 32-bit
3339 // pc-relative offset may not be large enough to hold the whole
3341 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3342 // If the callee is a GlobalAddress node (quite common, every direct call
3343 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3345 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3347 // We should use extra load for direct calls to dllimported functions in
3349 const GlobalValue *GV = G->getGlobal();
3350 if (!GV->hasDLLImportStorageClass()) {
3351 unsigned char OpFlags = 0;
3352 bool ExtraLoad = false;
3353 unsigned WrapperKind = ISD::DELETED_NODE;
3355 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3356 // external symbols most go through the PLT in PIC mode. If the symbol
3357 // has hidden or protected visibility, or if it is static or local, then
3358 // we don't need to use the PLT - we can directly call it.
3359 if (Subtarget->isTargetELF() &&
3360 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3361 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3362 OpFlags = X86II::MO_PLT;
3363 } else if (Subtarget->isPICStyleStubAny() &&
3364 !GV->isStrongDefinitionForLinker() &&
3365 (!Subtarget->getTargetTriple().isMacOSX() ||
3366 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3367 // PC-relative references to external symbols should go through $stub,
3368 // unless we're building with the leopard linker or later, which
3369 // automatically synthesizes these stubs.
3370 OpFlags = X86II::MO_DARWIN_STUB;
3371 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3372 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3373 // If the function is marked as non-lazy, generate an indirect call
3374 // which loads from the GOT directly. This avoids runtime overhead
3375 // at the cost of eager binding (and one extra byte of encoding).
3376 OpFlags = X86II::MO_GOTPCREL;
3377 WrapperKind = X86ISD::WrapperRIP;
3381 Callee = DAG.getTargetGlobalAddress(
3382 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3384 // Add a wrapper if needed.
3385 if (WrapperKind != ISD::DELETED_NODE)
3386 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3387 getPointerTy(DAG.getDataLayout()), Callee);
3388 // Add extra indirection if needed.
3390 Callee = DAG.getLoad(
3391 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3392 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3395 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3396 unsigned char OpFlags = 0;
3398 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3399 // external symbols should go through the PLT.
3400 if (Subtarget->isTargetELF() &&
3401 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3402 OpFlags = X86II::MO_PLT;
3403 } else if (Subtarget->isPICStyleStubAny() &&
3404 (!Subtarget->getTargetTriple().isMacOSX() ||
3405 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3406 // PC-relative references to external symbols should go through $stub,
3407 // unless we're building with the leopard linker or later, which
3408 // automatically synthesizes these stubs.
3409 OpFlags = X86II::MO_DARWIN_STUB;
3412 Callee = DAG.getTargetExternalSymbol(
3413 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3414 } else if (Subtarget->isTarget64BitILP32() &&
3415 Callee->getValueType(0) == MVT::i32) {
3416 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3417 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3420 // Returns a chain & a flag for retval copy to use.
3421 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3422 SmallVector<SDValue, 8> Ops;
3424 if (!IsSibcall && isTailCall) {
3425 Chain = DAG.getCALLSEQ_END(Chain,
3426 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3427 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3428 InFlag = Chain.getValue(1);
3431 Ops.push_back(Chain);
3432 Ops.push_back(Callee);
3435 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3437 // Add argument registers to the end of the list so that they are known live
3439 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3440 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3441 RegsToPass[i].second.getValueType()));
3443 // Add a register mask operand representing the call-preserved registers.
3444 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3445 assert(Mask && "Missing call preserved mask for calling convention");
3447 // If this is an invoke in a 32-bit function using a funclet-based
3448 // personality, assume the function clobbers all registers. If an exception
3449 // is thrown, the runtime will not restore CSRs.
3450 // FIXME: Model this more precisely so that we can register allocate across
3451 // the normal edge and spill and fill across the exceptional edge.
3452 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3453 const Function *CallerFn = MF.getFunction();
3454 EHPersonality Pers =
3455 CallerFn->hasPersonalityFn()
3456 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3457 : EHPersonality::Unknown;
3458 if (isFuncletEHPersonality(Pers))
3459 Mask = RegInfo->getNoPreservedMask();
3462 Ops.push_back(DAG.getRegisterMask(Mask));
3464 if (InFlag.getNode())
3465 Ops.push_back(InFlag);
3469 //// If this is the first return lowered for this function, add the regs
3470 //// to the liveout set for the function.
3471 // This isn't right, although it's probably harmless on x86; liveouts
3472 // should be computed from returns not tail calls. Consider a void
3473 // function making a tail call to a function returning int.
3474 MF.getFrameInfo()->setHasTailCall();
3475 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3478 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3479 InFlag = Chain.getValue(1);
3481 // Create the CALLSEQ_END node.
3482 unsigned NumBytesForCalleeToPop;
3483 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3484 DAG.getTarget().Options.GuaranteedTailCallOpt))
3485 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3486 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3487 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3488 SR == StackStructReturn)
3489 // If this is a call to a struct-return function, the callee
3490 // pops the hidden struct pointer, so we have to push it back.
3491 // This is common for Darwin/X86, Linux & Mingw32 targets.
3492 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3493 NumBytesForCalleeToPop = 4;
3495 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3497 // Returns a flag for retval copy to use.
3499 Chain = DAG.getCALLSEQ_END(Chain,
3500 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3501 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3504 InFlag = Chain.getValue(1);
3507 // Handle result values, copying them out of physregs into vregs that we
3509 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3510 Ins, dl, DAG, InVals);
3513 //===----------------------------------------------------------------------===//
3514 // Fast Calling Convention (tail call) implementation
3515 //===----------------------------------------------------------------------===//
3517 // Like std call, callee cleans arguments, convention except that ECX is
3518 // reserved for storing the tail called function address. Only 2 registers are
3519 // free for argument passing (inreg). Tail call optimization is performed
3521 // * tailcallopt is enabled
3522 // * caller/callee are fastcc
3523 // On X86_64 architecture with GOT-style position independent code only local
3524 // (within module) calls are supported at the moment.
3525 // To keep the stack aligned according to platform abi the function
3526 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3527 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3528 // If a tail called function callee has more arguments than the caller the
3529 // caller needs to make sure that there is room to move the RETADDR to. This is
3530 // achieved by reserving an area the size of the argument delta right after the
3531 // original RETADDR, but before the saved framepointer or the spilled registers
3532 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3544 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3547 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3548 SelectionDAG& DAG) const {
3549 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3550 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3551 unsigned StackAlignment = TFI.getStackAlignment();
3552 uint64_t AlignMask = StackAlignment - 1;
3553 int64_t Offset = StackSize;
3554 unsigned SlotSize = RegInfo->getSlotSize();
3555 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3556 // Number smaller than 12 so just add the difference.
3557 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3559 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3560 Offset = ((~AlignMask) & Offset) + StackAlignment +
3561 (StackAlignment-SlotSize);
3566 /// Return true if the given stack call argument is already available in the
3567 /// same position (relatively) of the caller's incoming argument stack.
3569 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3570 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3571 const X86InstrInfo *TII) {
3572 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3574 if (Arg.getOpcode() == ISD::CopyFromReg) {
3575 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3576 if (!TargetRegisterInfo::isVirtualRegister(VR))
3578 MachineInstr *Def = MRI->getVRegDef(VR);
3581 if (!Flags.isByVal()) {
3582 if (!TII->isLoadFromStackSlot(Def, FI))
3585 unsigned Opcode = Def->getOpcode();
3586 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3587 Opcode == X86::LEA64_32r) &&
3588 Def->getOperand(1).isFI()) {
3589 FI = Def->getOperand(1).getIndex();
3590 Bytes = Flags.getByValSize();
3594 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3595 if (Flags.isByVal())
3596 // ByVal argument is passed in as a pointer but it's now being
3597 // dereferenced. e.g.
3598 // define @foo(%struct.X* %A) {
3599 // tail call @bar(%struct.X* byval %A)
3602 SDValue Ptr = Ld->getBasePtr();
3603 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3606 FI = FINode->getIndex();
3607 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3608 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3609 FI = FINode->getIndex();
3610 Bytes = Flags.getByValSize();
3614 assert(FI != INT_MAX);
3615 if (!MFI->isFixedObjectIndex(FI))
3617 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3620 /// Check whether the call is eligible for tail call optimization. Targets
3621 /// that want to do tail call optimization should implement this function.
3622 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3623 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3624 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3625 const SmallVectorImpl<ISD::OutputArg> &Outs,
3626 const SmallVectorImpl<SDValue> &OutVals,
3627 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3628 if (!mayTailCallThisCC(CalleeCC))
3631 // If -tailcallopt is specified, make fastcc functions tail-callable.
3632 MachineFunction &MF = DAG.getMachineFunction();
3633 const Function *CallerF = MF.getFunction();
3635 // If the function return type is x86_fp80 and the callee return type is not,
3636 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3637 // perform a tailcall optimization here.
3638 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3641 CallingConv::ID CallerCC = CallerF->getCallingConv();
3642 bool CCMatch = CallerCC == CalleeCC;
3643 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3644 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3646 // Win64 functions have extra shadow space for argument homing. Don't do the
3647 // sibcall if the caller and callee have mismatched expectations for this
3649 if (IsCalleeWin64 != IsCallerWin64)
3652 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3653 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3658 // Look for obvious safe cases to perform tail call optimization that do not
3659 // require ABI changes. This is what gcc calls sibcall.
3661 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3662 // emit a special epilogue.
3663 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3664 if (RegInfo->needsStackRealignment(MF))
3667 // Also avoid sibcall optimization if either caller or callee uses struct
3668 // return semantics.
3669 if (isCalleeStructRet || isCallerStructRet)
3672 // Do not sibcall optimize vararg calls unless all arguments are passed via
3674 if (isVarArg && !Outs.empty()) {
3675 // Optimizing for varargs on Win64 is unlikely to be safe without
3676 // additional testing.
3677 if (IsCalleeWin64 || IsCallerWin64)
3680 SmallVector<CCValAssign, 16> ArgLocs;
3681 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3684 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3685 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3686 if (!ArgLocs[i].isRegLoc())
3690 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3691 // stack. Therefore, if it's not used by the call it is not safe to optimize
3692 // this into a sibcall.
3693 bool Unused = false;
3694 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3701 SmallVector<CCValAssign, 16> RVLocs;
3702 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3704 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3705 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3706 CCValAssign &VA = RVLocs[i];
3707 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3712 // If the calling conventions do not match, then we'd better make sure the
3713 // results are returned in the same way as what the caller expects.
3715 SmallVector<CCValAssign, 16> RVLocs1;
3716 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3718 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3720 SmallVector<CCValAssign, 16> RVLocs2;
3721 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3723 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3725 if (RVLocs1.size() != RVLocs2.size())
3727 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3728 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3730 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3732 if (RVLocs1[i].isRegLoc()) {
3733 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3736 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3742 unsigned StackArgsSize = 0;
3744 // If the callee takes no arguments then go on to check the results of the
3746 if (!Outs.empty()) {
3747 // Check if stack adjustment is needed. For now, do not do this if any
3748 // argument is passed on the stack.
3749 SmallVector<CCValAssign, 16> ArgLocs;
3750 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3753 // Allocate shadow area for Win64
3755 CCInfo.AllocateStack(32, 8);
3757 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3758 StackArgsSize = CCInfo.getNextStackOffset();
3760 if (CCInfo.getNextStackOffset()) {
3761 // Check if the arguments are already laid out in the right way as
3762 // the caller's fixed stack objects.
3763 MachineFrameInfo *MFI = MF.getFrameInfo();
3764 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3765 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3766 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3767 CCValAssign &VA = ArgLocs[i];
3768 SDValue Arg = OutVals[i];
3769 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3770 if (VA.getLocInfo() == CCValAssign::Indirect)
3772 if (!VA.isRegLoc()) {
3773 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3780 // If the tailcall address may be in a register, then make sure it's
3781 // possible to register allocate for it. In 32-bit, the call address can
3782 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3783 // callee-saved registers are restored. These happen to be the same
3784 // registers used to pass 'inreg' arguments so watch out for those.
3785 if (!Subtarget->is64Bit() &&
3786 ((!isa<GlobalAddressSDNode>(Callee) &&
3787 !isa<ExternalSymbolSDNode>(Callee)) ||
3788 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3789 unsigned NumInRegs = 0;
3790 // In PIC we need an extra register to formulate the address computation
3792 unsigned MaxInRegs =
3793 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3795 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3796 CCValAssign &VA = ArgLocs[i];
3799 unsigned Reg = VA.getLocReg();
3802 case X86::EAX: case X86::EDX: case X86::ECX:
3803 if (++NumInRegs == MaxInRegs)
3811 bool CalleeWillPop =
3812 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3813 MF.getTarget().Options.GuaranteedTailCallOpt);
3815 if (unsigned BytesToPop =
3816 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3817 // If we have bytes to pop, the callee must pop them.
3818 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3819 if (!CalleePopMatches)
3821 } else if (CalleeWillPop && StackArgsSize > 0) {
3822 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3830 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3831 const TargetLibraryInfo *libInfo) const {
3832 return X86::createFastISel(funcInfo, libInfo);
3835 //===----------------------------------------------------------------------===//
3836 // Other Lowering Hooks
3837 //===----------------------------------------------------------------------===//
3839 static bool MayFoldLoad(SDValue Op) {
3840 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3843 static bool MayFoldIntoStore(SDValue Op) {
3844 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3847 static bool isTargetShuffle(unsigned Opcode) {
3849 default: return false;
3850 case X86ISD::BLENDI:
3851 case X86ISD::PSHUFB:
3852 case X86ISD::PSHUFD:
3853 case X86ISD::PSHUFHW:
3854 case X86ISD::PSHUFLW:
3856 case X86ISD::PALIGNR:
3857 case X86ISD::MOVLHPS:
3858 case X86ISD::MOVLHPD:
3859 case X86ISD::MOVHLPS:
3860 case X86ISD::MOVLPS:
3861 case X86ISD::MOVLPD:
3862 case X86ISD::MOVSHDUP:
3863 case X86ISD::MOVSLDUP:
3864 case X86ISD::MOVDDUP:
3867 case X86ISD::UNPCKL:
3868 case X86ISD::UNPCKH:
3869 case X86ISD::VPERMILPI:
3870 case X86ISD::VPERM2X128:
3871 case X86ISD::VPERMI:
3872 case X86ISD::VPERMV:
3873 case X86ISD::VPERMV3:
3878 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3879 SDValue V1, unsigned TargetMask,
3880 SelectionDAG &DAG) {
3882 default: llvm_unreachable("Unknown x86 shuffle node");
3883 case X86ISD::PSHUFD:
3884 case X86ISD::PSHUFHW:
3885 case X86ISD::PSHUFLW:
3886 case X86ISD::VPERMILPI:
3887 case X86ISD::VPERMI:
3888 return DAG.getNode(Opc, dl, VT, V1,
3889 DAG.getConstant(TargetMask, dl, MVT::i8));
3893 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3894 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3896 default: llvm_unreachable("Unknown x86 shuffle node");
3897 case X86ISD::MOVLHPS:
3898 case X86ISD::MOVLHPD:
3899 case X86ISD::MOVHLPS:
3900 case X86ISD::MOVLPS:
3901 case X86ISD::MOVLPD:
3904 case X86ISD::UNPCKL:
3905 case X86ISD::UNPCKH:
3906 return DAG.getNode(Opc, dl, VT, V1, V2);
3910 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3911 MachineFunction &MF = DAG.getMachineFunction();
3912 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3913 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3914 int ReturnAddrIndex = FuncInfo->getRAIndex();
3916 if (ReturnAddrIndex == 0) {
3917 // Set up a frame object for the return address.
3918 unsigned SlotSize = RegInfo->getSlotSize();
3919 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3922 FuncInfo->setRAIndex(ReturnAddrIndex);
3925 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3928 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3929 bool hasSymbolicDisplacement) {
3930 // Offset should fit into 32 bit immediate field.
3931 if (!isInt<32>(Offset))
3934 // If we don't have a symbolic displacement - we don't have any extra
3936 if (!hasSymbolicDisplacement)
3939 // FIXME: Some tweaks might be needed for medium code model.
3940 if (M != CodeModel::Small && M != CodeModel::Kernel)
3943 // For small code model we assume that latest object is 16MB before end of 31
3944 // bits boundary. We may also accept pretty large negative constants knowing
3945 // that all objects are in the positive half of address space.
3946 if (M == CodeModel::Small && Offset < 16*1024*1024)
3949 // For kernel code model we know that all object resist in the negative half
3950 // of 32bits address space. We may not accept negative offsets, since they may
3951 // be just off and we may accept pretty large positive ones.
3952 if (M == CodeModel::Kernel && Offset >= 0)
3958 /// Determines whether the callee is required to pop its own arguments.
3959 /// Callee pop is necessary to support tail calls.
3960 bool X86::isCalleePop(CallingConv::ID CallingConv,
3961 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3962 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3963 // can guarantee TCO.
3964 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3967 switch (CallingConv) {
3970 case CallingConv::X86_StdCall:
3971 case CallingConv::X86_FastCall:
3972 case CallingConv::X86_ThisCall:
3973 case CallingConv::X86_VectorCall:
3978 /// \brief Return true if the condition is an unsigned comparison operation.
3979 static bool isX86CCUnsigned(unsigned X86CC) {
3981 default: llvm_unreachable("Invalid integer condition!");
3982 case X86::COND_E: return true;
3983 case X86::COND_G: return false;
3984 case X86::COND_GE: return false;
3985 case X86::COND_L: return false;
3986 case X86::COND_LE: return false;
3987 case X86::COND_NE: return true;
3988 case X86::COND_B: return true;
3989 case X86::COND_A: return true;
3990 case X86::COND_BE: return true;
3991 case X86::COND_AE: return true;
3995 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
3996 switch (SetCCOpcode) {
3997 default: llvm_unreachable("Invalid integer condition!");
3998 case ISD::SETEQ: return X86::COND_E;
3999 case ISD::SETGT: return X86::COND_G;
4000 case ISD::SETGE: return X86::COND_GE;
4001 case ISD::SETLT: return X86::COND_L;
4002 case ISD::SETLE: return X86::COND_LE;
4003 case ISD::SETNE: return X86::COND_NE;
4004 case ISD::SETULT: return X86::COND_B;
4005 case ISD::SETUGT: return X86::COND_A;
4006 case ISD::SETULE: return X86::COND_BE;
4007 case ISD::SETUGE: return X86::COND_AE;
4011 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
4012 /// condition code, returning the condition code and the LHS/RHS of the
4013 /// comparison to make.
4014 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4015 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4017 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4018 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4019 // X > -1 -> X == 0, jump !sign.
4020 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4021 return X86::COND_NS;
4023 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4024 // X < 0 -> X == 0, jump on sign.
4027 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4029 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4030 return X86::COND_LE;
4034 return TranslateIntegerX86CC(SetCCOpcode);
4037 // First determine if it is required or is profitable to flip the operands.
4039 // If LHS is a foldable load, but RHS is not, flip the condition.
4040 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4041 !ISD::isNON_EXTLoad(RHS.getNode())) {
4042 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4043 std::swap(LHS, RHS);
4046 switch (SetCCOpcode) {
4052 std::swap(LHS, RHS);
4056 // On a floating point condition, the flags are set as follows:
4058 // 0 | 0 | 0 | X > Y
4059 // 0 | 0 | 1 | X < Y
4060 // 1 | 0 | 0 | X == Y
4061 // 1 | 1 | 1 | unordered
4062 switch (SetCCOpcode) {
4063 default: llvm_unreachable("Condcode should be pre-legalized away");
4065 case ISD::SETEQ: return X86::COND_E;
4066 case ISD::SETOLT: // flipped
4068 case ISD::SETGT: return X86::COND_A;
4069 case ISD::SETOLE: // flipped
4071 case ISD::SETGE: return X86::COND_AE;
4072 case ISD::SETUGT: // flipped
4074 case ISD::SETLT: return X86::COND_B;
4075 case ISD::SETUGE: // flipped
4077 case ISD::SETLE: return X86::COND_BE;
4079 case ISD::SETNE: return X86::COND_NE;
4080 case ISD::SETUO: return X86::COND_P;
4081 case ISD::SETO: return X86::COND_NP;
4083 case ISD::SETUNE: return X86::COND_INVALID;
4087 /// Is there a floating point cmov for the specific X86 condition code?
4088 /// Current x86 isa includes the following FP cmov instructions:
4089 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4090 static bool hasFPCMov(unsigned X86CC) {
4106 /// Returns true if the target can instruction select the
4107 /// specified FP immediate natively. If false, the legalizer will
4108 /// materialize the FP immediate as a load from a constant pool.
4109 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4110 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4111 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4117 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4118 ISD::LoadExtType ExtTy,
4120 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4121 // relocation target a movq or addq instruction: don't let the load shrink.
4122 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4123 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4124 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4125 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4129 /// \brief Returns true if it is beneficial to convert a load of a constant
4130 /// to just the constant itself.
4131 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4133 assert(Ty->isIntegerTy());
4135 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4136 if (BitSize == 0 || BitSize > 64)
4141 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4142 unsigned Index) const {
4143 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4146 return (Index == 0 || Index == ResVT.getVectorNumElements());
4149 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4150 // Speculate cttz only if we can directly use TZCNT.
4151 return Subtarget->hasBMI();
4154 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4155 // Speculate ctlz only if we can directly use LZCNT.
4156 return Subtarget->hasLZCNT();
4159 /// Return true if every element in Mask, beginning
4160 /// from position Pos and ending in Pos+Size is undef.
4161 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4162 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4168 /// Return true if Val is undef or if its value falls within the
4169 /// specified range (L, H].
4170 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4171 return (Val < 0) || (Val >= Low && Val < Hi);
4174 /// Val is either less than zero (undef) or equal to the specified value.
4175 static bool isUndefOrEqual(int Val, int CmpVal) {
4176 return (Val < 0 || Val == CmpVal);
4179 /// Return true if every element in Mask, beginning
4180 /// from position Pos and ending in Pos+Size, falls within the specified
4181 /// sequential range (Low, Low+Size]. or is undef.
4182 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4183 unsigned Pos, unsigned Size, int Low) {
4184 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4185 if (!isUndefOrEqual(Mask[i], Low))
4190 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4191 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4192 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4193 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4194 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4197 // The index should be aligned on a vecWidth-bit boundary.
4199 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4201 MVT VT = N->getSimpleValueType(0);
4202 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4203 bool Result = (Index * ElSize) % vecWidth == 0;
4208 /// Return true if the specified INSERT_SUBVECTOR
4209 /// operand specifies a subvector insert that is suitable for input to
4210 /// insertion of 128 or 256-bit subvectors
4211 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4212 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4213 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4215 // The index should be aligned on a vecWidth-bit boundary.
4217 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4219 MVT VT = N->getSimpleValueType(0);
4220 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4221 bool Result = (Index * ElSize) % vecWidth == 0;
4226 bool X86::isVINSERT128Index(SDNode *N) {
4227 return isVINSERTIndex(N, 128);
4230 bool X86::isVINSERT256Index(SDNode *N) {
4231 return isVINSERTIndex(N, 256);
4234 bool X86::isVEXTRACT128Index(SDNode *N) {
4235 return isVEXTRACTIndex(N, 128);
4238 bool X86::isVEXTRACT256Index(SDNode *N) {
4239 return isVEXTRACTIndex(N, 256);
4242 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4243 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4244 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4245 "Illegal extract subvector for VEXTRACT");
4248 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4250 MVT VecVT = N->getOperand(0).getSimpleValueType();
4251 MVT ElVT = VecVT.getVectorElementType();
4253 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4254 return Index / NumElemsPerChunk;
4257 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4258 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4259 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4260 "Illegal insert subvector for VINSERT");
4263 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4265 MVT VecVT = N->getSimpleValueType(0);
4266 MVT ElVT = VecVT.getVectorElementType();
4268 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4269 return Index / NumElemsPerChunk;
4272 /// Return the appropriate immediate to extract the specified
4273 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4274 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4275 return getExtractVEXTRACTImmediate(N, 128);
4278 /// Return the appropriate immediate to extract the specified
4279 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4280 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4281 return getExtractVEXTRACTImmediate(N, 256);
4284 /// Return the appropriate immediate to insert at the specified
4285 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4286 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4287 return getInsertVINSERTImmediate(N, 128);
4290 /// Return the appropriate immediate to insert at the specified
4291 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4292 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4293 return getInsertVINSERTImmediate(N, 256);
4296 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4297 bool X86::isZeroNode(SDValue Elt) {
4298 return isNullConstant(Elt) || isNullFPConstant(Elt);
4301 // Build a vector of constants
4302 // Use an UNDEF node if MaskElt == -1.
4303 // Spilt 64-bit constants in the 32-bit mode.
4304 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4306 SDLoc dl, bool IsMask = false) {
4308 SmallVector<SDValue, 32> Ops;
4311 MVT ConstVecVT = VT;
4312 unsigned NumElts = VT.getVectorNumElements();
4313 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4314 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4315 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4319 MVT EltVT = ConstVecVT.getVectorElementType();
4320 for (unsigned i = 0; i < NumElts; ++i) {
4321 bool IsUndef = Values[i] < 0 && IsMask;
4322 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4323 DAG.getConstant(Values[i], dl, EltVT);
4324 Ops.push_back(OpNode);
4326 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4327 DAG.getConstant(0, dl, EltVT));
4329 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4331 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4335 /// Returns a vector of specified type with all zero elements.
4336 static SDValue getZeroVector(MVT VT, const X86Subtarget *Subtarget,
4337 SelectionDAG &DAG, SDLoc dl) {
4338 assert(VT.isVector() && "Expected a vector type");
4340 // Always build SSE zero vectors as <4 x i32> bitcasted
4341 // to their dest type. This ensures they get CSE'd.
4343 if (VT.is128BitVector()) { // SSE
4344 if (Subtarget->hasSSE2()) { // SSE2
4345 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4346 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4348 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4349 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4351 } else if (VT.is256BitVector()) { // AVX
4352 if (Subtarget->hasInt256()) { // AVX2
4353 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4354 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4355 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4357 // 256-bit logic and arithmetic instructions in AVX are all
4358 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4359 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4360 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4361 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4363 } else if (VT.is512BitVector()) { // AVX-512
4364 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4365 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4366 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4367 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4368 } else if (VT.getVectorElementType() == MVT::i1) {
4370 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4371 && "Unexpected vector type");
4372 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4373 && "Unexpected vector type");
4374 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4375 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4376 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4378 llvm_unreachable("Unexpected vector type");
4380 return DAG.getBitcast(VT, Vec);
4383 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4384 SelectionDAG &DAG, SDLoc dl,
4385 unsigned vectorWidth) {
4386 assert((vectorWidth == 128 || vectorWidth == 256) &&
4387 "Unsupported vector width");
4388 EVT VT = Vec.getValueType();
4389 EVT ElVT = VT.getVectorElementType();
4390 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4391 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4392 VT.getVectorNumElements()/Factor);
4394 // Extract from UNDEF is UNDEF.
4395 if (Vec.getOpcode() == ISD::UNDEF)
4396 return DAG.getUNDEF(ResultVT);
4398 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4399 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4400 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4402 // This is the index of the first element of the vectorWidth-bit chunk
4403 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4404 IdxVal &= ~(ElemsPerChunk - 1);
4406 // If the input is a buildvector just emit a smaller one.
4407 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4408 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4409 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4411 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4412 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4415 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4416 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4417 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4418 /// instructions or a simple subregister reference. Idx is an index in the
4419 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4420 /// lowering EXTRACT_VECTOR_ELT operations easier.
4421 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4422 SelectionDAG &DAG, SDLoc dl) {
4423 assert((Vec.getValueType().is256BitVector() ||
4424 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4425 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4428 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4429 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4430 SelectionDAG &DAG, SDLoc dl) {
4431 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4432 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4435 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4436 unsigned IdxVal, SelectionDAG &DAG,
4437 SDLoc dl, unsigned vectorWidth) {
4438 assert((vectorWidth == 128 || vectorWidth == 256) &&
4439 "Unsupported vector width");
4440 // Inserting UNDEF is Result
4441 if (Vec.getOpcode() == ISD::UNDEF)
4443 EVT VT = Vec.getValueType();
4444 EVT ElVT = VT.getVectorElementType();
4445 EVT ResultVT = Result.getValueType();
4447 // Insert the relevant vectorWidth bits.
4448 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4449 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4451 // This is the index of the first element of the vectorWidth-bit chunk
4452 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4453 IdxVal &= ~(ElemsPerChunk - 1);
4455 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4456 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4459 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4460 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4461 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4462 /// simple superregister reference. Idx is an index in the 128 bits
4463 /// we want. It need not be aligned to a 128-bit boundary. That makes
4464 /// lowering INSERT_VECTOR_ELT operations easier.
4465 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4466 SelectionDAG &DAG, SDLoc dl) {
4467 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4469 // For insertion into the zero index (low half) of a 256-bit vector, it is
4470 // more efficient to generate a blend with immediate instead of an insert*128.
4471 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4472 // extend the subvector to the size of the result vector. Make sure that
4473 // we are not recursing on that node by checking for undef here.
4474 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4475 Result.getOpcode() != ISD::UNDEF) {
4476 EVT ResultVT = Result.getValueType();
4477 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4478 SDValue Undef = DAG.getUNDEF(ResultVT);
4479 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4482 // The blend instruction, and therefore its mask, depend on the data type.
4483 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4484 if (ScalarType.isFloatingPoint()) {
4485 // Choose either vblendps (float) or vblendpd (double).
4486 unsigned ScalarSize = ScalarType.getSizeInBits();
4487 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4488 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4489 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4490 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4493 const X86Subtarget &Subtarget =
4494 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4496 // AVX2 is needed for 256-bit integer blend support.
4497 // Integers must be cast to 32-bit because there is only vpblendd;
4498 // vpblendw can't be used for this because it has a handicapped mask.
4500 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4501 // is still more efficient than using the wrong domain vinsertf128 that
4502 // will be created by InsertSubVector().
4503 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4505 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4506 Vec256 = DAG.getBitcast(CastVT, Vec256);
4507 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4508 return DAG.getBitcast(ResultVT, Vec256);
4511 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4514 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4515 SelectionDAG &DAG, SDLoc dl) {
4516 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4517 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4520 /// Insert i1-subvector to i1-vector.
4521 static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
4524 SDValue Vec = Op.getOperand(0);
4525 SDValue SubVec = Op.getOperand(1);
4526 SDValue Idx = Op.getOperand(2);
4528 if (!isa<ConstantSDNode>(Idx))
4531 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
4532 if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
4535 MVT OpVT = Op.getSimpleValueType();
4536 MVT SubVecVT = SubVec.getSimpleValueType();
4537 unsigned NumElems = OpVT.getVectorNumElements();
4538 unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
4540 assert(IdxVal + SubVecNumElems <= NumElems &&
4541 IdxVal % SubVecVT.getSizeInBits() == 0 &&
4542 "Unexpected index value in INSERT_SUBVECTOR");
4544 // There are 3 possible cases:
4545 // 1. Subvector should be inserted in the lower part (IdxVal == 0)
4546 // 2. Subvector should be inserted in the upper part
4547 // (IdxVal + SubVecNumElems == NumElems)
4548 // 3. Subvector should be inserted in the middle (for example v2i1
4549 // to v16i1, index 2)
4551 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
4552 SDValue Undef = DAG.getUNDEF(OpVT);
4553 SDValue WideSubVec =
4554 DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
4556 return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4557 DAG.getConstant(IdxVal, dl, MVT::i8));
4559 if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
4560 unsigned ShiftLeft = NumElems - SubVecNumElems;
4561 unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
4562 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4563 DAG.getConstant(ShiftLeft, dl, MVT::i8));
4564 return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
4565 DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
4569 // Zero lower bits of the Vec
4570 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4571 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4572 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4573 // Merge them together
4574 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4577 // Simple case when we put subvector in the upper part
4578 if (IdxVal + SubVecNumElems == NumElems) {
4579 // Zero upper bits of the Vec
4580 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
4581 DAG.getConstant(IdxVal, dl, MVT::i8));
4582 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4583 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4584 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4585 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4587 // Subvector should be inserted in the middle - use shuffle
4588 SmallVector<int, 64> Mask;
4589 for (unsigned i = 0; i < NumElems; ++i)
4590 Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
4592 return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
4595 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4596 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4597 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4598 /// large BUILD_VECTORS.
4599 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4600 unsigned NumElems, SelectionDAG &DAG,
4602 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4603 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4606 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4607 unsigned NumElems, SelectionDAG &DAG,
4609 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4610 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4613 /// Returns a vector of specified type with all bits set.
4614 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4615 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4616 /// Then bitcast to their original type, ensuring they get CSE'd.
4617 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4618 SelectionDAG &DAG, SDLoc dl) {
4619 assert(VT.isVector() && "Expected a vector type");
4621 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4623 if (VT.is512BitVector()) {
4624 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4625 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4626 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4627 } else if (VT.is256BitVector()) {
4628 if (Subtarget->hasInt256()) { // AVX2
4629 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4630 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4632 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4633 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4635 } else if (VT.is128BitVector()) {
4636 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4638 llvm_unreachable("Unexpected vector type");
4640 return DAG.getBitcast(VT, Vec);
4643 /// Returns a vector_shuffle node for an unpackl operation.
4644 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4646 unsigned NumElems = VT.getVectorNumElements();
4647 SmallVector<int, 8> Mask;
4648 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4650 Mask.push_back(i + NumElems);
4652 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4655 /// Returns a vector_shuffle node for an unpackh operation.
4656 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4658 unsigned NumElems = VT.getVectorNumElements();
4659 SmallVector<int, 8> Mask;
4660 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4661 Mask.push_back(i + Half);
4662 Mask.push_back(i + NumElems + Half);
4664 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4667 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4668 /// This produces a shuffle where the low element of V2 is swizzled into the
4669 /// zero/undef vector, landing at element Idx.
4670 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4671 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4673 const X86Subtarget *Subtarget,
4674 SelectionDAG &DAG) {
4675 MVT VT = V2.getSimpleValueType();
4677 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4678 unsigned NumElems = VT.getVectorNumElements();
4679 SmallVector<int, 16> MaskVec;
4680 for (unsigned i = 0; i != NumElems; ++i)
4681 // If this is the insertion idx, put the low elt of V2 here.
4682 MaskVec.push_back(i == Idx ? NumElems : i);
4683 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4686 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4687 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4688 /// uses one source. Note that this will set IsUnary for shuffles which use a
4689 /// single input multiple times, and in those cases it will
4690 /// adjust the mask to only have indices within that single input.
4691 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4692 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4693 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4694 unsigned NumElems = VT.getVectorNumElements();
4698 bool IsFakeUnary = false;
4699 switch(N->getOpcode()) {
4700 case X86ISD::BLENDI:
4701 ImmN = N->getOperand(N->getNumOperands()-1);
4702 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4705 ImmN = N->getOperand(N->getNumOperands()-1);
4706 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4707 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4709 case X86ISD::UNPCKH:
4710 DecodeUNPCKHMask(VT, Mask);
4711 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4713 case X86ISD::UNPCKL:
4714 DecodeUNPCKLMask(VT, Mask);
4715 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4717 case X86ISD::MOVHLPS:
4718 DecodeMOVHLPSMask(NumElems, Mask);
4719 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4721 case X86ISD::MOVLHPS:
4722 DecodeMOVLHPSMask(NumElems, Mask);
4723 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4725 case X86ISD::PALIGNR:
4726 ImmN = N->getOperand(N->getNumOperands()-1);
4727 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4729 case X86ISD::PSHUFD:
4730 case X86ISD::VPERMILPI:
4731 ImmN = N->getOperand(N->getNumOperands()-1);
4732 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4735 case X86ISD::PSHUFHW:
4736 ImmN = N->getOperand(N->getNumOperands()-1);
4737 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4740 case X86ISD::PSHUFLW:
4741 ImmN = N->getOperand(N->getNumOperands()-1);
4742 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4745 case X86ISD::PSHUFB: {
4747 SDValue MaskNode = N->getOperand(1);
4748 while (MaskNode->getOpcode() == ISD::BITCAST)
4749 MaskNode = MaskNode->getOperand(0);
4751 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4752 // If we have a build-vector, then things are easy.
4753 MVT VT = MaskNode.getSimpleValueType();
4754 assert(VT.isVector() &&
4755 "Can't produce a non-vector with a build_vector!");
4756 if (!VT.isInteger())
4759 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4761 SmallVector<uint64_t, 32> RawMask;
4762 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4763 SDValue Op = MaskNode->getOperand(i);
4764 if (Op->getOpcode() == ISD::UNDEF) {
4765 RawMask.push_back((uint64_t)SM_SentinelUndef);
4768 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4771 APInt MaskElement = CN->getAPIntValue();
4773 // We now have to decode the element which could be any integer size and
4774 // extract each byte of it.
4775 for (int j = 0; j < NumBytesPerElement; ++j) {
4776 // Note that this is x86 and so always little endian: the low byte is
4777 // the first byte of the mask.
4778 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4779 MaskElement = MaskElement.lshr(8);
4782 DecodePSHUFBMask(RawMask, Mask);
4786 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4790 SDValue Ptr = MaskLoad->getBasePtr();
4791 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4792 Ptr->getOpcode() == X86ISD::WrapperRIP)
4793 Ptr = Ptr->getOperand(0);
4795 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4796 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4799 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4800 DecodePSHUFBMask(C, Mask);
4808 case X86ISD::VPERMI:
4809 ImmN = N->getOperand(N->getNumOperands()-1);
4810 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4815 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4817 case X86ISD::VPERM2X128:
4818 ImmN = N->getOperand(N->getNumOperands()-1);
4819 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4820 if (Mask.empty()) return false;
4821 // Mask only contains negative index if an element is zero.
4822 if (std::any_of(Mask.begin(), Mask.end(),
4823 [](int M){ return M == SM_SentinelZero; }))
4826 case X86ISD::MOVSLDUP:
4827 DecodeMOVSLDUPMask(VT, Mask);
4830 case X86ISD::MOVSHDUP:
4831 DecodeMOVSHDUPMask(VT, Mask);
4834 case X86ISD::MOVDDUP:
4835 DecodeMOVDDUPMask(VT, Mask);
4838 case X86ISD::MOVLHPD:
4839 case X86ISD::MOVLPD:
4840 case X86ISD::MOVLPS:
4841 // Not yet implemented
4843 case X86ISD::VPERMV: {
4845 SDValue MaskNode = N->getOperand(0);
4846 while (MaskNode->getOpcode() == ISD::BITCAST)
4847 MaskNode = MaskNode->getOperand(0);
4849 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4850 SmallVector<uint64_t, 32> RawMask;
4851 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4852 // If we have a build-vector, then things are easy.
4853 assert(MaskNode.getSimpleValueType().isInteger() &&
4854 MaskNode.getSimpleValueType().getVectorNumElements() ==
4855 VT.getVectorNumElements());
4857 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4858 SDValue Op = MaskNode->getOperand(i);
4859 if (Op->getOpcode() == ISD::UNDEF)
4860 RawMask.push_back((uint64_t)SM_SentinelUndef);
4861 else if (isa<ConstantSDNode>(Op)) {
4862 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4863 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4867 DecodeVPERMVMask(RawMask, Mask);
4870 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4871 unsigned NumEltsInMask = MaskNode->getNumOperands();
4872 MaskNode = MaskNode->getOperand(0);
4873 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4875 APInt MaskEltValue = CN->getAPIntValue();
4876 for (unsigned i = 0; i < NumEltsInMask; ++i)
4877 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4878 DecodeVPERMVMask(RawMask, Mask);
4881 // It may be a scalar load
4884 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4888 SDValue Ptr = MaskLoad->getBasePtr();
4889 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4890 Ptr->getOpcode() == X86ISD::WrapperRIP)
4891 Ptr = Ptr->getOperand(0);
4893 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4894 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4897 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4899 DecodeVPERMVMask(C, VT, Mask);
4906 case X86ISD::VPERMV3: {
4908 SDValue MaskNode = N->getOperand(1);
4909 while (MaskNode->getOpcode() == ISD::BITCAST)
4910 MaskNode = MaskNode->getOperand(1);
4912 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4913 // If we have a build-vector, then things are easy.
4914 assert(MaskNode.getSimpleValueType().isInteger() &&
4915 MaskNode.getSimpleValueType().getVectorNumElements() ==
4916 VT.getVectorNumElements());
4918 SmallVector<uint64_t, 32> RawMask;
4919 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4921 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4922 SDValue Op = MaskNode->getOperand(i);
4923 if (Op->getOpcode() == ISD::UNDEF)
4924 RawMask.push_back((uint64_t)SM_SentinelUndef);
4926 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4929 APInt MaskElement = CN->getAPIntValue();
4930 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4933 DecodeVPERMV3Mask(RawMask, Mask);
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 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4952 DecodeVPERMV3Mask(C, VT, Mask);
4959 default: llvm_unreachable("unknown target shuffle node");
4962 // If we have a fake unary shuffle, the shuffle mask is spread across two
4963 // inputs that are actually the same node. Re-map the mask to always point
4964 // into the first input.
4967 if (M >= (int)Mask.size())
4973 /// Returns the scalar element that will make up the ith
4974 /// element of the result of the vector shuffle.
4975 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4978 return SDValue(); // Limit search depth.
4980 SDValue V = SDValue(N, 0);
4981 EVT VT = V.getValueType();
4982 unsigned Opcode = V.getOpcode();
4984 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4985 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4986 int Elt = SV->getMaskElt(Index);
4989 return DAG.getUNDEF(VT.getVectorElementType());
4991 unsigned NumElems = VT.getVectorNumElements();
4992 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4993 : SV->getOperand(1);
4994 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4997 // Recurse into target specific vector shuffles to find scalars.
4998 if (isTargetShuffle(Opcode)) {
4999 MVT ShufVT = V.getSimpleValueType();
5000 unsigned NumElems = ShufVT.getVectorNumElements();
5001 SmallVector<int, 16> ShuffleMask;
5004 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5007 int Elt = ShuffleMask[Index];
5009 return DAG.getUNDEF(ShufVT.getVectorElementType());
5011 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5013 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5017 // Actual nodes that may contain scalar elements
5018 if (Opcode == ISD::BITCAST) {
5019 V = V.getOperand(0);
5020 EVT SrcVT = V.getValueType();
5021 unsigned NumElems = VT.getVectorNumElements();
5023 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5027 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5028 return (Index == 0) ? V.getOperand(0)
5029 : DAG.getUNDEF(VT.getVectorElementType());
5031 if (V.getOpcode() == ISD::BUILD_VECTOR)
5032 return V.getOperand(Index);
5037 /// Custom lower build_vector of v16i8.
5038 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5039 unsigned NumNonZero, unsigned NumZero,
5041 const X86Subtarget* Subtarget,
5042 const TargetLowering &TLI) {
5050 // SSE4.1 - use PINSRB to insert each byte directly.
5051 if (Subtarget->hasSSE41()) {
5052 for (unsigned i = 0; i < 16; ++i) {
5053 bool isNonZero = (NonZeros & (1 << i)) != 0;
5057 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
5059 V = DAG.getUNDEF(MVT::v16i8);
5062 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5063 MVT::v16i8, V, Op.getOperand(i),
5064 DAG.getIntPtrConstant(i, dl));
5071 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5072 for (unsigned i = 0; i < 16; ++i) {
5073 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5074 if (ThisIsNonZero && First) {
5076 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5078 V = DAG.getUNDEF(MVT::v8i16);
5083 SDValue ThisElt, LastElt;
5084 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5085 if (LastIsNonZero) {
5086 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5087 MVT::i16, Op.getOperand(i-1));
5089 if (ThisIsNonZero) {
5090 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5091 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5092 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5094 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5098 if (ThisElt.getNode())
5099 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5100 DAG.getIntPtrConstant(i/2, dl));
5104 return DAG.getBitcast(MVT::v16i8, V);
5107 /// Custom lower build_vector of v8i16.
5108 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5109 unsigned NumNonZero, unsigned NumZero,
5111 const X86Subtarget* Subtarget,
5112 const TargetLowering &TLI) {
5119 for (unsigned i = 0; i < 8; ++i) {
5120 bool isNonZero = (NonZeros & (1 << i)) != 0;
5124 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5126 V = DAG.getUNDEF(MVT::v8i16);
5129 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5130 MVT::v8i16, V, Op.getOperand(i),
5131 DAG.getIntPtrConstant(i, dl));
5138 /// Custom lower build_vector of v4i32 or v4f32.
5139 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5140 const X86Subtarget *Subtarget,
5141 const TargetLowering &TLI) {
5142 // Find all zeroable elements.
5143 std::bitset<4> Zeroable;
5144 for (int i=0; i < 4; ++i) {
5145 SDValue Elt = Op->getOperand(i);
5146 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5148 assert(Zeroable.size() - Zeroable.count() > 1 &&
5149 "We expect at least two non-zero elements!");
5151 // We only know how to deal with build_vector nodes where elements are either
5152 // zeroable or extract_vector_elt with constant index.
5153 SDValue FirstNonZero;
5154 unsigned FirstNonZeroIdx;
5155 for (unsigned i=0; i < 4; ++i) {
5158 SDValue Elt = Op->getOperand(i);
5159 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5160 !isa<ConstantSDNode>(Elt.getOperand(1)))
5162 // Make sure that this node is extracting from a 128-bit vector.
5163 MVT VT = Elt.getOperand(0).getSimpleValueType();
5164 if (!VT.is128BitVector())
5166 if (!FirstNonZero.getNode()) {
5168 FirstNonZeroIdx = i;
5172 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5173 SDValue V1 = FirstNonZero.getOperand(0);
5174 MVT VT = V1.getSimpleValueType();
5176 // See if this build_vector can be lowered as a blend with zero.
5178 unsigned EltMaskIdx, EltIdx;
5180 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5181 if (Zeroable[EltIdx]) {
5182 // The zero vector will be on the right hand side.
5183 Mask[EltIdx] = EltIdx+4;
5187 Elt = Op->getOperand(EltIdx);
5188 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5189 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5190 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5192 Mask[EltIdx] = EltIdx;
5196 // Let the shuffle legalizer deal with blend operations.
5197 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5198 if (V1.getSimpleValueType() != VT)
5199 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5200 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5203 // See if we can lower this build_vector to a INSERTPS.
5204 if (!Subtarget->hasSSE41())
5207 SDValue V2 = Elt.getOperand(0);
5208 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5211 bool CanFold = true;
5212 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5216 SDValue Current = Op->getOperand(i);
5217 SDValue SrcVector = Current->getOperand(0);
5220 CanFold = SrcVector == V1 &&
5221 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5227 assert(V1.getNode() && "Expected at least two non-zero elements!");
5228 if (V1.getSimpleValueType() != MVT::v4f32)
5229 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5230 if (V2.getSimpleValueType() != MVT::v4f32)
5231 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5233 // Ok, we can emit an INSERTPS instruction.
5234 unsigned ZMask = Zeroable.to_ulong();
5236 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5237 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5239 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5240 DAG.getIntPtrConstant(InsertPSMask, DL));
5241 return DAG.getBitcast(VT, Result);
5244 /// Return a vector logical shift node.
5245 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5246 unsigned NumBits, SelectionDAG &DAG,
5247 const TargetLowering &TLI, SDLoc dl) {
5248 assert(VT.is128BitVector() && "Unknown type for VShift");
5249 MVT ShVT = MVT::v2i64;
5250 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5251 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5252 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5253 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5254 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5255 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5259 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5261 // Check if the scalar load can be widened into a vector load. And if
5262 // the address is "base + cst" see if the cst can be "absorbed" into
5263 // the shuffle mask.
5264 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5265 SDValue Ptr = LD->getBasePtr();
5266 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5268 EVT PVT = LD->getValueType(0);
5269 if (PVT != MVT::i32 && PVT != MVT::f32)
5274 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5275 FI = FINode->getIndex();
5277 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5278 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5279 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5280 Offset = Ptr.getConstantOperandVal(1);
5281 Ptr = Ptr.getOperand(0);
5286 // FIXME: 256-bit vector instructions don't require a strict alignment,
5287 // improve this code to support it better.
5288 unsigned RequiredAlign = VT.getSizeInBits()/8;
5289 SDValue Chain = LD->getChain();
5290 // Make sure the stack object alignment is at least 16 or 32.
5291 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5292 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5293 if (MFI->isFixedObjectIndex(FI)) {
5294 // Can't change the alignment. FIXME: It's possible to compute
5295 // the exact stack offset and reference FI + adjust offset instead.
5296 // If someone *really* cares about this. That's the way to implement it.
5299 MFI->setObjectAlignment(FI, RequiredAlign);
5303 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5304 // Ptr + (Offset & ~15).
5307 if ((Offset % RequiredAlign) & 3)
5309 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5312 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5313 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5316 int EltNo = (Offset - StartOffset) >> 2;
5317 unsigned NumElems = VT.getVectorNumElements();
5319 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5320 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5321 LD->getPointerInfo().getWithOffset(StartOffset),
5322 false, false, false, 0);
5324 SmallVector<int, 8> Mask(NumElems, EltNo);
5326 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5332 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5333 /// elements can be replaced by a single large load which has the same value as
5334 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5336 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5338 /// FIXME: we'd also like to handle the case where the last elements are zero
5339 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5340 /// There's even a handy isZeroNode for that purpose.
5341 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5342 SDLoc &DL, SelectionDAG &DAG,
5343 bool isAfterLegalize) {
5344 unsigned NumElems = Elts.size();
5346 LoadSDNode *LDBase = nullptr;
5347 unsigned LastLoadedElt = -1U;
5349 // For each element in the initializer, see if we've found a load or an undef.
5350 // If we don't find an initial load element, or later load elements are
5351 // non-consecutive, bail out.
5352 for (unsigned i = 0; i < NumElems; ++i) {
5353 SDValue Elt = Elts[i];
5354 // Look through a bitcast.
5355 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5356 Elt = Elt.getOperand(0);
5357 if (!Elt.getNode() ||
5358 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5361 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5363 LDBase = cast<LoadSDNode>(Elt.getNode());
5367 if (Elt.getOpcode() == ISD::UNDEF)
5370 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5371 EVT LdVT = Elt.getValueType();
5372 // Each loaded element must be the correct fractional portion of the
5373 // requested vector load.
5374 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5376 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5381 // If we have found an entire vector of loads and undefs, then return a large
5382 // load of the entire vector width starting at the base pointer. If we found
5383 // consecutive loads for the low half, generate a vzext_load node.
5384 if (LastLoadedElt == NumElems - 1) {
5385 assert(LDBase && "Did not find base load for merging consecutive loads");
5386 EVT EltVT = LDBase->getValueType(0);
5387 // Ensure that the input vector size for the merged loads matches the
5388 // cumulative size of the input elements.
5389 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5392 if (isAfterLegalize &&
5393 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5396 SDValue NewLd = SDValue();
5398 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5399 LDBase->getPointerInfo(), LDBase->isVolatile(),
5400 LDBase->isNonTemporal(), LDBase->isInvariant(),
5401 LDBase->getAlignment());
5403 if (LDBase->hasAnyUseOfValue(1)) {
5404 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5406 SDValue(NewLd.getNode(), 1));
5407 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5408 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5409 SDValue(NewLd.getNode(), 1));
5415 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5416 //of a v4i32 / v4f32. It's probably worth generalizing.
5417 EVT EltVT = VT.getVectorElementType();
5418 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5419 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5420 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5421 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5423 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5424 LDBase->getPointerInfo(),
5425 LDBase->getAlignment(),
5426 false/*isVolatile*/, true/*ReadMem*/,
5429 // Make sure the newly-created LOAD is in the same position as LDBase in
5430 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5431 // update uses of LDBase's output chain to use the TokenFactor.
5432 if (LDBase->hasAnyUseOfValue(1)) {
5433 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5434 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5435 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5436 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5437 SDValue(ResNode.getNode(), 1));
5440 return DAG.getBitcast(VT, ResNode);
5445 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5446 /// to generate a splat value for the following cases:
5447 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5448 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5449 /// a scalar load, or a constant.
5450 /// The VBROADCAST node is returned when a pattern is found,
5451 /// or SDValue() otherwise.
5452 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5453 SelectionDAG &DAG) {
5454 // VBROADCAST requires AVX.
5455 // TODO: Splats could be generated for non-AVX CPUs using SSE
5456 // instructions, but there's less potential gain for only 128-bit vectors.
5457 if (!Subtarget->hasAVX())
5460 MVT VT = Op.getSimpleValueType();
5463 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5464 "Unsupported vector type for broadcast.");
5469 switch (Op.getOpcode()) {
5471 // Unknown pattern found.
5474 case ISD::BUILD_VECTOR: {
5475 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5476 BitVector UndefElements;
5477 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5479 // We need a splat of a single value to use broadcast, and it doesn't
5480 // make any sense if the value is only in one element of the vector.
5481 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5485 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5486 Ld.getOpcode() == ISD::ConstantFP);
5488 // Make sure that all of the users of a non-constant load are from the
5489 // BUILD_VECTOR node.
5490 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5495 case ISD::VECTOR_SHUFFLE: {
5496 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5498 // Shuffles must have a splat mask where the first element is
5500 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5503 SDValue Sc = Op.getOperand(0);
5504 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5505 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5507 if (!Subtarget->hasInt256())
5510 // Use the register form of the broadcast instruction available on AVX2.
5511 if (VT.getSizeInBits() >= 256)
5512 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5513 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5516 Ld = Sc.getOperand(0);
5517 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5518 Ld.getOpcode() == ISD::ConstantFP);
5520 // The scalar_to_vector node and the suspected
5521 // load node must have exactly one user.
5522 // Constants may have multiple users.
5524 // AVX-512 has register version of the broadcast
5525 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5526 Ld.getValueType().getSizeInBits() >= 32;
5527 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5534 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5535 bool IsGE256 = (VT.getSizeInBits() >= 256);
5537 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5538 // instruction to save 8 or more bytes of constant pool data.
5539 // TODO: If multiple splats are generated to load the same constant,
5540 // it may be detrimental to overall size. There needs to be a way to detect
5541 // that condition to know if this is truly a size win.
5542 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5544 // Handle broadcasting a single constant scalar from the constant pool
5546 // On Sandybridge (no AVX2), it is still better to load a constant vector
5547 // from the constant pool and not to broadcast it from a scalar.
5548 // But override that restriction when optimizing for size.
5549 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5550 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5551 EVT CVT = Ld.getValueType();
5552 assert(!CVT.isVector() && "Must not broadcast a vector type");
5554 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5555 // For size optimization, also splat v2f64 and v2i64, and for size opt
5556 // with AVX2, also splat i8 and i16.
5557 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5558 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5559 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5560 const Constant *C = nullptr;
5561 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5562 C = CI->getConstantIntValue();
5563 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5564 C = CF->getConstantFPValue();
5566 assert(C && "Invalid constant type");
5568 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5570 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5571 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5573 CVT, dl, DAG.getEntryNode(), CP,
5574 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5575 false, false, Alignment);
5577 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5581 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5583 // Handle AVX2 in-register broadcasts.
5584 if (!IsLoad && Subtarget->hasInt256() &&
5585 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5586 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5588 // The scalar source must be a normal load.
5592 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5593 (Subtarget->hasVLX() && ScalarSize == 64))
5594 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5596 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5597 // double since there is no vbroadcastsd xmm
5598 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5599 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5600 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5603 // Unsupported broadcast.
5607 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5608 /// underlying vector and index.
5610 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5612 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5614 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5615 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5618 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5620 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5622 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5623 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5626 // In this case the vector is the extract_subvector expression and the index
5627 // is 2, as specified by the shuffle.
5628 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5629 SDValue ShuffleVec = SVOp->getOperand(0);
5630 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5631 assert(ShuffleVecVT.getVectorElementType() ==
5632 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5634 int ShuffleIdx = SVOp->getMaskElt(Idx);
5635 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5636 ExtractedFromVec = ShuffleVec;
5642 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5643 MVT VT = Op.getSimpleValueType();
5645 // Skip if insert_vec_elt is not supported.
5646 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5647 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5651 unsigned NumElems = Op.getNumOperands();
5655 SmallVector<unsigned, 4> InsertIndices;
5656 SmallVector<int, 8> Mask(NumElems, -1);
5658 for (unsigned i = 0; i != NumElems; ++i) {
5659 unsigned Opc = Op.getOperand(i).getOpcode();
5661 if (Opc == ISD::UNDEF)
5664 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5665 // Quit if more than 1 elements need inserting.
5666 if (InsertIndices.size() > 1)
5669 InsertIndices.push_back(i);
5673 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5674 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5675 // Quit if non-constant index.
5676 if (!isa<ConstantSDNode>(ExtIdx))
5678 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5680 // Quit if extracted from vector of different type.
5681 if (ExtractedFromVec.getValueType() != VT)
5684 if (!VecIn1.getNode())
5685 VecIn1 = ExtractedFromVec;
5686 else if (VecIn1 != ExtractedFromVec) {
5687 if (!VecIn2.getNode())
5688 VecIn2 = ExtractedFromVec;
5689 else if (VecIn2 != ExtractedFromVec)
5690 // Quit if more than 2 vectors to shuffle
5694 if (ExtractedFromVec == VecIn1)
5696 else if (ExtractedFromVec == VecIn2)
5697 Mask[i] = Idx + NumElems;
5700 if (!VecIn1.getNode())
5703 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5704 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5705 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5706 unsigned Idx = InsertIndices[i];
5707 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5708 DAG.getIntPtrConstant(Idx, DL));
5714 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5715 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5716 Op.getScalarValueSizeInBits() == 1 &&
5717 "Can not convert non-constant vector");
5718 uint64_t Immediate = 0;
5719 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5720 SDValue In = Op.getOperand(idx);
5721 if (In.getOpcode() != ISD::UNDEF)
5722 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5726 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5727 return DAG.getConstant(Immediate, dl, VT);
5729 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5731 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5733 MVT VT = Op.getSimpleValueType();
5734 assert((VT.getVectorElementType() == MVT::i1) &&
5735 "Unexpected type in LowerBUILD_VECTORvXi1!");
5738 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5739 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5740 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5741 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5744 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5745 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5746 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5747 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5750 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5751 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5752 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5753 return DAG.getBitcast(VT, Imm);
5754 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5755 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5756 DAG.getIntPtrConstant(0, dl));
5759 // Vector has one or more non-const elements
5760 uint64_t Immediate = 0;
5761 SmallVector<unsigned, 16> NonConstIdx;
5762 bool IsSplat = true;
5763 bool HasConstElts = false;
5765 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5766 SDValue In = Op.getOperand(idx);
5767 if (In.getOpcode() == ISD::UNDEF)
5769 if (!isa<ConstantSDNode>(In))
5770 NonConstIdx.push_back(idx);
5772 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5773 HasConstElts = true;
5777 else if (In != Op.getOperand(SplatIdx))
5781 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5783 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5784 DAG.getConstant(1, dl, VT),
5785 DAG.getConstant(0, dl, VT));
5787 // insert elements one by one
5791 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5792 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5794 else if (HasConstElts)
5795 Imm = DAG.getConstant(0, dl, VT);
5797 Imm = DAG.getUNDEF(VT);
5798 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5799 DstVec = DAG.getBitcast(VT, Imm);
5801 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5802 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5803 DAG.getIntPtrConstant(0, dl));
5806 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5807 unsigned InsertIdx = NonConstIdx[i];
5808 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5809 Op.getOperand(InsertIdx),
5810 DAG.getIntPtrConstant(InsertIdx, dl));
5815 /// \brief Return true if \p N implements a horizontal binop and return the
5816 /// operands for the horizontal binop into V0 and V1.
5818 /// This is a helper function of LowerToHorizontalOp().
5819 /// This function checks that the build_vector \p N in input implements a
5820 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5821 /// operation to match.
5822 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5823 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5824 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5827 /// This function only analyzes elements of \p N whose indices are
5828 /// in range [BaseIdx, LastIdx).
5829 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5831 unsigned BaseIdx, unsigned LastIdx,
5832 SDValue &V0, SDValue &V1) {
5833 EVT VT = N->getValueType(0);
5835 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5836 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5837 "Invalid Vector in input!");
5839 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5840 bool CanFold = true;
5841 unsigned ExpectedVExtractIdx = BaseIdx;
5842 unsigned NumElts = LastIdx - BaseIdx;
5843 V0 = DAG.getUNDEF(VT);
5844 V1 = DAG.getUNDEF(VT);
5846 // Check if N implements a horizontal binop.
5847 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5848 SDValue Op = N->getOperand(i + BaseIdx);
5851 if (Op->getOpcode() == ISD::UNDEF) {
5852 // Update the expected vector extract index.
5853 if (i * 2 == NumElts)
5854 ExpectedVExtractIdx = BaseIdx;
5855 ExpectedVExtractIdx += 2;
5859 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5864 SDValue Op0 = Op.getOperand(0);
5865 SDValue Op1 = Op.getOperand(1);
5867 // Try to match the following pattern:
5868 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5869 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5870 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5871 Op0.getOperand(0) == Op1.getOperand(0) &&
5872 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5873 isa<ConstantSDNode>(Op1.getOperand(1)));
5877 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5878 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5880 if (i * 2 < NumElts) {
5881 if (V0.getOpcode() == ISD::UNDEF) {
5882 V0 = Op0.getOperand(0);
5883 if (V0.getValueType() != VT)
5887 if (V1.getOpcode() == ISD::UNDEF) {
5888 V1 = Op0.getOperand(0);
5889 if (V1.getValueType() != VT)
5892 if (i * 2 == NumElts)
5893 ExpectedVExtractIdx = BaseIdx;
5896 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5897 if (I0 == ExpectedVExtractIdx)
5898 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5899 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5900 // Try to match the following dag sequence:
5901 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5902 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5906 ExpectedVExtractIdx += 2;
5912 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5913 /// a concat_vector.
5915 /// This is a helper function of LowerToHorizontalOp().
5916 /// This function expects two 256-bit vectors called V0 and V1.
5917 /// At first, each vector is split into two separate 128-bit vectors.
5918 /// Then, the resulting 128-bit vectors are used to implement two
5919 /// horizontal binary operations.
5921 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5923 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5924 /// the two new horizontal binop.
5925 /// When Mode is set, the first horizontal binop dag node would take as input
5926 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5927 /// horizontal binop dag node would take as input the lower 128-bit of V1
5928 /// and the upper 128-bit of V1.
5930 /// HADD V0_LO, V0_HI
5931 /// HADD V1_LO, V1_HI
5933 /// Otherwise, the first horizontal binop dag node takes as input the lower
5934 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5935 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5937 /// HADD V0_LO, V1_LO
5938 /// HADD V0_HI, V1_HI
5940 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5941 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5942 /// the upper 128-bits of the result.
5943 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5944 SDLoc DL, SelectionDAG &DAG,
5945 unsigned X86Opcode, bool Mode,
5946 bool isUndefLO, bool isUndefHI) {
5947 EVT VT = V0.getValueType();
5948 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5949 "Invalid nodes in input!");
5951 unsigned NumElts = VT.getVectorNumElements();
5952 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5953 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5954 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5955 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5956 EVT NewVT = V0_LO.getValueType();
5958 SDValue LO = DAG.getUNDEF(NewVT);
5959 SDValue HI = DAG.getUNDEF(NewVT);
5962 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5963 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5964 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5965 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5966 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5968 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5969 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5970 V1_LO->getOpcode() != ISD::UNDEF))
5971 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5973 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5974 V1_HI->getOpcode() != ISD::UNDEF))
5975 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5978 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5981 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5983 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5984 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5985 MVT VT = BV->getSimpleValueType(0);
5986 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5987 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5991 unsigned NumElts = VT.getVectorNumElements();
5992 SDValue InVec0 = DAG.getUNDEF(VT);
5993 SDValue InVec1 = DAG.getUNDEF(VT);
5995 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5996 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5998 // Odd-numbered elements in the input build vector are obtained from
5999 // adding two integer/float elements.
6000 // Even-numbered elements in the input build vector are obtained from
6001 // subtracting two integer/float elements.
6002 unsigned ExpectedOpcode = ISD::FSUB;
6003 unsigned NextExpectedOpcode = ISD::FADD;
6004 bool AddFound = false;
6005 bool SubFound = false;
6007 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6008 SDValue Op = BV->getOperand(i);
6010 // Skip 'undef' values.
6011 unsigned Opcode = Op.getOpcode();
6012 if (Opcode == ISD::UNDEF) {
6013 std::swap(ExpectedOpcode, NextExpectedOpcode);
6017 // Early exit if we found an unexpected opcode.
6018 if (Opcode != ExpectedOpcode)
6021 SDValue Op0 = Op.getOperand(0);
6022 SDValue Op1 = Op.getOperand(1);
6024 // Try to match the following pattern:
6025 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6026 // Early exit if we cannot match that sequence.
6027 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6028 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6029 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6030 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6031 Op0.getOperand(1) != Op1.getOperand(1))
6034 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6038 // We found a valid add/sub node. Update the information accordingly.
6044 // Update InVec0 and InVec1.
6045 if (InVec0.getOpcode() == ISD::UNDEF) {
6046 InVec0 = Op0.getOperand(0);
6047 if (InVec0.getSimpleValueType() != VT)
6050 if (InVec1.getOpcode() == ISD::UNDEF) {
6051 InVec1 = Op1.getOperand(0);
6052 if (InVec1.getSimpleValueType() != VT)
6056 // Make sure that operands in input to each add/sub node always
6057 // come from a same pair of vectors.
6058 if (InVec0 != Op0.getOperand(0)) {
6059 if (ExpectedOpcode == ISD::FSUB)
6062 // FADD is commutable. Try to commute the operands
6063 // and then test again.
6064 std::swap(Op0, Op1);
6065 if (InVec0 != Op0.getOperand(0))
6069 if (InVec1 != Op1.getOperand(0))
6072 // Update the pair of expected opcodes.
6073 std::swap(ExpectedOpcode, NextExpectedOpcode);
6076 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6077 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6078 InVec1.getOpcode() != ISD::UNDEF)
6079 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6084 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6085 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6086 const X86Subtarget *Subtarget,
6087 SelectionDAG &DAG) {
6088 MVT VT = BV->getSimpleValueType(0);
6089 unsigned NumElts = VT.getVectorNumElements();
6090 unsigned NumUndefsLO = 0;
6091 unsigned NumUndefsHI = 0;
6092 unsigned Half = NumElts/2;
6094 // Count the number of UNDEF operands in the build_vector in input.
6095 for (unsigned i = 0, e = Half; i != e; ++i)
6096 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6099 for (unsigned i = Half, e = NumElts; i != e; ++i)
6100 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6103 // Early exit if this is either a build_vector of all UNDEFs or all the
6104 // operands but one are UNDEF.
6105 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6109 SDValue InVec0, InVec1;
6110 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6111 // Try to match an SSE3 float HADD/HSUB.
6112 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6113 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6115 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6116 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6117 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6118 // Try to match an SSSE3 integer HADD/HSUB.
6119 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6120 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6122 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6123 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6126 if (!Subtarget->hasAVX())
6129 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6130 // Try to match an AVX horizontal add/sub of packed single/double
6131 // precision floating point values from 256-bit vectors.
6132 SDValue InVec2, InVec3;
6133 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6134 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6135 ((InVec0.getOpcode() == ISD::UNDEF ||
6136 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6137 ((InVec1.getOpcode() == ISD::UNDEF ||
6138 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6139 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6141 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6142 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6143 ((InVec0.getOpcode() == ISD::UNDEF ||
6144 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6145 ((InVec1.getOpcode() == ISD::UNDEF ||
6146 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6147 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6148 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6149 // Try to match an AVX2 horizontal add/sub of signed integers.
6150 SDValue InVec2, InVec3;
6152 bool CanFold = true;
6154 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6155 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6156 ((InVec0.getOpcode() == ISD::UNDEF ||
6157 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6158 ((InVec1.getOpcode() == ISD::UNDEF ||
6159 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6160 X86Opcode = X86ISD::HADD;
6161 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6162 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6163 ((InVec0.getOpcode() == ISD::UNDEF ||
6164 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6165 ((InVec1.getOpcode() == ISD::UNDEF ||
6166 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6167 X86Opcode = X86ISD::HSUB;
6172 // Fold this build_vector into a single horizontal add/sub.
6173 // Do this only if the target has AVX2.
6174 if (Subtarget->hasAVX2())
6175 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6177 // Do not try to expand this build_vector into a pair of horizontal
6178 // add/sub if we can emit a pair of scalar add/sub.
6179 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6182 // Convert this build_vector into a pair of horizontal binop followed by
6184 bool isUndefLO = NumUndefsLO == Half;
6185 bool isUndefHI = NumUndefsHI == Half;
6186 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6187 isUndefLO, isUndefHI);
6191 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6192 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6194 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6195 X86Opcode = X86ISD::HADD;
6196 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6197 X86Opcode = X86ISD::HSUB;
6198 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6199 X86Opcode = X86ISD::FHADD;
6200 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6201 X86Opcode = X86ISD::FHSUB;
6205 // Don't try to expand this build_vector into a pair of horizontal add/sub
6206 // if we can simply emit a pair of scalar add/sub.
6207 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6210 // Convert this build_vector into two horizontal add/sub followed by
6212 bool isUndefLO = NumUndefsLO == Half;
6213 bool isUndefHI = NumUndefsHI == Half;
6214 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6215 isUndefLO, isUndefHI);
6222 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6225 MVT VT = Op.getSimpleValueType();
6226 MVT ExtVT = VT.getVectorElementType();
6227 unsigned NumElems = Op.getNumOperands();
6229 // Generate vectors for predicate vectors.
6230 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6231 return LowerBUILD_VECTORvXi1(Op, DAG);
6233 // Vectors containing all zeros can be matched by pxor and xorps later
6234 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6235 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6236 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6237 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6240 return getZeroVector(VT, Subtarget, DAG, dl);
6243 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6244 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6245 // vpcmpeqd on 256-bit vectors.
6246 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6247 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6250 if (!VT.is512BitVector())
6251 return getOnesVector(VT, Subtarget, DAG, dl);
6254 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6255 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6257 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6258 return HorizontalOp;
6259 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6262 unsigned EVTBits = ExtVT.getSizeInBits();
6264 unsigned NumZero = 0;
6265 unsigned NumNonZero = 0;
6266 uint64_t NonZeros = 0;
6267 bool IsAllConstants = true;
6268 SmallSet<SDValue, 8> Values;
6269 for (unsigned i = 0; i < NumElems; ++i) {
6270 SDValue Elt = Op.getOperand(i);
6271 if (Elt.getOpcode() == ISD::UNDEF)
6274 if (Elt.getOpcode() != ISD::Constant &&
6275 Elt.getOpcode() != ISD::ConstantFP)
6276 IsAllConstants = false;
6277 if (X86::isZeroNode(Elt))
6280 assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
6281 NonZeros |= ((uint64_t)1 << i);
6286 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6287 if (NumNonZero == 0)
6288 return DAG.getUNDEF(VT);
6290 // Special case for single non-zero, non-undef, element.
6291 if (NumNonZero == 1) {
6292 unsigned Idx = countTrailingZeros(NonZeros);
6293 SDValue Item = Op.getOperand(Idx);
6295 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6296 // the value are obviously zero, truncate the value to i32 and do the
6297 // insertion that way. Only do this if the value is non-constant or if the
6298 // value is a constant being inserted into element 0. It is cheaper to do
6299 // a constant pool load than it is to do a movd + shuffle.
6300 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6301 (!IsAllConstants || Idx == 0)) {
6302 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6304 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6305 MVT VecVT = MVT::v4i32;
6307 // Truncate the value (which may itself be a constant) to i32, and
6308 // convert it to a vector with movd (S2V+shuffle to zero extend).
6309 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6310 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6311 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6312 Item, Idx * 2, true, Subtarget, DAG));
6316 // If we have a constant or non-constant insertion into the low element of
6317 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6318 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6319 // depending on what the source datatype is.
6322 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6324 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6325 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6326 if (VT.is512BitVector()) {
6327 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6328 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6329 Item, DAG.getIntPtrConstant(0, dl));
6331 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6332 "Expected an SSE value type!");
6333 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6334 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6335 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6338 // We can't directly insert an i8 or i16 into a vector, so zero extend
6340 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6341 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6342 if (VT.is256BitVector()) {
6343 if (Subtarget->hasAVX()) {
6344 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6345 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6347 // Without AVX, we need to extend to a 128-bit vector and then
6348 // insert into the 256-bit vector.
6349 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6350 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6351 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6354 assert(VT.is128BitVector() && "Expected an SSE value type!");
6355 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6356 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6358 return DAG.getBitcast(VT, Item);
6362 // Is it a vector logical left shift?
6363 if (NumElems == 2 && Idx == 1 &&
6364 X86::isZeroNode(Op.getOperand(0)) &&
6365 !X86::isZeroNode(Op.getOperand(1))) {
6366 unsigned NumBits = VT.getSizeInBits();
6367 return getVShift(true, VT,
6368 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6369 VT, Op.getOperand(1)),
6370 NumBits/2, DAG, *this, dl);
6373 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6376 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6377 // is a non-constant being inserted into an element other than the low one,
6378 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6379 // movd/movss) to move this into the low element, then shuffle it into
6381 if (EVTBits == 32) {
6382 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6383 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6387 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6388 if (Values.size() == 1) {
6389 if (EVTBits == 32) {
6390 // Instead of a shuffle like this:
6391 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6392 // Check if it's possible to issue this instead.
6393 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6394 unsigned Idx = countTrailingZeros(NonZeros);
6395 SDValue Item = Op.getOperand(Idx);
6396 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6397 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6402 // A vector full of immediates; various special cases are already
6403 // handled, so this is best done with a single constant-pool load.
6407 // For AVX-length vectors, see if we can use a vector load to get all of the
6408 // elements, otherwise build the individual 128-bit pieces and use
6409 // shuffles to put them in place.
6410 if (VT.is256BitVector() || VT.is512BitVector()) {
6411 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6413 // Check for a build vector of consecutive loads.
6414 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6417 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6419 // Build both the lower and upper subvector.
6420 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6421 makeArrayRef(&V[0], NumElems/2));
6422 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6423 makeArrayRef(&V[NumElems / 2], NumElems/2));
6425 // Recreate the wider vector with the lower and upper part.
6426 if (VT.is256BitVector())
6427 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6428 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6431 // Let legalizer expand 2-wide build_vectors.
6432 if (EVTBits == 64) {
6433 if (NumNonZero == 1) {
6434 // One half is zero or undef.
6435 unsigned Idx = countTrailingZeros(NonZeros);
6436 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6437 Op.getOperand(Idx));
6438 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6443 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6444 if (EVTBits == 8 && NumElems == 16)
6445 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
6446 DAG, Subtarget, *this))
6449 if (EVTBits == 16 && NumElems == 8)
6450 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
6451 DAG, Subtarget, *this))
6454 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6455 if (EVTBits == 32 && NumElems == 4)
6456 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6459 // If element VT is == 32 bits, turn it into a number of shuffles.
6460 SmallVector<SDValue, 8> V(NumElems);
6461 if (NumElems == 4 && NumZero > 0) {
6462 for (unsigned i = 0; i < 4; ++i) {
6463 bool isZero = !(NonZeros & (1ULL << i));
6465 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6467 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6470 for (unsigned i = 0; i < 2; ++i) {
6471 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6474 V[i] = V[i*2]; // Must be a zero vector.
6477 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6480 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6483 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6488 bool Reverse1 = (NonZeros & 0x3) == 2;
6489 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6493 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6494 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6496 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6499 if (Values.size() > 1 && VT.is128BitVector()) {
6500 // Check for a build vector of consecutive loads.
6501 for (unsigned i = 0; i < NumElems; ++i)
6502 V[i] = Op.getOperand(i);
6504 // Check for elements which are consecutive loads.
6505 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6508 // Check for a build vector from mostly shuffle plus few inserting.
6509 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6512 // For SSE 4.1, use insertps to put the high elements into the low element.
6513 if (Subtarget->hasSSE41()) {
6515 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6516 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6518 Result = DAG.getUNDEF(VT);
6520 for (unsigned i = 1; i < NumElems; ++i) {
6521 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6522 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6523 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6528 // Otherwise, expand into a number of unpckl*, start by extending each of
6529 // our (non-undef) elements to the full vector width with the element in the
6530 // bottom slot of the vector (which generates no code for SSE).
6531 for (unsigned i = 0; i < NumElems; ++i) {
6532 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6533 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6535 V[i] = DAG.getUNDEF(VT);
6538 // Next, we iteratively mix elements, e.g. for v4f32:
6539 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6540 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6541 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6542 unsigned EltStride = NumElems >> 1;
6543 while (EltStride != 0) {
6544 for (unsigned i = 0; i < EltStride; ++i) {
6545 // If V[i+EltStride] is undef and this is the first round of mixing,
6546 // then it is safe to just drop this shuffle: V[i] is already in the
6547 // right place, the one element (since it's the first round) being
6548 // inserted as undef can be dropped. This isn't safe for successive
6549 // rounds because they will permute elements within both vectors.
6550 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6551 EltStride == NumElems/2)
6554 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6563 // 256-bit AVX can use the vinsertf128 instruction
6564 // to create 256-bit vectors from two other 128-bit ones.
6565 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6567 MVT ResVT = Op.getSimpleValueType();
6569 assert((ResVT.is256BitVector() ||
6570 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6572 SDValue V1 = Op.getOperand(0);
6573 SDValue V2 = Op.getOperand(1);
6574 unsigned NumElems = ResVT.getVectorNumElements();
6575 if (ResVT.is256BitVector())
6576 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6578 if (Op.getNumOperands() == 4) {
6579 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6580 ResVT.getVectorNumElements()/2);
6581 SDValue V3 = Op.getOperand(2);
6582 SDValue V4 = Op.getOperand(3);
6583 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6584 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6586 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6589 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6590 const X86Subtarget *Subtarget,
6591 SelectionDAG & DAG) {
6593 MVT ResVT = Op.getSimpleValueType();
6594 unsigned NumOfOperands = Op.getNumOperands();
6596 assert(isPowerOf2_32(NumOfOperands) &&
6597 "Unexpected number of operands in CONCAT_VECTORS");
6599 SDValue Undef = DAG.getUNDEF(ResVT);
6600 if (NumOfOperands > 2) {
6601 // Specialize the cases when all, or all but one, of the operands are undef.
6602 unsigned NumOfDefinedOps = 0;
6604 for (unsigned i = 0; i < NumOfOperands; i++)
6605 if (!Op.getOperand(i).isUndef()) {
6609 if (NumOfDefinedOps == 0)
6611 if (NumOfDefinedOps == 1) {
6612 unsigned SubVecNumElts =
6613 Op.getOperand(OpIdx).getValueType().getVectorNumElements();
6614 SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
6615 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
6616 Op.getOperand(OpIdx), IdxVal);
6619 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6620 ResVT.getVectorNumElements()/2);
6621 SmallVector<SDValue, 2> Ops;
6622 for (unsigned i = 0; i < NumOfOperands/2; i++)
6623 Ops.push_back(Op.getOperand(i));
6624 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6626 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6627 Ops.push_back(Op.getOperand(i));
6628 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6629 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6633 SDValue V1 = Op.getOperand(0);
6634 SDValue V2 = Op.getOperand(1);
6635 unsigned NumElems = ResVT.getVectorNumElements();
6636 assert(V1.getValueType() == V2.getValueType() &&
6637 V1.getValueType().getVectorNumElements() == NumElems/2 &&
6638 "Unexpected operands in CONCAT_VECTORS");
6640 if (ResVT.getSizeInBits() >= 16)
6641 return Op; // The operation is legal with KUNPCK
6643 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6644 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6645 SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
6646 if (IsZeroV1 && IsZeroV2)
6649 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6651 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6653 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
6655 SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
6657 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
6660 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
6662 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6663 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
6666 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6667 const X86Subtarget *Subtarget,
6668 SelectionDAG &DAG) {
6669 MVT VT = Op.getSimpleValueType();
6670 if (VT.getVectorElementType() == MVT::i1)
6671 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6673 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6674 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6675 Op.getNumOperands() == 4)));
6677 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6678 // from two other 128-bit ones.
6680 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6681 return LowerAVXCONCAT_VECTORS(Op, DAG);
6684 //===----------------------------------------------------------------------===//
6685 // Vector shuffle lowering
6687 // This is an experimental code path for lowering vector shuffles on x86. It is
6688 // designed to handle arbitrary vector shuffles and blends, gracefully
6689 // degrading performance as necessary. It works hard to recognize idiomatic
6690 // shuffles and lower them to optimal instruction patterns without leaving
6691 // a framework that allows reasonably efficient handling of all vector shuffle
6693 //===----------------------------------------------------------------------===//
6695 /// \brief Tiny helper function to identify a no-op mask.
6697 /// This is a somewhat boring predicate function. It checks whether the mask
6698 /// array input, which is assumed to be a single-input shuffle mask of the kind
6699 /// used by the X86 shuffle instructions (not a fully general
6700 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6701 /// in-place shuffle are 'no-op's.
6702 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6703 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6704 if (Mask[i] != -1 && Mask[i] != i)
6709 /// \brief Helper function to classify a mask as a single-input mask.
6711 /// This isn't a generic single-input test because in the vector shuffle
6712 /// lowering we canonicalize single inputs to be the first input operand. This
6713 /// means we can more quickly test for a single input by only checking whether
6714 /// an input from the second operand exists. We also assume that the size of
6715 /// mask corresponds to the size of the input vectors which isn't true in the
6716 /// fully general case.
6717 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6719 if (M >= (int)Mask.size())
6724 /// \brief Test whether there are elements crossing 128-bit lanes in this
6727 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6728 /// and we routinely test for these.
6729 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6730 int LaneSize = 128 / VT.getScalarSizeInBits();
6731 int Size = Mask.size();
6732 for (int i = 0; i < Size; ++i)
6733 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6738 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6740 /// This checks a shuffle mask to see if it is performing the same
6741 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6742 /// that it is also not lane-crossing. It may however involve a blend from the
6743 /// same lane of a second vector.
6745 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6746 /// non-trivial to compute in the face of undef lanes. The representation is
6747 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6748 /// entries from both V1 and V2 inputs to the wider mask.
6750 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6751 SmallVectorImpl<int> &RepeatedMask) {
6752 int LaneSize = 128 / VT.getScalarSizeInBits();
6753 RepeatedMask.resize(LaneSize, -1);
6754 int Size = Mask.size();
6755 for (int i = 0; i < Size; ++i) {
6758 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6759 // This entry crosses lanes, so there is no way to model this shuffle.
6762 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6763 if (RepeatedMask[i % LaneSize] == -1)
6764 // This is the first non-undef entry in this slot of a 128-bit lane.
6765 RepeatedMask[i % LaneSize] =
6766 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6767 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6768 // Found a mismatch with the repeated mask.
6774 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6777 /// This is a fast way to test a shuffle mask against a fixed pattern:
6779 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6781 /// It returns true if the mask is exactly as wide as the argument list, and
6782 /// each element of the mask is either -1 (signifying undef) or the value given
6783 /// in the argument.
6784 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6785 ArrayRef<int> ExpectedMask) {
6786 if (Mask.size() != ExpectedMask.size())
6789 int Size = Mask.size();
6791 // If the values are build vectors, we can look through them to find
6792 // equivalent inputs that make the shuffles equivalent.
6793 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6794 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6796 for (int i = 0; i < Size; ++i)
6797 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6798 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6799 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6800 if (!MaskBV || !ExpectedBV ||
6801 MaskBV->getOperand(Mask[i] % Size) !=
6802 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6809 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6811 /// This helper function produces an 8-bit shuffle immediate corresponding to
6812 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6813 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6816 /// NB: We rely heavily on "undef" masks preserving the input lane.
6817 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6818 SelectionDAG &DAG) {
6819 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6820 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6821 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6822 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6823 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6826 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6827 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6828 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6829 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6830 return DAG.getConstant(Imm, DL, MVT::i8);
6833 /// \brief Compute whether each element of a shuffle is zeroable.
6835 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6836 /// Either it is an undef element in the shuffle mask, the element of the input
6837 /// referenced is undef, or the element of the input referenced is known to be
6838 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6839 /// as many lanes with this technique as possible to simplify the remaining
6841 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6842 SDValue V1, SDValue V2) {
6843 SmallBitVector Zeroable(Mask.size(), false);
6845 while (V1.getOpcode() == ISD::BITCAST)
6846 V1 = V1->getOperand(0);
6847 while (V2.getOpcode() == ISD::BITCAST)
6848 V2 = V2->getOperand(0);
6850 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6851 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6853 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6855 // Handle the easy cases.
6856 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6861 // If this is an index into a build_vector node (which has the same number
6862 // of elements), dig out the input value and use it.
6863 SDValue V = M < Size ? V1 : V2;
6864 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6867 SDValue Input = V.getOperand(M % Size);
6868 // The UNDEF opcode check really should be dead code here, but not quite
6869 // worth asserting on (it isn't invalid, just unexpected).
6870 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6877 // X86 has dedicated unpack instructions that can handle specific blend
6878 // operations: UNPCKH and UNPCKL.
6879 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6880 SDValue V1, SDValue V2,
6881 SelectionDAG &DAG) {
6882 int NumElts = VT.getVectorNumElements();
6883 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6884 SmallVector<int, 8> Unpckl;
6885 SmallVector<int, 8> Unpckh;
6887 for (int i = 0; i < NumElts; ++i) {
6888 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6889 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6890 int HiPos = LoPos + NumEltsInLane / 2;
6891 Unpckl.push_back(LoPos);
6892 Unpckh.push_back(HiPos);
6895 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6896 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6897 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6898 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6900 // Commute and try again.
6901 ShuffleVectorSDNode::commuteMask(Unpckl);
6902 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6903 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6905 ShuffleVectorSDNode::commuteMask(Unpckh);
6906 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6907 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6912 /// \brief Try to emit a bitmask instruction for a shuffle.
6914 /// This handles cases where we can model a blend exactly as a bitmask due to
6915 /// one of the inputs being zeroable.
6916 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6917 SDValue V2, ArrayRef<int> Mask,
6918 SelectionDAG &DAG) {
6919 MVT EltVT = VT.getVectorElementType();
6920 int NumEltBits = EltVT.getSizeInBits();
6921 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6922 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6923 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6925 if (EltVT.isFloatingPoint()) {
6926 Zero = DAG.getBitcast(EltVT, Zero);
6927 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6929 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6930 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6932 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6935 if (Mask[i] % Size != i)
6936 return SDValue(); // Not a blend.
6938 V = Mask[i] < Size ? V1 : V2;
6939 else if (V != (Mask[i] < Size ? V1 : V2))
6940 return SDValue(); // Can only let one input through the mask.
6942 VMaskOps[i] = AllOnes;
6945 return SDValue(); // No non-zeroable elements!
6947 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6948 V = DAG.getNode(VT.isFloatingPoint()
6949 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6954 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6956 /// This is used as a fallback approach when first class blend instructions are
6957 /// unavailable. Currently it is only suitable for integer vectors, but could
6958 /// be generalized for floating point vectors if desirable.
6959 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6960 SDValue V2, ArrayRef<int> Mask,
6961 SelectionDAG &DAG) {
6962 assert(VT.isInteger() && "Only supports integer vector types!");
6963 MVT EltVT = VT.getVectorElementType();
6964 int NumEltBits = EltVT.getSizeInBits();
6965 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6966 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6968 SmallVector<SDValue, 16> MaskOps;
6969 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6970 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6971 return SDValue(); // Shuffled input!
6972 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6975 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6976 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6977 // We have to cast V2 around.
6978 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6979 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6980 DAG.getBitcast(MaskVT, V1Mask),
6981 DAG.getBitcast(MaskVT, V2)));
6982 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6985 /// \brief Try to emit a blend instruction for a shuffle.
6987 /// This doesn't do any checks for the availability of instructions for blending
6988 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6989 /// be matched in the backend with the type given. What it does check for is
6990 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6991 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6992 SDValue V2, ArrayRef<int> Original,
6993 const X86Subtarget *Subtarget,
6994 SelectionDAG &DAG) {
6995 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6996 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6997 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6998 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6999 bool ForceV1Zero = false, ForceV2Zero = false;
7001 // Attempt to generate the binary blend mask. If an input is zero then
7002 // we can use any lane.
7003 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
7004 unsigned BlendMask = 0;
7005 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7011 if (M == i + Size) {
7012 BlendMask |= 1u << i;
7023 BlendMask |= 1u << i;
7028 return SDValue(); // Shuffled input!
7031 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
7033 V1 = getZeroVector(VT, Subtarget, DAG, DL);
7035 V2 = getZeroVector(VT, Subtarget, DAG, DL);
7037 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
7038 unsigned ScaledMask = 0;
7039 for (int i = 0; i != Size; ++i)
7040 if (BlendMask & (1u << i))
7041 for (int j = 0; j != Scale; ++j)
7042 ScaledMask |= 1u << (i * Scale + j);
7046 switch (VT.SimpleTy) {
7051 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7052 DAG.getConstant(BlendMask, DL, MVT::i8));
7056 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7060 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7061 // that instruction.
7062 if (Subtarget->hasAVX2()) {
7063 // Scale the blend by the number of 32-bit dwords per element.
7064 int Scale = VT.getScalarSizeInBits() / 32;
7065 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7066 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7067 V1 = DAG.getBitcast(BlendVT, V1);
7068 V2 = DAG.getBitcast(BlendVT, V2);
7069 return DAG.getBitcast(
7070 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7071 DAG.getConstant(BlendMask, DL, MVT::i8)));
7075 // For integer shuffles we need to expand the mask and cast the inputs to
7076 // v8i16s prior to blending.
7077 int Scale = 8 / VT.getVectorNumElements();
7078 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7079 V1 = DAG.getBitcast(MVT::v8i16, V1);
7080 V2 = DAG.getBitcast(MVT::v8i16, V2);
7081 return DAG.getBitcast(VT,
7082 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7083 DAG.getConstant(BlendMask, DL, MVT::i8)));
7087 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7088 SmallVector<int, 8> RepeatedMask;
7089 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7090 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7091 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7093 for (int i = 0; i < 8; ++i)
7094 if (RepeatedMask[i] >= 16)
7095 BlendMask |= 1u << i;
7096 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7097 DAG.getConstant(BlendMask, DL, MVT::i8));
7103 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
7104 "256-bit byte-blends require AVX2 support!");
7106 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
7107 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
7110 // Scale the blend by the number of bytes per element.
7111 int Scale = VT.getScalarSizeInBits() / 8;
7113 // This form of blend is always done on bytes. Compute the byte vector
7115 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7117 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7118 // mix of LLVM's code generator and the x86 backend. We tell the code
7119 // generator that boolean values in the elements of an x86 vector register
7120 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7121 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7122 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7123 // of the element (the remaining are ignored) and 0 in that high bit would
7124 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7125 // the LLVM model for boolean values in vector elements gets the relevant
7126 // bit set, it is set backwards and over constrained relative to x86's
7128 SmallVector<SDValue, 32> VSELECTMask;
7129 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7130 for (int j = 0; j < Scale; ++j)
7131 VSELECTMask.push_back(
7132 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7133 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7136 V1 = DAG.getBitcast(BlendVT, V1);
7137 V2 = DAG.getBitcast(BlendVT, V2);
7138 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7139 DAG.getNode(ISD::BUILD_VECTOR, DL,
7140 BlendVT, VSELECTMask),
7145 llvm_unreachable("Not a supported integer vector type!");
7149 /// \brief Try to lower as a blend of elements from two inputs followed by
7150 /// a single-input permutation.
7152 /// This matches the pattern where we can blend elements from two inputs and
7153 /// then reduce the shuffle to a single-input permutation.
7154 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7157 SelectionDAG &DAG) {
7158 // We build up the blend mask while checking whether a blend is a viable way
7159 // to reduce the shuffle.
7160 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7161 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7163 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7167 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7169 if (BlendMask[Mask[i] % Size] == -1)
7170 BlendMask[Mask[i] % Size] = Mask[i];
7171 else if (BlendMask[Mask[i] % Size] != Mask[i])
7172 return SDValue(); // Can't blend in the needed input!
7174 PermuteMask[i] = Mask[i] % Size;
7177 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7178 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7181 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7182 /// blends and permutes.
7184 /// This matches the extremely common pattern for handling combined
7185 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7186 /// operations. It will try to pick the best arrangement of shuffles and
7188 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7192 SelectionDAG &DAG) {
7193 // Shuffle the input elements into the desired positions in V1 and V2 and
7194 // blend them together.
7195 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7196 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7197 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7198 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7199 if (Mask[i] >= 0 && Mask[i] < Size) {
7200 V1Mask[i] = Mask[i];
7202 } else if (Mask[i] >= Size) {
7203 V2Mask[i] = Mask[i] - Size;
7204 BlendMask[i] = i + Size;
7207 // Try to lower with the simpler initial blend strategy unless one of the
7208 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7209 // shuffle may be able to fold with a load or other benefit. However, when
7210 // we'll have to do 2x as many shuffles in order to achieve this, blending
7211 // first is a better strategy.
7212 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7213 if (SDValue BlendPerm =
7214 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7217 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7218 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7219 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7222 /// \brief Try to lower a vector shuffle as a byte rotation.
7224 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7225 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7226 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7227 /// try to generically lower a vector shuffle through such an pattern. It
7228 /// does not check for the profitability of lowering either as PALIGNR or
7229 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7230 /// This matches shuffle vectors that look like:
7232 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7234 /// Essentially it concatenates V1 and V2, shifts right by some number of
7235 /// elements, and takes the low elements as the result. Note that while this is
7236 /// specified as a *right shift* because x86 is little-endian, it is a *left
7237 /// rotate* of the vector lanes.
7238 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7241 const X86Subtarget *Subtarget,
7242 SelectionDAG &DAG) {
7243 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7245 int NumElts = Mask.size();
7246 int NumLanes = VT.getSizeInBits() / 128;
7247 int NumLaneElts = NumElts / NumLanes;
7249 // We need to detect various ways of spelling a rotation:
7250 // [11, 12, 13, 14, 15, 0, 1, 2]
7251 // [-1, 12, 13, 14, -1, -1, 1, -1]
7252 // [-1, -1, -1, -1, -1, -1, 1, 2]
7253 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7254 // [-1, 4, 5, 6, -1, -1, 9, -1]
7255 // [-1, 4, 5, 6, -1, -1, -1, -1]
7258 for (int l = 0; l < NumElts; l += NumLaneElts) {
7259 for (int i = 0; i < NumLaneElts; ++i) {
7260 if (Mask[l + i] == -1)
7262 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7264 // Get the mod-Size index and lane correct it.
7265 int LaneIdx = (Mask[l + i] % NumElts) - l;
7266 // Make sure it was in this lane.
7267 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7270 // Determine where a rotated vector would have started.
7271 int StartIdx = i - LaneIdx;
7273 // The identity rotation isn't interesting, stop.
7276 // If we found the tail of a vector the rotation must be the missing
7277 // front. If we found the head of a vector, it must be how much of the
7279 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7282 Rotation = CandidateRotation;
7283 else if (Rotation != CandidateRotation)
7284 // The rotations don't match, so we can't match this mask.
7287 // Compute which value this mask is pointing at.
7288 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7290 // Compute which of the two target values this index should be assigned
7291 // to. This reflects whether the high elements are remaining or the low
7292 // elements are remaining.
7293 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7295 // Either set up this value if we've not encountered it before, or check
7296 // that it remains consistent.
7299 else if (TargetV != MaskV)
7300 // This may be a rotation, but it pulls from the inputs in some
7301 // unsupported interleaving.
7306 // Check that we successfully analyzed the mask, and normalize the results.
7307 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7308 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7314 // The actual rotate instruction rotates bytes, so we need to scale the
7315 // rotation based on how many bytes are in the vector lane.
7316 int Scale = 16 / NumLaneElts;
7318 // SSSE3 targets can use the palignr instruction.
7319 if (Subtarget->hasSSSE3()) {
7320 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7321 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7322 Lo = DAG.getBitcast(AlignVT, Lo);
7323 Hi = DAG.getBitcast(AlignVT, Hi);
7325 return DAG.getBitcast(
7326 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7327 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7330 assert(VT.is128BitVector() &&
7331 "Rotate-based lowering only supports 128-bit lowering!");
7332 assert(Mask.size() <= 16 &&
7333 "Can shuffle at most 16 bytes in a 128-bit vector!");
7335 // Default SSE2 implementation
7336 int LoByteShift = 16 - Rotation * Scale;
7337 int HiByteShift = Rotation * Scale;
7339 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7340 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7341 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7343 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7344 DAG.getConstant(LoByteShift, DL, MVT::i8));
7345 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7346 DAG.getConstant(HiByteShift, DL, MVT::i8));
7347 return DAG.getBitcast(VT,
7348 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7351 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7353 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7354 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7355 /// matches elements from one of the input vectors shuffled to the left or
7356 /// right with zeroable elements 'shifted in'. It handles both the strictly
7357 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7360 /// PSHL : (little-endian) left bit shift.
7361 /// [ zz, 0, zz, 2 ]
7362 /// [ -1, 4, zz, -1 ]
7363 /// PSRL : (little-endian) right bit shift.
7365 /// [ -1, -1, 7, zz]
7366 /// PSLLDQ : (little-endian) left byte shift
7367 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7368 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7369 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7370 /// PSRLDQ : (little-endian) right byte shift
7371 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7372 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7373 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7374 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7375 SDValue V2, ArrayRef<int> Mask,
7376 SelectionDAG &DAG) {
7377 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7379 int Size = Mask.size();
7380 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7382 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7383 for (int i = 0; i < Size; i += Scale)
7384 for (int j = 0; j < Shift; ++j)
7385 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7391 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7392 for (int i = 0; i != Size; i += Scale) {
7393 unsigned Pos = Left ? i + Shift : i;
7394 unsigned Low = Left ? i : i + Shift;
7395 unsigned Len = Scale - Shift;
7396 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7397 Low + (V == V1 ? 0 : Size)))
7401 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7402 bool ByteShift = ShiftEltBits > 64;
7403 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7404 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7405 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7407 // Normalize the scale for byte shifts to still produce an i64 element
7409 Scale = ByteShift ? Scale / 2 : Scale;
7411 // We need to round trip through the appropriate type for the shift.
7412 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7413 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7414 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7415 "Illegal integer vector type");
7416 V = DAG.getBitcast(ShiftVT, V);
7418 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7419 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7420 return DAG.getBitcast(VT, V);
7423 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7424 // keep doubling the size of the integer elements up to that. We can
7425 // then shift the elements of the integer vector by whole multiples of
7426 // their width within the elements of the larger integer vector. Test each
7427 // multiple to see if we can find a match with the moved element indices
7428 // and that the shifted in elements are all zeroable.
7429 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7430 for (int Shift = 1; Shift != Scale; ++Shift)
7431 for (bool Left : {true, false})
7432 if (CheckZeros(Shift, Scale, Left))
7433 for (SDValue V : {V1, V2})
7434 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7441 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7442 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7443 SDValue V2, ArrayRef<int> Mask,
7444 SelectionDAG &DAG) {
7445 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7446 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7448 int Size = Mask.size();
7449 int HalfSize = Size / 2;
7450 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7452 // Upper half must be undefined.
7453 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7456 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7457 // Remainder of lower half result is zero and upper half is all undef.
7458 auto LowerAsEXTRQ = [&]() {
7459 // Determine the extraction length from the part of the
7460 // lower half that isn't zeroable.
7462 for (; Len > 0; --Len)
7463 if (!Zeroable[Len - 1])
7465 assert(Len > 0 && "Zeroable shuffle mask");
7467 // Attempt to match first Len sequential elements from the lower half.
7470 for (int i = 0; i != Len; ++i) {
7474 SDValue &V = (M < Size ? V1 : V2);
7477 // The extracted elements must start at a valid index and all mask
7478 // elements must be in the lower half.
7479 if (i > M || M >= HalfSize)
7482 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7493 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7494 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7495 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7496 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7497 DAG.getConstant(BitLen, DL, MVT::i8),
7498 DAG.getConstant(BitIdx, DL, MVT::i8));
7501 if (SDValue ExtrQ = LowerAsEXTRQ())
7504 // INSERTQ: Extract lowest Len elements from lower half of second source and
7505 // insert over first source, starting at Idx.
7506 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7507 auto LowerAsInsertQ = [&]() {
7508 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7511 // Attempt to match first source from mask before insertion point.
7512 if (isUndefInRange(Mask, 0, Idx)) {
7514 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7516 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7522 // Extend the extraction length looking to match both the insertion of
7523 // the second source and the remaining elements of the first.
7524 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7529 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7531 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7537 // Match the remaining elements of the lower half.
7538 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7540 } else if ((!Base || (Base == V1)) &&
7541 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7543 } else if ((!Base || (Base == V2)) &&
7544 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7551 // We may not have a base (first source) - this can safely be undefined.
7553 Base = DAG.getUNDEF(VT);
7555 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7556 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7557 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7558 DAG.getConstant(BitLen, DL, MVT::i8),
7559 DAG.getConstant(BitIdx, DL, MVT::i8));
7566 if (SDValue InsertQ = LowerAsInsertQ())
7572 /// \brief Lower a vector shuffle as a zero or any extension.
7574 /// Given a specific number of elements, element bit width, and extension
7575 /// stride, produce either a zero or any extension based on the available
7576 /// features of the subtarget. The extended elements are consecutive and
7577 /// begin and can start from an offseted element index in the input; to
7578 /// avoid excess shuffling the offset must either being in the bottom lane
7579 /// or at the start of a higher lane. All extended elements must be from
7581 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7582 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7583 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7584 assert(Scale > 1 && "Need a scale to extend.");
7585 int EltBits = VT.getScalarSizeInBits();
7586 int NumElements = VT.getVectorNumElements();
7587 int NumEltsPerLane = 128 / EltBits;
7588 int OffsetLane = Offset / NumEltsPerLane;
7589 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7590 "Only 8, 16, and 32 bit elements can be extended.");
7591 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7592 assert(0 <= Offset && "Extension offset must be positive.");
7593 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7594 "Extension offset must be in the first lane or start an upper lane.");
7596 // Check that an index is in same lane as the base offset.
7597 auto SafeOffset = [&](int Idx) {
7598 return OffsetLane == (Idx / NumEltsPerLane);
7601 // Shift along an input so that the offset base moves to the first element.
7602 auto ShuffleOffset = [&](SDValue V) {
7606 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7607 for (int i = 0; i * Scale < NumElements; ++i) {
7608 int SrcIdx = i + Offset;
7609 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7611 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7614 // Found a valid zext mask! Try various lowering strategies based on the
7615 // input type and available ISA extensions.
7616 if (Subtarget->hasSSE41()) {
7617 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7618 // PUNPCK will catch this in a later shuffle match.
7619 if (Offset && Scale == 2 && VT.is128BitVector())
7621 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7622 NumElements / Scale);
7623 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7624 return DAG.getBitcast(VT, InputV);
7627 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7629 // For any extends we can cheat for larger element sizes and use shuffle
7630 // instructions that can fold with a load and/or copy.
7631 if (AnyExt && EltBits == 32) {
7632 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7634 return DAG.getBitcast(
7635 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7636 DAG.getBitcast(MVT::v4i32, InputV),
7637 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7639 if (AnyExt && EltBits == 16 && Scale > 2) {
7640 int PSHUFDMask[4] = {Offset / 2, -1,
7641 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7642 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7643 DAG.getBitcast(MVT::v4i32, InputV),
7644 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7645 int PSHUFWMask[4] = {1, -1, -1, -1};
7646 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7647 return DAG.getBitcast(
7648 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7649 DAG.getBitcast(MVT::v8i16, InputV),
7650 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7653 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7655 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7656 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7657 assert(VT.is128BitVector() && "Unexpected vector width!");
7659 int LoIdx = Offset * EltBits;
7660 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7661 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7662 DAG.getConstant(EltBits, DL, MVT::i8),
7663 DAG.getConstant(LoIdx, DL, MVT::i8)));
7665 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7666 !SafeOffset(Offset + 1))
7667 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7669 int HiIdx = (Offset + 1) * EltBits;
7670 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7671 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7672 DAG.getConstant(EltBits, DL, MVT::i8),
7673 DAG.getConstant(HiIdx, DL, MVT::i8)));
7674 return DAG.getNode(ISD::BITCAST, DL, VT,
7675 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7678 // If this would require more than 2 unpack instructions to expand, use
7679 // pshufb when available. We can only use more than 2 unpack instructions
7680 // when zero extending i8 elements which also makes it easier to use pshufb.
7681 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7682 assert(NumElements == 16 && "Unexpected byte vector width!");
7683 SDValue PSHUFBMask[16];
7684 for (int i = 0; i < 16; ++i) {
7685 int Idx = Offset + (i / Scale);
7686 PSHUFBMask[i] = DAG.getConstant(
7687 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7689 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7690 return DAG.getBitcast(VT,
7691 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7692 DAG.getNode(ISD::BUILD_VECTOR, DL,
7693 MVT::v16i8, PSHUFBMask)));
7696 // If we are extending from an offset, ensure we start on a boundary that
7697 // we can unpack from.
7698 int AlignToUnpack = Offset % (NumElements / Scale);
7699 if (AlignToUnpack) {
7700 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7701 for (int i = AlignToUnpack; i < NumElements; ++i)
7702 ShMask[i - AlignToUnpack] = i;
7703 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7704 Offset -= AlignToUnpack;
7707 // Otherwise emit a sequence of unpacks.
7709 unsigned UnpackLoHi = X86ISD::UNPCKL;
7710 if (Offset >= (NumElements / 2)) {
7711 UnpackLoHi = X86ISD::UNPCKH;
7712 Offset -= (NumElements / 2);
7715 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7716 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7717 : getZeroVector(InputVT, Subtarget, DAG, DL);
7718 InputV = DAG.getBitcast(InputVT, InputV);
7719 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7723 } while (Scale > 1);
7724 return DAG.getBitcast(VT, InputV);
7727 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7729 /// This routine will try to do everything in its power to cleverly lower
7730 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7731 /// check for the profitability of this lowering, it tries to aggressively
7732 /// match this pattern. It will use all of the micro-architectural details it
7733 /// can to emit an efficient lowering. It handles both blends with all-zero
7734 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7735 /// masking out later).
7737 /// The reason we have dedicated lowering for zext-style shuffles is that they
7738 /// are both incredibly common and often quite performance sensitive.
7739 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7740 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7741 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7742 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7744 int Bits = VT.getSizeInBits();
7745 int NumLanes = Bits / 128;
7746 int NumElements = VT.getVectorNumElements();
7747 int NumEltsPerLane = NumElements / NumLanes;
7748 assert(VT.getScalarSizeInBits() <= 32 &&
7749 "Exceeds 32-bit integer zero extension limit");
7750 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7752 // Define a helper function to check a particular ext-scale and lower to it if
7754 auto Lower = [&](int Scale) -> SDValue {
7759 for (int i = 0; i < NumElements; ++i) {
7762 continue; // Valid anywhere but doesn't tell us anything.
7763 if (i % Scale != 0) {
7764 // Each of the extended elements need to be zeroable.
7768 // We no longer are in the anyext case.
7773 // Each of the base elements needs to be consecutive indices into the
7774 // same input vector.
7775 SDValue V = M < NumElements ? V1 : V2;
7776 M = M % NumElements;
7779 Offset = M - (i / Scale);
7780 } else if (InputV != V)
7781 return SDValue(); // Flip-flopping inputs.
7783 // Offset must start in the lowest 128-bit lane or at the start of an
7785 // FIXME: Is it ever worth allowing a negative base offset?
7786 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7787 (Offset % NumEltsPerLane) == 0))
7790 // If we are offsetting, all referenced entries must come from the same
7792 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7795 if ((M % NumElements) != (Offset + (i / Scale)))
7796 return SDValue(); // Non-consecutive strided elements.
7800 // If we fail to find an input, we have a zero-shuffle which should always
7801 // have already been handled.
7802 // FIXME: Maybe handle this here in case during blending we end up with one?
7806 // If we are offsetting, don't extend if we only match a single input, we
7807 // can always do better by using a basic PSHUF or PUNPCK.
7808 if (Offset != 0 && Matches < 2)
7811 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7812 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7815 // The widest scale possible for extending is to a 64-bit integer.
7816 assert(Bits % 64 == 0 &&
7817 "The number of bits in a vector must be divisible by 64 on x86!");
7818 int NumExtElements = Bits / 64;
7820 // Each iteration, try extending the elements half as much, but into twice as
7822 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7823 assert(NumElements % NumExtElements == 0 &&
7824 "The input vector size must be divisible by the extended size.");
7825 if (SDValue V = Lower(NumElements / NumExtElements))
7829 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7833 // Returns one of the source operands if the shuffle can be reduced to a
7834 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7835 auto CanZExtLowHalf = [&]() {
7836 for (int i = NumElements / 2; i != NumElements; ++i)
7839 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7841 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7846 if (SDValue V = CanZExtLowHalf()) {
7847 V = DAG.getBitcast(MVT::v2i64, V);
7848 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7849 return DAG.getBitcast(VT, V);
7852 // No viable ext lowering found.
7856 /// \brief Try to get a scalar value for a specific element of a vector.
7858 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7859 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7860 SelectionDAG &DAG) {
7861 MVT VT = V.getSimpleValueType();
7862 MVT EltVT = VT.getVectorElementType();
7863 while (V.getOpcode() == ISD::BITCAST)
7864 V = V.getOperand(0);
7865 // If the bitcasts shift the element size, we can't extract an equivalent
7867 MVT NewVT = V.getSimpleValueType();
7868 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7871 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7872 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7873 // Ensure the scalar operand is the same size as the destination.
7874 // FIXME: Add support for scalar truncation where possible.
7875 SDValue S = V.getOperand(Idx);
7876 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7877 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7883 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7885 /// This is particularly important because the set of instructions varies
7886 /// significantly based on whether the operand is a load or not.
7887 static bool isShuffleFoldableLoad(SDValue V) {
7888 while (V.getOpcode() == ISD::BITCAST)
7889 V = V.getOperand(0);
7891 return ISD::isNON_EXTLoad(V.getNode());
7894 /// \brief Try to lower insertion of a single element into a zero vector.
7896 /// This is a common pattern that we have especially efficient patterns to lower
7897 /// across all subtarget feature sets.
7898 static SDValue lowerVectorShuffleAsElementInsertion(
7899 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7900 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7901 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7903 MVT EltVT = VT.getVectorElementType();
7905 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7906 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7908 bool IsV1Zeroable = true;
7909 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7910 if (i != V2Index && !Zeroable[i]) {
7911 IsV1Zeroable = false;
7915 // Check for a single input from a SCALAR_TO_VECTOR node.
7916 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7917 // all the smarts here sunk into that routine. However, the current
7918 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7919 // vector shuffle lowering is dead.
7920 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7922 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7923 // We need to zext the scalar if it is smaller than an i32.
7924 V2S = DAG.getBitcast(EltVT, V2S);
7925 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7926 // Using zext to expand a narrow element won't work for non-zero
7931 // Zero-extend directly to i32.
7933 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7935 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7936 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7937 EltVT == MVT::i16) {
7938 // Either not inserting from the low element of the input or the input
7939 // element size is too small to use VZEXT_MOVL to clear the high bits.
7943 if (!IsV1Zeroable) {
7944 // If V1 can't be treated as a zero vector we have fewer options to lower
7945 // this. We can't support integer vectors or non-zero targets cheaply, and
7946 // the V1 elements can't be permuted in any way.
7947 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7948 if (!VT.isFloatingPoint() || V2Index != 0)
7950 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7951 V1Mask[V2Index] = -1;
7952 if (!isNoopShuffleMask(V1Mask))
7954 // This is essentially a special case blend operation, but if we have
7955 // general purpose blend operations, they are always faster. Bail and let
7956 // the rest of the lowering handle these as blends.
7957 if (Subtarget->hasSSE41())
7960 // Otherwise, use MOVSD or MOVSS.
7961 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7962 "Only two types of floating point element types to handle!");
7963 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7967 // This lowering only works for the low element with floating point vectors.
7968 if (VT.isFloatingPoint() && V2Index != 0)
7971 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7973 V2 = DAG.getBitcast(VT, V2);
7976 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7977 // the desired position. Otherwise it is more efficient to do a vector
7978 // shift left. We know that we can do a vector shift left because all
7979 // the inputs are zero.
7980 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7981 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7982 V2Shuffle[V2Index] = 0;
7983 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7985 V2 = DAG.getBitcast(MVT::v2i64, V2);
7987 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7988 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7989 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7990 DAG.getDataLayout(), VT)));
7991 V2 = DAG.getBitcast(VT, V2);
7997 /// \brief Try to lower broadcast of a single - truncated - integer element,
7998 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
8000 /// This assumes we have AVX2.
8001 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
8003 const X86Subtarget *Subtarget,
8004 SelectionDAG &DAG) {
8005 assert(Subtarget->hasAVX2() &&
8006 "We can only lower integer broadcasts with AVX2!");
8008 EVT EltVT = VT.getVectorElementType();
8009 EVT V0VT = V0.getValueType();
8011 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
8012 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
8014 EVT V0EltVT = V0VT.getVectorElementType();
8015 if (!V0EltVT.isInteger())
8018 const unsigned EltSize = EltVT.getSizeInBits();
8019 const unsigned V0EltSize = V0EltVT.getSizeInBits();
8021 // This is only a truncation if the original element type is larger.
8022 if (V0EltSize <= EltSize)
8025 assert(((V0EltSize % EltSize) == 0) &&
8026 "Scalar type sizes must all be powers of 2 on x86!");
8028 const unsigned V0Opc = V0.getOpcode();
8029 const unsigned Scale = V0EltSize / EltSize;
8030 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
8032 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
8033 V0Opc != ISD::BUILD_VECTOR)
8036 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
8038 // If we're extracting non-least-significant bits, shift so we can truncate.
8039 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
8040 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
8041 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
8042 if (const int OffsetIdx = BroadcastIdx % Scale)
8043 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
8044 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
8046 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
8047 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
8050 /// \brief Try to lower broadcast of a single element.
8052 /// For convenience, this code also bundles all of the subtarget feature set
8053 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8054 /// a convenient way to factor it out.
8055 /// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
8056 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
8058 const X86Subtarget *Subtarget,
8059 SelectionDAG &DAG) {
8060 if (!Subtarget->hasAVX())
8062 if (VT.isInteger() && !Subtarget->hasAVX2())
8065 // Check that the mask is a broadcast.
8066 int BroadcastIdx = -1;
8068 if (M >= 0 && BroadcastIdx == -1)
8070 else if (M >= 0 && M != BroadcastIdx)
8073 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8074 "a sorted mask where the broadcast "
8077 // Go up the chain of (vector) values to find a scalar load that we can
8078 // combine with the broadcast.
8080 switch (V.getOpcode()) {
8081 case ISD::CONCAT_VECTORS: {
8082 int OperandSize = Mask.size() / V.getNumOperands();
8083 V = V.getOperand(BroadcastIdx / OperandSize);
8084 BroadcastIdx %= OperandSize;
8088 case ISD::INSERT_SUBVECTOR: {
8089 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8090 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8094 int BeginIdx = (int)ConstantIdx->getZExtValue();
8096 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
8097 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8098 BroadcastIdx -= BeginIdx;
8109 // Check if this is a broadcast of a scalar. We special case lowering
8110 // for scalars so that we can more effectively fold with loads.
8111 // First, look through bitcast: if the original value has a larger element
8112 // type than the shuffle, the broadcast element is in essence truncated.
8113 // Make that explicit to ease folding.
8114 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
8115 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
8116 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
8117 return TruncBroadcast;
8119 // Also check the simpler case, where we can directly reuse the scalar.
8120 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8121 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8122 V = V.getOperand(BroadcastIdx);
8124 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8125 // Only AVX2 has register broadcasts.
8126 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8128 } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
8129 // If we are broadcasting a load that is only used by the shuffle
8130 // then we can reduce the vector load to the broadcasted scalar load.
8131 LoadSDNode *Ld = cast<LoadSDNode>(V);
8132 SDValue BaseAddr = Ld->getOperand(1);
8133 EVT AddrVT = BaseAddr.getValueType();
8134 EVT SVT = VT.getScalarType();
8135 unsigned Offset = BroadcastIdx * SVT.getStoreSize();
8136 SDValue NewAddr = DAG.getNode(
8137 ISD::ADD, DL, AddrVT, BaseAddr,
8138 DAG.getConstant(Offset, DL, AddrVT));
8139 V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
8140 DAG.getMachineFunction().getMachineMemOperand(
8141 Ld->getMemOperand(), Offset, SVT.getStoreSize()));
8142 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8143 // We can't broadcast from a vector register without AVX2, and we can only
8144 // broadcast from the zero-element of a vector register.
8148 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8151 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8152 // INSERTPS when the V1 elements are already in the correct locations
8153 // because otherwise we can just always use two SHUFPS instructions which
8154 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8155 // perform INSERTPS if a single V1 element is out of place and all V2
8156 // elements are zeroable.
8157 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8159 SelectionDAG &DAG) {
8160 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8161 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8162 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8163 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8165 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8168 int V1DstIndex = -1;
8169 int V2DstIndex = -1;
8170 bool V1UsedInPlace = false;
8172 for (int i = 0; i < 4; ++i) {
8173 // Synthesize a zero mask from the zeroable elements (includes undefs).
8179 // Flag if we use any V1 inputs in place.
8181 V1UsedInPlace = true;
8185 // We can only insert a single non-zeroable element.
8186 if (V1DstIndex != -1 || V2DstIndex != -1)
8190 // V1 input out of place for insertion.
8193 // V2 input for insertion.
8198 // Don't bother if we have no (non-zeroable) element for insertion.
8199 if (V1DstIndex == -1 && V2DstIndex == -1)
8202 // Determine element insertion src/dst indices. The src index is from the
8203 // start of the inserted vector, not the start of the concatenated vector.
8204 unsigned V2SrcIndex = 0;
8205 if (V1DstIndex != -1) {
8206 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8207 // and don't use the original V2 at all.
8208 V2SrcIndex = Mask[V1DstIndex];
8209 V2DstIndex = V1DstIndex;
8212 V2SrcIndex = Mask[V2DstIndex] - 4;
8215 // If no V1 inputs are used in place, then the result is created only from
8216 // the zero mask and the V2 insertion - so remove V1 dependency.
8218 V1 = DAG.getUNDEF(MVT::v4f32);
8220 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8221 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8223 // Insert the V2 element into the desired position.
8225 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8226 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8229 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8230 /// UNPCK instruction.
8232 /// This specifically targets cases where we end up with alternating between
8233 /// the two inputs, and so can permute them into something that feeds a single
8234 /// UNPCK instruction. Note that this routine only targets integer vectors
8235 /// because for floating point vectors we have a generalized SHUFPS lowering
8236 /// strategy that handles everything that doesn't *exactly* match an unpack,
8237 /// making this clever lowering unnecessary.
8238 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8239 SDValue V1, SDValue V2,
8241 SelectionDAG &DAG) {
8242 assert(!VT.isFloatingPoint() &&
8243 "This routine only supports integer vectors.");
8244 assert(!isSingleInputShuffleMask(Mask) &&
8245 "This routine should only be used when blending two inputs.");
8246 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8248 int Size = Mask.size();
8250 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8251 return M >= 0 && M % Size < Size / 2;
8253 int NumHiInputs = std::count_if(
8254 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8256 bool UnpackLo = NumLoInputs >= NumHiInputs;
8258 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8259 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8260 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8262 for (int i = 0; i < Size; ++i) {
8266 // Each element of the unpack contains Scale elements from this mask.
8267 int UnpackIdx = i / Scale;
8269 // We only handle the case where V1 feeds the first slots of the unpack.
8270 // We rely on canonicalization to ensure this is the case.
8271 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8274 // Setup the mask for this input. The indexing is tricky as we have to
8275 // handle the unpack stride.
8276 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8277 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8281 // If we will have to shuffle both inputs to use the unpack, check whether
8282 // we can just unpack first and shuffle the result. If so, skip this unpack.
8283 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8284 !isNoopShuffleMask(V2Mask))
8287 // Shuffle the inputs into place.
8288 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8289 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8291 // Cast the inputs to the type we will use to unpack them.
8292 V1 = DAG.getBitcast(UnpackVT, V1);
8293 V2 = DAG.getBitcast(UnpackVT, V2);
8295 // Unpack the inputs and cast the result back to the desired type.
8296 return DAG.getBitcast(
8297 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8301 // We try each unpack from the largest to the smallest to try and find one
8302 // that fits this mask.
8303 int OrigNumElements = VT.getVectorNumElements();
8304 int OrigScalarSize = VT.getScalarSizeInBits();
8305 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8306 int Scale = ScalarSize / OrigScalarSize;
8307 int NumElements = OrigNumElements / Scale;
8308 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8309 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8313 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8315 if (NumLoInputs == 0 || NumHiInputs == 0) {
8316 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8317 "We have to have *some* inputs!");
8318 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8320 // FIXME: We could consider the total complexity of the permute of each
8321 // possible unpacking. Or at the least we should consider how many
8322 // half-crossings are created.
8323 // FIXME: We could consider commuting the unpacks.
8325 SmallVector<int, 32> PermMask;
8326 PermMask.assign(Size, -1);
8327 for (int i = 0; i < Size; ++i) {
8331 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8334 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8336 return DAG.getVectorShuffle(
8337 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8339 DAG.getUNDEF(VT), PermMask);
8345 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8347 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8348 /// support for floating point shuffles but not integer shuffles. These
8349 /// instructions will incur a domain crossing penalty on some chips though so
8350 /// it is better to avoid lowering through this for integer vectors where
8352 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8353 const X86Subtarget *Subtarget,
8354 SelectionDAG &DAG) {
8356 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8357 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8358 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8359 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8360 ArrayRef<int> Mask = SVOp->getMask();
8361 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8363 if (isSingleInputShuffleMask(Mask)) {
8364 // Use low duplicate instructions for masks that match their pattern.
8365 if (Subtarget->hasSSE3())
8366 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8367 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8369 // Straight shuffle of a single input vector. Simulate this by using the
8370 // single input as both of the "inputs" to this instruction..
8371 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8373 if (Subtarget->hasAVX()) {
8374 // If we have AVX, we can use VPERMILPS which will allow folding a load
8375 // into the shuffle.
8376 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8377 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8380 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8381 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8383 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8384 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8386 // If we have a single input, insert that into V1 if we can do so cheaply.
8387 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8388 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8389 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8391 // Try inverting the insertion since for v2 masks it is easy to do and we
8392 // can't reliably sort the mask one way or the other.
8393 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8394 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8395 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8396 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8400 // Try to use one of the special instruction patterns to handle two common
8401 // blend patterns if a zero-blend above didn't work.
8402 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8403 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8404 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8405 // We can either use a special instruction to load over the low double or
8406 // to move just the low double.
8408 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8410 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8412 if (Subtarget->hasSSE41())
8413 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8417 // Use dedicated unpack instructions for masks that match their pattern.
8419 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8422 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8423 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8424 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8427 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8429 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8430 /// the integer unit to minimize domain crossing penalties. However, for blends
8431 /// it falls back to the floating point shuffle operation with appropriate bit
8433 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8434 const X86Subtarget *Subtarget,
8435 SelectionDAG &DAG) {
8437 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8438 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8439 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8440 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8441 ArrayRef<int> Mask = SVOp->getMask();
8442 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8444 if (isSingleInputShuffleMask(Mask)) {
8445 // Check for being able to broadcast a single element.
8446 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8447 Mask, Subtarget, DAG))
8450 // Straight shuffle of a single input vector. For everything from SSE2
8451 // onward this has a single fast instruction with no scary immediates.
8452 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8453 V1 = DAG.getBitcast(MVT::v4i32, V1);
8454 int WidenedMask[4] = {
8455 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8456 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8457 return DAG.getBitcast(
8459 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8460 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8462 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8463 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8464 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8465 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8467 // If we have a blend of two PACKUS operations an the blend aligns with the
8468 // low and half halves, we can just merge the PACKUS operations. This is
8469 // particularly important as it lets us merge shuffles that this routine itself
8471 auto GetPackNode = [](SDValue V) {
8472 while (V.getOpcode() == ISD::BITCAST)
8473 V = V.getOperand(0);
8475 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8477 if (SDValue V1Pack = GetPackNode(V1))
8478 if (SDValue V2Pack = GetPackNode(V2))
8479 return DAG.getBitcast(MVT::v2i64,
8480 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8481 Mask[0] == 0 ? V1Pack.getOperand(0)
8482 : V1Pack.getOperand(1),
8483 Mask[1] == 2 ? V2Pack.getOperand(0)
8484 : V2Pack.getOperand(1)));
8486 // Try to use shift instructions.
8488 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8491 // When loading a scalar and then shuffling it into a vector we can often do
8492 // the insertion cheaply.
8493 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8494 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8496 // Try inverting the insertion since for v2 masks it is easy to do and we
8497 // can't reliably sort the mask one way or the other.
8498 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8499 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8500 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8503 // We have different paths for blend lowering, but they all must use the
8504 // *exact* same predicate.
8505 bool IsBlendSupported = Subtarget->hasSSE41();
8506 if (IsBlendSupported)
8507 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8511 // Use dedicated unpack instructions for masks that match their pattern.
8513 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8516 // Try to use byte rotation instructions.
8517 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8518 if (Subtarget->hasSSSE3())
8519 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8520 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8523 // If we have direct support for blends, we should lower by decomposing into
8524 // a permute. That will be faster than the domain cross.
8525 if (IsBlendSupported)
8526 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8529 // We implement this with SHUFPD which is pretty lame because it will likely
8530 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8531 // However, all the alternatives are still more cycles and newer chips don't
8532 // have this problem. It would be really nice if x86 had better shuffles here.
8533 V1 = DAG.getBitcast(MVT::v2f64, V1);
8534 V2 = DAG.getBitcast(MVT::v2f64, V2);
8535 return DAG.getBitcast(MVT::v2i64,
8536 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8539 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8541 /// This is used to disable more specialized lowerings when the shufps lowering
8542 /// will happen to be efficient.
8543 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8544 // This routine only handles 128-bit shufps.
8545 assert(Mask.size() == 4 && "Unsupported mask size!");
8547 // To lower with a single SHUFPS we need to have the low half and high half
8548 // each requiring a single input.
8549 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8551 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8557 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8559 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8560 /// It makes no assumptions about whether this is the *best* lowering, it simply
8562 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8563 ArrayRef<int> Mask, SDValue V1,
8564 SDValue V2, SelectionDAG &DAG) {
8565 SDValue LowV = V1, HighV = V2;
8566 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8569 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8571 if (NumV2Elements == 1) {
8573 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8576 // Compute the index adjacent to V2Index and in the same half by toggling
8578 int V2AdjIndex = V2Index ^ 1;
8580 if (Mask[V2AdjIndex] == -1) {
8581 // Handles all the cases where we have a single V2 element and an undef.
8582 // This will only ever happen in the high lanes because we commute the
8583 // vector otherwise.
8585 std::swap(LowV, HighV);
8586 NewMask[V2Index] -= 4;
8588 // Handle the case where the V2 element ends up adjacent to a V1 element.
8589 // To make this work, blend them together as the first step.
8590 int V1Index = V2AdjIndex;
8591 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8592 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8593 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8595 // Now proceed to reconstruct the final blend as we have the necessary
8596 // high or low half formed.
8603 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8604 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8606 } else if (NumV2Elements == 2) {
8607 if (Mask[0] < 4 && Mask[1] < 4) {
8608 // Handle the easy case where we have V1 in the low lanes and V2 in the
8612 } else if (Mask[2] < 4 && Mask[3] < 4) {
8613 // We also handle the reversed case because this utility may get called
8614 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8615 // arrange things in the right direction.
8621 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8622 // trying to place elements directly, just blend them and set up the final
8623 // shuffle to place them.
8625 // The first two blend mask elements are for V1, the second two are for
8627 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8628 Mask[2] < 4 ? Mask[2] : Mask[3],
8629 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8630 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8631 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8632 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8634 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8637 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8638 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8639 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8640 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8643 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8644 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8647 /// \brief Lower 4-lane 32-bit floating point shuffles.
8649 /// Uses instructions exclusively from the floating point unit to minimize
8650 /// domain crossing penalties, as these are sufficient to implement all v4f32
8652 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8653 const X86Subtarget *Subtarget,
8654 SelectionDAG &DAG) {
8656 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8657 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8658 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8660 ArrayRef<int> Mask = SVOp->getMask();
8661 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8664 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8666 if (NumV2Elements == 0) {
8667 // Check for being able to broadcast a single element.
8668 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8669 Mask, Subtarget, DAG))
8672 // Use even/odd duplicate instructions for masks that match their pattern.
8673 if (Subtarget->hasSSE3()) {
8674 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8675 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8676 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8677 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8680 if (Subtarget->hasAVX()) {
8681 // If we have AVX, we can use VPERMILPS which will allow folding a load
8682 // into the shuffle.
8683 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8684 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8687 // Otherwise, use a straight shuffle of a single input vector. We pass the
8688 // input vector to both operands to simulate this with a SHUFPS.
8689 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8690 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8693 // There are special ways we can lower some single-element blends. However, we
8694 // have custom ways we can lower more complex single-element blends below that
8695 // we defer to if both this and BLENDPS fail to match, so restrict this to
8696 // when the V2 input is targeting element 0 of the mask -- that is the fast
8698 if (NumV2Elements == 1 && Mask[0] >= 4)
8699 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8700 Mask, Subtarget, DAG))
8703 if (Subtarget->hasSSE41()) {
8704 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8708 // Use INSERTPS if we can complete the shuffle efficiently.
8709 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8712 if (!isSingleSHUFPSMask(Mask))
8713 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8714 DL, MVT::v4f32, V1, V2, Mask, DAG))
8718 // Use dedicated unpack instructions for masks that match their pattern.
8720 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8723 // Otherwise fall back to a SHUFPS lowering strategy.
8724 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8727 /// \brief Lower 4-lane i32 vector shuffles.
8729 /// We try to handle these with integer-domain shuffles where we can, but for
8730 /// blends we use the floating point domain blend instructions.
8731 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8732 const X86Subtarget *Subtarget,
8733 SelectionDAG &DAG) {
8735 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8736 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8737 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8738 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8739 ArrayRef<int> Mask = SVOp->getMask();
8740 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8742 // Whenever we can lower this as a zext, that instruction is strictly faster
8743 // than any alternative. It also allows us to fold memory operands into the
8744 // shuffle in many cases.
8745 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8746 Mask, Subtarget, DAG))
8750 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8752 if (NumV2Elements == 0) {
8753 // Check for being able to broadcast a single element.
8754 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8755 Mask, Subtarget, DAG))
8758 // Straight shuffle of a single input vector. For everything from SSE2
8759 // onward this has a single fast instruction with no scary immediates.
8760 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8761 // but we aren't actually going to use the UNPCK instruction because doing
8762 // so prevents folding a load into this instruction or making a copy.
8763 const int UnpackLoMask[] = {0, 0, 1, 1};
8764 const int UnpackHiMask[] = {2, 2, 3, 3};
8765 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8766 Mask = UnpackLoMask;
8767 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8768 Mask = UnpackHiMask;
8770 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8771 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8774 // Try to use shift instructions.
8776 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8779 // There are special ways we can lower some single-element blends.
8780 if (NumV2Elements == 1)
8781 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8782 Mask, Subtarget, DAG))
8785 // We have different paths for blend lowering, but they all must use the
8786 // *exact* same predicate.
8787 bool IsBlendSupported = Subtarget->hasSSE41();
8788 if (IsBlendSupported)
8789 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8793 if (SDValue Masked =
8794 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8797 // Use dedicated unpack instructions for masks that match their pattern.
8799 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8802 // Try to use byte rotation instructions.
8803 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8804 if (Subtarget->hasSSSE3())
8805 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8806 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8809 // If we have direct support for blends, we should lower by decomposing into
8810 // a permute. That will be faster than the domain cross.
8811 if (IsBlendSupported)
8812 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8815 // Try to lower by permuting the inputs into an unpack instruction.
8816 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8820 // We implement this with SHUFPS because it can blend from two vectors.
8821 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8822 // up the inputs, bypassing domain shift penalties that we would encur if we
8823 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8825 return DAG.getBitcast(
8827 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8828 DAG.getBitcast(MVT::v4f32, V2), Mask));
8831 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8832 /// shuffle lowering, and the most complex part.
8834 /// The lowering strategy is to try to form pairs of input lanes which are
8835 /// targeted at the same half of the final vector, and then use a dword shuffle
8836 /// to place them onto the right half, and finally unpack the paired lanes into
8837 /// their final position.
8839 /// The exact breakdown of how to form these dword pairs and align them on the
8840 /// correct sides is really tricky. See the comments within the function for
8841 /// more of the details.
8843 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8844 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8845 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8846 /// vector, form the analogous 128-bit 8-element Mask.
8847 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8848 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8849 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8850 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8851 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8853 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8854 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8855 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8857 SmallVector<int, 4> LoInputs;
8858 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8859 [](int M) { return M >= 0; });
8860 std::sort(LoInputs.begin(), LoInputs.end());
8861 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8862 SmallVector<int, 4> HiInputs;
8863 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8864 [](int M) { return M >= 0; });
8865 std::sort(HiInputs.begin(), HiInputs.end());
8866 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8868 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8869 int NumHToL = LoInputs.size() - NumLToL;
8871 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8872 int NumHToH = HiInputs.size() - NumLToH;
8873 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8874 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8875 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8876 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8878 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8879 // such inputs we can swap two of the dwords across the half mark and end up
8880 // with <=2 inputs to each half in each half. Once there, we can fall through
8881 // to the generic code below. For example:
8883 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8884 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8886 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8887 // and an existing 2-into-2 on the other half. In this case we may have to
8888 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8889 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8890 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8891 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8892 // half than the one we target for fixing) will be fixed when we re-enter this
8893 // path. We will also combine away any sequence of PSHUFD instructions that
8894 // result into a single instruction. Here is an example of the tricky case:
8896 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8897 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8899 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8901 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8902 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8904 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8905 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8907 // The result is fine to be handled by the generic logic.
8908 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8909 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8910 int AOffset, int BOffset) {
8911 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8912 "Must call this with A having 3 or 1 inputs from the A half.");
8913 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8914 "Must call this with B having 1 or 3 inputs from the B half.");
8915 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8916 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8918 bool ThreeAInputs = AToAInputs.size() == 3;
8920 // Compute the index of dword with only one word among the three inputs in
8921 // a half by taking the sum of the half with three inputs and subtracting
8922 // the sum of the actual three inputs. The difference is the remaining
8925 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8926 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8927 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8928 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8929 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8930 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8931 int TripleNonInputIdx =
8932 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8933 TripleDWord = TripleNonInputIdx / 2;
8935 // We use xor with one to compute the adjacent DWord to whichever one the
8937 OneInputDWord = (OneInput / 2) ^ 1;
8939 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8940 // and BToA inputs. If there is also such a problem with the BToB and AToB
8941 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8942 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8943 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8944 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8945 // Compute how many inputs will be flipped by swapping these DWords. We
8947 // to balance this to ensure we don't form a 3-1 shuffle in the other
8949 int NumFlippedAToBInputs =
8950 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8951 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8952 int NumFlippedBToBInputs =
8953 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8954 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8955 if ((NumFlippedAToBInputs == 1 &&
8956 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8957 (NumFlippedBToBInputs == 1 &&
8958 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8959 // We choose whether to fix the A half or B half based on whether that
8960 // half has zero flipped inputs. At zero, we may not be able to fix it
8961 // with that half. We also bias towards fixing the B half because that
8962 // will more commonly be the high half, and we have to bias one way.
8963 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8964 ArrayRef<int> Inputs) {
8965 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8966 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8967 PinnedIdx ^ 1) != Inputs.end();
8968 // Determine whether the free index is in the flipped dword or the
8969 // unflipped dword based on where the pinned index is. We use this bit
8970 // in an xor to conditionally select the adjacent dword.
8971 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8972 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8973 FixFreeIdx) != Inputs.end();
8974 if (IsFixIdxInput == IsFixFreeIdxInput)
8976 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8977 FixFreeIdx) != Inputs.end();
8978 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8979 "We need to be changing the number of flipped inputs!");
8980 int PSHUFHalfMask[] = {0, 1, 2, 3};
8981 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8982 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8984 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8987 if (M != -1 && M == FixIdx)
8989 else if (M != -1 && M == FixFreeIdx)
8992 if (NumFlippedBToBInputs != 0) {
8994 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8995 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8997 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8998 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8999 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9004 int PSHUFDMask[] = {0, 1, 2, 3};
9005 PSHUFDMask[ADWord] = BDWord;
9006 PSHUFDMask[BDWord] = ADWord;
9009 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9010 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9012 // Adjust the mask to match the new locations of A and B.
9014 if (M != -1 && M/2 == ADWord)
9015 M = 2 * BDWord + M % 2;
9016 else if (M != -1 && M/2 == BDWord)
9017 M = 2 * ADWord + M % 2;
9019 // Recurse back into this routine to re-compute state now that this isn't
9020 // a 3 and 1 problem.
9021 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
9024 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9025 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9026 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9027 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9029 // At this point there are at most two inputs to the low and high halves from
9030 // each half. That means the inputs can always be grouped into dwords and
9031 // those dwords can then be moved to the correct half with a dword shuffle.
9032 // We use at most one low and one high word shuffle to collect these paired
9033 // inputs into dwords, and finally a dword shuffle to place them.
9034 int PSHUFLMask[4] = {-1, -1, -1, -1};
9035 int PSHUFHMask[4] = {-1, -1, -1, -1};
9036 int PSHUFDMask[4] = {-1, -1, -1, -1};
9038 // First fix the masks for all the inputs that are staying in their
9039 // original halves. This will then dictate the targets of the cross-half
9041 auto fixInPlaceInputs =
9042 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9043 MutableArrayRef<int> SourceHalfMask,
9044 MutableArrayRef<int> HalfMask, int HalfOffset) {
9045 if (InPlaceInputs.empty())
9047 if (InPlaceInputs.size() == 1) {
9048 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9049 InPlaceInputs[0] - HalfOffset;
9050 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9053 if (IncomingInputs.empty()) {
9054 // Just fix all of the in place inputs.
9055 for (int Input : InPlaceInputs) {
9056 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9057 PSHUFDMask[Input / 2] = Input / 2;
9062 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9063 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9064 InPlaceInputs[0] - HalfOffset;
9065 // Put the second input next to the first so that they are packed into
9066 // a dword. We find the adjacent index by toggling the low bit.
9067 int AdjIndex = InPlaceInputs[0] ^ 1;
9068 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9069 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9070 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9072 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9073 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9075 // Now gather the cross-half inputs and place them into a free dword of
9076 // their target half.
9077 // FIXME: This operation could almost certainly be simplified dramatically to
9078 // look more like the 3-1 fixing operation.
9079 auto moveInputsToRightHalf = [&PSHUFDMask](
9080 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9081 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9082 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9084 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9085 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9087 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9089 int LowWord = Word & ~1;
9090 int HighWord = Word | 1;
9091 return isWordClobbered(SourceHalfMask, LowWord) ||
9092 isWordClobbered(SourceHalfMask, HighWord);
9095 if (IncomingInputs.empty())
9098 if (ExistingInputs.empty()) {
9099 // Map any dwords with inputs from them into the right half.
9100 for (int Input : IncomingInputs) {
9101 // If the source half mask maps over the inputs, turn those into
9102 // swaps and use the swapped lane.
9103 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9104 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9105 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9106 Input - SourceOffset;
9107 // We have to swap the uses in our half mask in one sweep.
9108 for (int &M : HalfMask)
9109 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9111 else if (M == Input)
9112 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9114 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9115 Input - SourceOffset &&
9116 "Previous placement doesn't match!");
9118 // Note that this correctly re-maps both when we do a swap and when
9119 // we observe the other side of the swap above. We rely on that to
9120 // avoid swapping the members of the input list directly.
9121 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9124 // Map the input's dword into the correct half.
9125 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9126 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9128 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9130 "Previous placement doesn't match!");
9133 // And just directly shift any other-half mask elements to be same-half
9134 // as we will have mirrored the dword containing the element into the
9135 // same position within that half.
9136 for (int &M : HalfMask)
9137 if (M >= SourceOffset && M < SourceOffset + 4) {
9138 M = M - SourceOffset + DestOffset;
9139 assert(M >= 0 && "This should never wrap below zero!");
9144 // Ensure we have the input in a viable dword of its current half. This
9145 // is particularly tricky because the original position may be clobbered
9146 // by inputs being moved and *staying* in that half.
9147 if (IncomingInputs.size() == 1) {
9148 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9149 int InputFixed = std::find(std::begin(SourceHalfMask),
9150 std::end(SourceHalfMask), -1) -
9151 std::begin(SourceHalfMask) + SourceOffset;
9152 SourceHalfMask[InputFixed - SourceOffset] =
9153 IncomingInputs[0] - SourceOffset;
9154 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9156 IncomingInputs[0] = InputFixed;
9158 } else if (IncomingInputs.size() == 2) {
9159 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9160 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9161 // We have two non-adjacent or clobbered inputs we need to extract from
9162 // the source half. To do this, we need to map them into some adjacent
9163 // dword slot in the source mask.
9164 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9165 IncomingInputs[1] - SourceOffset};
9167 // If there is a free slot in the source half mask adjacent to one of
9168 // the inputs, place the other input in it. We use (Index XOR 1) to
9169 // compute an adjacent index.
9170 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9171 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9172 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9173 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9174 InputsFixed[1] = InputsFixed[0] ^ 1;
9175 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9176 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9177 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9178 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9179 InputsFixed[0] = InputsFixed[1] ^ 1;
9180 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9181 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9182 // The two inputs are in the same DWord but it is clobbered and the
9183 // adjacent DWord isn't used at all. Move both inputs to the free
9185 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9186 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9187 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9188 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9190 // The only way we hit this point is if there is no clobbering
9191 // (because there are no off-half inputs to this half) and there is no
9192 // free slot adjacent to one of the inputs. In this case, we have to
9193 // swap an input with a non-input.
9194 for (int i = 0; i < 4; ++i)
9195 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9196 "We can't handle any clobbers here!");
9197 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9198 "Cannot have adjacent inputs here!");
9200 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9201 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9203 // We also have to update the final source mask in this case because
9204 // it may need to undo the above swap.
9205 for (int &M : FinalSourceHalfMask)
9206 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9207 M = InputsFixed[1] + SourceOffset;
9208 else if (M == InputsFixed[1] + SourceOffset)
9209 M = (InputsFixed[0] ^ 1) + SourceOffset;
9211 InputsFixed[1] = InputsFixed[0] ^ 1;
9214 // Point everything at the fixed inputs.
9215 for (int &M : HalfMask)
9216 if (M == IncomingInputs[0])
9217 M = InputsFixed[0] + SourceOffset;
9218 else if (M == IncomingInputs[1])
9219 M = InputsFixed[1] + SourceOffset;
9221 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9222 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9225 llvm_unreachable("Unhandled input size!");
9228 // Now hoist the DWord down to the right half.
9229 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9230 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9231 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9232 for (int &M : HalfMask)
9233 for (int Input : IncomingInputs)
9235 M = FreeDWord * 2 + Input % 2;
9237 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9238 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9239 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9240 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9242 // Now enact all the shuffles we've computed to move the inputs into their
9244 if (!isNoopShuffleMask(PSHUFLMask))
9245 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9246 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9247 if (!isNoopShuffleMask(PSHUFHMask))
9248 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9249 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9250 if (!isNoopShuffleMask(PSHUFDMask))
9253 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9254 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9256 // At this point, each half should contain all its inputs, and we can then
9257 // just shuffle them into their final position.
9258 assert(std::count_if(LoMask.begin(), LoMask.end(),
9259 [](int M) { return M >= 4; }) == 0 &&
9260 "Failed to lift all the high half inputs to the low mask!");
9261 assert(std::count_if(HiMask.begin(), HiMask.end(),
9262 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9263 "Failed to lift all the low half inputs to the high mask!");
9265 // Do a half shuffle for the low mask.
9266 if (!isNoopShuffleMask(LoMask))
9267 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9268 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9270 // Do a half shuffle with the high mask after shifting its values down.
9271 for (int &M : HiMask)
9274 if (!isNoopShuffleMask(HiMask))
9275 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9276 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9281 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9282 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9283 SDValue V2, ArrayRef<int> Mask,
9284 SelectionDAG &DAG, bool &V1InUse,
9286 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9292 int Size = Mask.size();
9293 int Scale = 16 / Size;
9294 for (int i = 0; i < 16; ++i) {
9295 if (Mask[i / Scale] == -1) {
9296 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9298 const int ZeroMask = 0x80;
9299 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9301 int V2Idx = Mask[i / Scale] < Size
9303 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9304 if (Zeroable[i / Scale])
9305 V1Idx = V2Idx = ZeroMask;
9306 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9307 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9308 V1InUse |= (ZeroMask != V1Idx);
9309 V2InUse |= (ZeroMask != V2Idx);
9314 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9315 DAG.getBitcast(MVT::v16i8, V1),
9316 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9318 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9319 DAG.getBitcast(MVT::v16i8, V2),
9320 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9322 // If we need shuffled inputs from both, blend the two.
9324 if (V1InUse && V2InUse)
9325 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9327 V = V1InUse ? V1 : V2;
9329 // Cast the result back to the correct type.
9330 return DAG.getBitcast(VT, V);
9333 /// \brief Generic lowering of 8-lane i16 shuffles.
9335 /// This handles both single-input shuffles and combined shuffle/blends with
9336 /// two inputs. The single input shuffles are immediately delegated to
9337 /// a dedicated lowering routine.
9339 /// The blends are lowered in one of three fundamental ways. If there are few
9340 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9341 /// of the input is significantly cheaper when lowered as an interleaving of
9342 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9343 /// halves of the inputs separately (making them have relatively few inputs)
9344 /// and then concatenate them.
9345 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9346 const X86Subtarget *Subtarget,
9347 SelectionDAG &DAG) {
9349 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9350 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9351 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9352 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9353 ArrayRef<int> OrigMask = SVOp->getMask();
9354 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9355 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9356 MutableArrayRef<int> Mask(MaskStorage);
9358 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9360 // Whenever we can lower this as a zext, that instruction is strictly faster
9361 // than any alternative.
9362 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9363 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9366 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9368 auto isV2 = [](int M) { return M >= 8; };
9370 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9372 if (NumV2Inputs == 0) {
9373 // Check for being able to broadcast a single element.
9374 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9375 Mask, Subtarget, DAG))
9378 // Try to use shift instructions.
9380 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9383 // Use dedicated unpack instructions for masks that match their pattern.
9385 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9388 // Try to use byte rotation instructions.
9389 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9390 Mask, Subtarget, DAG))
9393 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9397 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9398 "All single-input shuffles should be canonicalized to be V1-input "
9401 // Try to use shift instructions.
9403 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9406 // See if we can use SSE4A Extraction / Insertion.
9407 if (Subtarget->hasSSE4A())
9408 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9411 // There are special ways we can lower some single-element blends.
9412 if (NumV2Inputs == 1)
9413 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9414 Mask, Subtarget, DAG))
9417 // We have different paths for blend lowering, but they all must use the
9418 // *exact* same predicate.
9419 bool IsBlendSupported = Subtarget->hasSSE41();
9420 if (IsBlendSupported)
9421 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9425 if (SDValue Masked =
9426 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9429 // Use dedicated unpack instructions for masks that match their pattern.
9431 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9434 // Try to use byte rotation instructions.
9435 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9436 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9439 if (SDValue BitBlend =
9440 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9443 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9447 // If we can't directly blend but can use PSHUFB, that will be better as it
9448 // can both shuffle and set up the inefficient blend.
9449 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9450 bool V1InUse, V2InUse;
9451 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9455 // We can always bit-blend if we have to so the fallback strategy is to
9456 // decompose into single-input permutes and blends.
9457 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9461 /// \brief Check whether a compaction lowering can be done by dropping even
9462 /// elements and compute how many times even elements must be dropped.
9464 /// This handles shuffles which take every Nth element where N is a power of
9465 /// two. Example shuffle masks:
9467 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9468 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9469 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9470 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9471 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9472 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9474 /// Any of these lanes can of course be undef.
9476 /// This routine only supports N <= 3.
9477 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9480 /// \returns N above, or the number of times even elements must be dropped if
9481 /// there is such a number. Otherwise returns zero.
9482 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9483 // Figure out whether we're looping over two inputs or just one.
9484 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9486 // The modulus for the shuffle vector entries is based on whether this is
9487 // a single input or not.
9488 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9489 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9490 "We should only be called with masks with a power-of-2 size!");
9492 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9494 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9495 // and 2^3 simultaneously. This is because we may have ambiguity with
9496 // partially undef inputs.
9497 bool ViableForN[3] = {true, true, true};
9499 for (int i = 0, e = Mask.size(); i < e; ++i) {
9500 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9505 bool IsAnyViable = false;
9506 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9507 if (ViableForN[j]) {
9510 // The shuffle mask must be equal to (i * 2^N) % M.
9511 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9514 ViableForN[j] = false;
9516 // Early exit if we exhaust the possible powers of two.
9521 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9525 // Return 0 as there is no viable power of two.
9529 /// \brief Generic lowering of v16i8 shuffles.
9531 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9532 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9533 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9534 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9536 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9537 const X86Subtarget *Subtarget,
9538 SelectionDAG &DAG) {
9540 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9541 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9542 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9543 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9544 ArrayRef<int> Mask = SVOp->getMask();
9545 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9547 // Try to use shift instructions.
9549 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9552 // Try to use byte rotation instructions.
9553 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9554 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9557 // Try to use a zext lowering.
9558 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9559 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9562 // See if we can use SSE4A Extraction / Insertion.
9563 if (Subtarget->hasSSE4A())
9564 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9568 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9570 // For single-input shuffles, there are some nicer lowering tricks we can use.
9571 if (NumV2Elements == 0) {
9572 // Check for being able to broadcast a single element.
9573 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9574 Mask, Subtarget, DAG))
9577 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9578 // Notably, this handles splat and partial-splat shuffles more efficiently.
9579 // However, it only makes sense if the pre-duplication shuffle simplifies
9580 // things significantly. Currently, this means we need to be able to
9581 // express the pre-duplication shuffle as an i16 shuffle.
9583 // FIXME: We should check for other patterns which can be widened into an
9584 // i16 shuffle as well.
9585 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9586 for (int i = 0; i < 16; i += 2)
9587 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9592 auto tryToWidenViaDuplication = [&]() -> SDValue {
9593 if (!canWidenViaDuplication(Mask))
9595 SmallVector<int, 4> LoInputs;
9596 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9597 [](int M) { return M >= 0 && M < 8; });
9598 std::sort(LoInputs.begin(), LoInputs.end());
9599 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9601 SmallVector<int, 4> HiInputs;
9602 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9603 [](int M) { return M >= 8; });
9604 std::sort(HiInputs.begin(), HiInputs.end());
9605 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9608 bool TargetLo = LoInputs.size() >= HiInputs.size();
9609 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9610 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9612 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9613 SmallDenseMap<int, int, 8> LaneMap;
9614 for (int I : InPlaceInputs) {
9615 PreDupI16Shuffle[I/2] = I/2;
9618 int j = TargetLo ? 0 : 4, je = j + 4;
9619 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9620 // Check if j is already a shuffle of this input. This happens when
9621 // there are two adjacent bytes after we move the low one.
9622 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9623 // If we haven't yet mapped the input, search for a slot into which
9625 while (j < je && PreDupI16Shuffle[j] != -1)
9629 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9632 // Map this input with the i16 shuffle.
9633 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9636 // Update the lane map based on the mapping we ended up with.
9637 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9639 V1 = DAG.getBitcast(
9641 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9642 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9644 // Unpack the bytes to form the i16s that will be shuffled into place.
9645 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9646 MVT::v16i8, V1, V1);
9648 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9649 for (int i = 0; i < 16; ++i)
9650 if (Mask[i] != -1) {
9651 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9652 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9653 if (PostDupI16Shuffle[i / 2] == -1)
9654 PostDupI16Shuffle[i / 2] = MappedMask;
9656 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9657 "Conflicting entrties in the original shuffle!");
9659 return DAG.getBitcast(
9661 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9662 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9664 if (SDValue V = tryToWidenViaDuplication())
9668 if (SDValue Masked =
9669 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9672 // Use dedicated unpack instructions for masks that match their pattern.
9674 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9677 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9678 // with PSHUFB. It is important to do this before we attempt to generate any
9679 // blends but after all of the single-input lowerings. If the single input
9680 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9681 // want to preserve that and we can DAG combine any longer sequences into
9682 // a PSHUFB in the end. But once we start blending from multiple inputs,
9683 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9684 // and there are *very* few patterns that would actually be faster than the
9685 // PSHUFB approach because of its ability to zero lanes.
9687 // FIXME: The only exceptions to the above are blends which are exact
9688 // interleavings with direct instructions supporting them. We currently don't
9689 // handle those well here.
9690 if (Subtarget->hasSSSE3()) {
9691 bool V1InUse = false;
9692 bool V2InUse = false;
9694 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9695 DAG, V1InUse, V2InUse);
9697 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9698 // do so. This avoids using them to handle blends-with-zero which is
9699 // important as a single pshufb is significantly faster for that.
9700 if (V1InUse && V2InUse) {
9701 if (Subtarget->hasSSE41())
9702 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9703 Mask, Subtarget, DAG))
9706 // We can use an unpack to do the blending rather than an or in some
9707 // cases. Even though the or may be (very minorly) more efficient, we
9708 // preference this lowering because there are common cases where part of
9709 // the complexity of the shuffles goes away when we do the final blend as
9711 // FIXME: It might be worth trying to detect if the unpack-feeding
9712 // shuffles will both be pshufb, in which case we shouldn't bother with
9714 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9715 DL, MVT::v16i8, V1, V2, Mask, DAG))
9722 // There are special ways we can lower some single-element blends.
9723 if (NumV2Elements == 1)
9724 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9725 Mask, Subtarget, DAG))
9728 if (SDValue BitBlend =
9729 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9732 // Check whether a compaction lowering can be done. This handles shuffles
9733 // which take every Nth element for some even N. See the helper function for
9736 // We special case these as they can be particularly efficiently handled with
9737 // the PACKUSB instruction on x86 and they show up in common patterns of
9738 // rearranging bytes to truncate wide elements.
9739 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9740 // NumEvenDrops is the power of two stride of the elements. Another way of
9741 // thinking about it is that we need to drop the even elements this many
9742 // times to get the original input.
9743 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9745 // First we need to zero all the dropped bytes.
9746 assert(NumEvenDrops <= 3 &&
9747 "No support for dropping even elements more than 3 times.");
9748 // We use the mask type to pick which bytes are preserved based on how many
9749 // elements are dropped.
9750 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9751 SDValue ByteClearMask = DAG.getBitcast(
9752 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9753 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9755 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9757 // Now pack things back together.
9758 V1 = DAG.getBitcast(MVT::v8i16, V1);
9759 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9760 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9761 for (int i = 1; i < NumEvenDrops; ++i) {
9762 Result = DAG.getBitcast(MVT::v8i16, Result);
9763 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9769 // Handle multi-input cases by blending single-input shuffles.
9770 if (NumV2Elements > 0)
9771 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9774 // The fallback path for single-input shuffles widens this into two v8i16
9775 // vectors with unpacks, shuffles those, and then pulls them back together
9779 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9780 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9781 for (int i = 0; i < 16; ++i)
9783 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9785 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9787 SDValue VLoHalf, VHiHalf;
9788 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9789 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9791 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9792 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9793 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9794 [](int M) { return M >= 0 && M % 2 == 1; })) {
9795 // Use a mask to drop the high bytes.
9796 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9797 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9798 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9800 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9801 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9803 // Squash the masks to point directly into VLoHalf.
9804 for (int &M : LoBlendMask)
9807 for (int &M : HiBlendMask)
9811 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9812 // VHiHalf so that we can blend them as i16s.
9813 VLoHalf = DAG.getBitcast(
9814 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9815 VHiHalf = DAG.getBitcast(
9816 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9819 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9820 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9822 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9825 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9827 /// This routine breaks down the specific type of 128-bit shuffle and
9828 /// dispatches to the lowering routines accordingly.
9829 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9830 MVT VT, const X86Subtarget *Subtarget,
9831 SelectionDAG &DAG) {
9832 switch (VT.SimpleTy) {
9834 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9836 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9838 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9840 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9842 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9844 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9847 llvm_unreachable("Unimplemented!");
9851 /// \brief Helper function to test whether a shuffle mask could be
9852 /// simplified by widening the elements being shuffled.
9854 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9855 /// leaves it in an unspecified state.
9857 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9858 /// shuffle masks. The latter have the special property of a '-2' representing
9859 /// a zero-ed lane of a vector.
9860 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9861 SmallVectorImpl<int> &WidenedMask) {
9862 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9863 // If both elements are undef, its trivial.
9864 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9865 WidenedMask.push_back(SM_SentinelUndef);
9869 // Check for an undef mask and a mask value properly aligned to fit with
9870 // a pair of values. If we find such a case, use the non-undef mask's value.
9871 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9872 WidenedMask.push_back(Mask[i + 1] / 2);
9875 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9876 WidenedMask.push_back(Mask[i] / 2);
9880 // When zeroing, we need to spread the zeroing across both lanes to widen.
9881 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9882 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9883 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9884 WidenedMask.push_back(SM_SentinelZero);
9890 // Finally check if the two mask values are adjacent and aligned with
9892 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9893 WidenedMask.push_back(Mask[i] / 2);
9897 // Otherwise we can't safely widen the elements used in this shuffle.
9900 assert(WidenedMask.size() == Mask.size() / 2 &&
9901 "Incorrect size of mask after widening the elements!");
9906 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9908 /// This routine just extracts two subvectors, shuffles them independently, and
9909 /// then concatenates them back together. This should work effectively with all
9910 /// AVX vector shuffle types.
9911 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9912 SDValue V2, ArrayRef<int> Mask,
9913 SelectionDAG &DAG) {
9914 assert(VT.getSizeInBits() >= 256 &&
9915 "Only for 256-bit or wider vector shuffles!");
9916 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9917 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9919 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9920 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9922 int NumElements = VT.getVectorNumElements();
9923 int SplitNumElements = NumElements / 2;
9924 MVT ScalarVT = VT.getVectorElementType();
9925 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9927 // Rather than splitting build-vectors, just build two narrower build
9928 // vectors. This helps shuffling with splats and zeros.
9929 auto SplitVector = [&](SDValue V) {
9930 while (V.getOpcode() == ISD::BITCAST)
9931 V = V->getOperand(0);
9933 MVT OrigVT = V.getSimpleValueType();
9934 int OrigNumElements = OrigVT.getVectorNumElements();
9935 int OrigSplitNumElements = OrigNumElements / 2;
9936 MVT OrigScalarVT = OrigVT.getVectorElementType();
9937 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9941 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9943 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9944 DAG.getIntPtrConstant(0, DL));
9945 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9946 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9949 SmallVector<SDValue, 16> LoOps, HiOps;
9950 for (int i = 0; i < OrigSplitNumElements; ++i) {
9951 LoOps.push_back(BV->getOperand(i));
9952 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9954 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9955 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9957 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9958 DAG.getBitcast(SplitVT, HiV));
9961 SDValue LoV1, HiV1, LoV2, HiV2;
9962 std::tie(LoV1, HiV1) = SplitVector(V1);
9963 std::tie(LoV2, HiV2) = SplitVector(V2);
9965 // Now create two 4-way blends of these half-width vectors.
9966 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9967 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9968 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9969 for (int i = 0; i < SplitNumElements; ++i) {
9970 int M = HalfMask[i];
9971 if (M >= NumElements) {
9972 if (M >= NumElements + SplitNumElements)
9976 V2BlendMask.push_back(M - NumElements);
9977 V1BlendMask.push_back(-1);
9978 BlendMask.push_back(SplitNumElements + i);
9979 } else if (M >= 0) {
9980 if (M >= SplitNumElements)
9984 V2BlendMask.push_back(-1);
9985 V1BlendMask.push_back(M);
9986 BlendMask.push_back(i);
9988 V2BlendMask.push_back(-1);
9989 V1BlendMask.push_back(-1);
9990 BlendMask.push_back(-1);
9994 // Because the lowering happens after all combining takes place, we need to
9995 // manually combine these blend masks as much as possible so that we create
9996 // a minimal number of high-level vector shuffle nodes.
9998 // First try just blending the halves of V1 or V2.
9999 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10000 return DAG.getUNDEF(SplitVT);
10001 if (!UseLoV2 && !UseHiV2)
10002 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10003 if (!UseLoV1 && !UseHiV1)
10004 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10006 SDValue V1Blend, V2Blend;
10007 if (UseLoV1 && UseHiV1) {
10009 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10011 // We only use half of V1 so map the usage down into the final blend mask.
10012 V1Blend = UseLoV1 ? LoV1 : HiV1;
10013 for (int i = 0; i < SplitNumElements; ++i)
10014 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10015 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10017 if (UseLoV2 && UseHiV2) {
10019 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10021 // We only use half of V2 so map the usage down into the final blend mask.
10022 V2Blend = UseLoV2 ? LoV2 : HiV2;
10023 for (int i = 0; i < SplitNumElements; ++i)
10024 if (BlendMask[i] >= SplitNumElements)
10025 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10027 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10029 SDValue Lo = HalfBlend(LoMask);
10030 SDValue Hi = HalfBlend(HiMask);
10031 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10034 /// \brief Either split a vector in halves or decompose the shuffles and the
10037 /// This is provided as a good fallback for many lowerings of non-single-input
10038 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10039 /// between splitting the shuffle into 128-bit components and stitching those
10040 /// back together vs. extracting the single-input shuffles and blending those
10042 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10043 SDValue V2, ArrayRef<int> Mask,
10044 SelectionDAG &DAG) {
10045 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10046 "lower single-input shuffles as it "
10047 "could then recurse on itself.");
10048 int Size = Mask.size();
10050 // If this can be modeled as a broadcast of two elements followed by a blend,
10051 // prefer that lowering. This is especially important because broadcasts can
10052 // often fold with memory operands.
10053 auto DoBothBroadcast = [&] {
10054 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10057 if (V2BroadcastIdx == -1)
10058 V2BroadcastIdx = M - Size;
10059 else if (M - Size != V2BroadcastIdx)
10061 } else if (M >= 0) {
10062 if (V1BroadcastIdx == -1)
10063 V1BroadcastIdx = M;
10064 else if (M != V1BroadcastIdx)
10069 if (DoBothBroadcast())
10070 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10073 // If the inputs all stem from a single 128-bit lane of each input, then we
10074 // split them rather than blending because the split will decompose to
10075 // unusually few instructions.
10076 int LaneCount = VT.getSizeInBits() / 128;
10077 int LaneSize = Size / LaneCount;
10078 SmallBitVector LaneInputs[2];
10079 LaneInputs[0].resize(LaneCount, false);
10080 LaneInputs[1].resize(LaneCount, false);
10081 for (int i = 0; i < Size; ++i)
10083 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10084 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10085 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10087 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10088 // that the decomposed single-input shuffles don't end up here.
10089 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10092 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10093 /// a permutation and blend of those lanes.
10095 /// This essentially blends the out-of-lane inputs to each lane into the lane
10096 /// from a permuted copy of the vector. This lowering strategy results in four
10097 /// instructions in the worst case for a single-input cross lane shuffle which
10098 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10099 /// of. Special cases for each particular shuffle pattern should be handled
10100 /// prior to trying this lowering.
10101 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10102 SDValue V1, SDValue V2,
10103 ArrayRef<int> Mask,
10104 SelectionDAG &DAG) {
10105 // FIXME: This should probably be generalized for 512-bit vectors as well.
10106 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
10107 int LaneSize = Mask.size() / 2;
10109 // If there are only inputs from one 128-bit lane, splitting will in fact be
10110 // less expensive. The flags track whether the given lane contains an element
10111 // that crosses to another lane.
10112 bool LaneCrossing[2] = {false, false};
10113 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10114 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10115 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10116 if (!LaneCrossing[0] || !LaneCrossing[1])
10117 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10119 if (isSingleInputShuffleMask(Mask)) {
10120 SmallVector<int, 32> FlippedBlendMask;
10121 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10122 FlippedBlendMask.push_back(
10123 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10125 : Mask[i] % LaneSize +
10126 (i / LaneSize) * LaneSize + Size));
10128 // Flip the vector, and blend the results which should now be in-lane. The
10129 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10130 // 5 for the high source. The value 3 selects the high half of source 2 and
10131 // the value 2 selects the low half of source 2. We only use source 2 to
10132 // allow folding it into a memory operand.
10133 unsigned PERMMask = 3 | 2 << 4;
10134 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10135 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10136 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10139 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10140 // will be handled by the above logic and a blend of the results, much like
10141 // other patterns in AVX.
10142 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10145 /// \brief Handle lowering 2-lane 128-bit shuffles.
10146 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10147 SDValue V2, ArrayRef<int> Mask,
10148 const X86Subtarget *Subtarget,
10149 SelectionDAG &DAG) {
10150 // TODO: If minimizing size and one of the inputs is a zero vector and the
10151 // the zero vector has only one use, we could use a VPERM2X128 to save the
10152 // instruction bytes needed to explicitly generate the zero vector.
10154 // Blends are faster and handle all the non-lane-crossing cases.
10155 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10159 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10160 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10162 // If either input operand is a zero vector, use VPERM2X128 because its mask
10163 // allows us to replace the zero input with an implicit zero.
10164 if (!IsV1Zero && !IsV2Zero) {
10165 // Check for patterns which can be matched with a single insert of a 128-bit
10167 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10168 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10169 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10170 VT.getVectorNumElements() / 2);
10171 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10172 DAG.getIntPtrConstant(0, DL));
10173 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10174 OnlyUsesV1 ? V1 : V2,
10175 DAG.getIntPtrConstant(0, DL));
10176 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10180 // Otherwise form a 128-bit permutation. After accounting for undefs,
10181 // convert the 64-bit shuffle mask selection values into 128-bit
10182 // selection bits by dividing the indexes by 2 and shifting into positions
10183 // defined by a vperm2*128 instruction's immediate control byte.
10185 // The immediate permute control byte looks like this:
10186 // [1:0] - select 128 bits from sources for low half of destination
10188 // [3] - zero low half of destination
10189 // [5:4] - select 128 bits from sources for high half of destination
10191 // [7] - zero high half of destination
10193 int MaskLO = Mask[0];
10194 if (MaskLO == SM_SentinelUndef)
10195 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10197 int MaskHI = Mask[2];
10198 if (MaskHI == SM_SentinelUndef)
10199 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10201 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10203 // If either input is a zero vector, replace it with an undef input.
10204 // Shuffle mask values < 4 are selecting elements of V1.
10205 // Shuffle mask values >= 4 are selecting elements of V2.
10206 // Adjust each half of the permute mask by clearing the half that was
10207 // selecting the zero vector and setting the zero mask bit.
10209 V1 = DAG.getUNDEF(VT);
10211 PermMask = (PermMask & 0xf0) | 0x08;
10213 PermMask = (PermMask & 0x0f) | 0x80;
10216 V2 = DAG.getUNDEF(VT);
10218 PermMask = (PermMask & 0xf0) | 0x08;
10220 PermMask = (PermMask & 0x0f) | 0x80;
10223 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10224 DAG.getConstant(PermMask, DL, MVT::i8));
10227 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10228 /// shuffling each lane.
10230 /// This will only succeed when the result of fixing the 128-bit lanes results
10231 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10232 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10233 /// the lane crosses early and then use simpler shuffles within each lane.
10235 /// FIXME: It might be worthwhile at some point to support this without
10236 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10237 /// in x86 only floating point has interesting non-repeating shuffles, and even
10238 /// those are still *marginally* more expensive.
10239 static SDValue lowerVectorShuffleByMerging128BitLanes(
10240 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10241 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10242 assert(!isSingleInputShuffleMask(Mask) &&
10243 "This is only useful with multiple inputs.");
10245 int Size = Mask.size();
10246 int LaneSize = 128 / VT.getScalarSizeInBits();
10247 int NumLanes = Size / LaneSize;
10248 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10250 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10251 // check whether the in-128-bit lane shuffles share a repeating pattern.
10252 SmallVector<int, 4> Lanes;
10253 Lanes.resize(NumLanes, -1);
10254 SmallVector<int, 4> InLaneMask;
10255 InLaneMask.resize(LaneSize, -1);
10256 for (int i = 0; i < Size; ++i) {
10260 int j = i / LaneSize;
10262 if (Lanes[j] < 0) {
10263 // First entry we've seen for this lane.
10264 Lanes[j] = Mask[i] / LaneSize;
10265 } else if (Lanes[j] != Mask[i] / LaneSize) {
10266 // This doesn't match the lane selected previously!
10270 // Check that within each lane we have a consistent shuffle mask.
10271 int k = i % LaneSize;
10272 if (InLaneMask[k] < 0) {
10273 InLaneMask[k] = Mask[i] % LaneSize;
10274 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10275 // This doesn't fit a repeating in-lane mask.
10280 // First shuffle the lanes into place.
10281 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10282 VT.getSizeInBits() / 64);
10283 SmallVector<int, 8> LaneMask;
10284 LaneMask.resize(NumLanes * 2, -1);
10285 for (int i = 0; i < NumLanes; ++i)
10286 if (Lanes[i] >= 0) {
10287 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10288 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10291 V1 = DAG.getBitcast(LaneVT, V1);
10292 V2 = DAG.getBitcast(LaneVT, V2);
10293 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10295 // Cast it back to the type we actually want.
10296 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10298 // Now do a simple shuffle that isn't lane crossing.
10299 SmallVector<int, 8> NewMask;
10300 NewMask.resize(Size, -1);
10301 for (int i = 0; i < Size; ++i)
10303 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10304 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10305 "Must not introduce lane crosses at this point!");
10307 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10310 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10313 /// This returns true if the elements from a particular input are already in the
10314 /// slot required by the given mask and require no permutation.
10315 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10316 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10317 int Size = Mask.size();
10318 for (int i = 0; i < Size; ++i)
10319 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10325 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10326 ArrayRef<int> Mask, SDValue V1,
10327 SDValue V2, SelectionDAG &DAG) {
10329 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10330 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10331 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10332 int NumElts = VT.getVectorNumElements();
10333 bool ShufpdMask = true;
10334 bool CommutableMask = true;
10335 unsigned Immediate = 0;
10336 for (int i = 0; i < NumElts; ++i) {
10339 int Val = (i & 6) + NumElts * (i & 1);
10340 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10341 if (Mask[i] < Val || Mask[i] > Val + 1)
10342 ShufpdMask = false;
10343 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10344 CommutableMask = false;
10345 Immediate |= (Mask[i] % 2) << i;
10348 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10349 DAG.getConstant(Immediate, DL, MVT::i8));
10350 if (CommutableMask)
10351 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10352 DAG.getConstant(Immediate, DL, MVT::i8));
10356 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10358 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10359 /// isn't available.
10360 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10361 const X86Subtarget *Subtarget,
10362 SelectionDAG &DAG) {
10364 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10365 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10366 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10367 ArrayRef<int> Mask = SVOp->getMask();
10368 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10370 SmallVector<int, 4> WidenedMask;
10371 if (canWidenShuffleElements(Mask, WidenedMask))
10372 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10375 if (isSingleInputShuffleMask(Mask)) {
10376 // Check for being able to broadcast a single element.
10377 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10378 Mask, Subtarget, DAG))
10381 // Use low duplicate instructions for masks that match their pattern.
10382 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10383 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10385 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10386 // Non-half-crossing single input shuffles can be lowerid with an
10387 // interleaved permutation.
10388 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10389 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10390 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10391 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10394 // With AVX2 we have direct support for this permutation.
10395 if (Subtarget->hasAVX2())
10396 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10397 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10399 // Otherwise, fall back.
10400 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10404 // Use dedicated unpack instructions for masks that match their pattern.
10406 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10409 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10413 // Check if the blend happens to exactly fit that of SHUFPD.
10415 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10418 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10419 // shuffle. However, if we have AVX2 and either inputs are already in place,
10420 // we will be able to shuffle even across lanes the other input in a single
10421 // instruction so skip this pattern.
10422 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10423 isShuffleMaskInputInPlace(1, Mask))))
10424 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10425 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10428 // If we have AVX2 then we always want to lower with a blend because an v4 we
10429 // can fully permute the elements.
10430 if (Subtarget->hasAVX2())
10431 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10434 // Otherwise fall back on generic lowering.
10435 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10438 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10440 /// This routine is only called when we have AVX2 and thus a reasonable
10441 /// instruction set for v4i64 shuffling..
10442 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10443 const X86Subtarget *Subtarget,
10444 SelectionDAG &DAG) {
10446 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10447 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10448 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10449 ArrayRef<int> Mask = SVOp->getMask();
10450 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10451 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10453 SmallVector<int, 4> WidenedMask;
10454 if (canWidenShuffleElements(Mask, WidenedMask))
10455 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10458 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10462 // Check for being able to broadcast a single element.
10463 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10464 Mask, Subtarget, DAG))
10467 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10468 // use lower latency instructions that will operate on both 128-bit lanes.
10469 SmallVector<int, 2> RepeatedMask;
10470 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10471 if (isSingleInputShuffleMask(Mask)) {
10472 int PSHUFDMask[] = {-1, -1, -1, -1};
10473 for (int i = 0; i < 2; ++i)
10474 if (RepeatedMask[i] >= 0) {
10475 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10476 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10478 return DAG.getBitcast(
10480 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10481 DAG.getBitcast(MVT::v8i32, V1),
10482 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10486 // AVX2 provides a direct instruction for permuting a single input across
10488 if (isSingleInputShuffleMask(Mask))
10489 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10490 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10492 // Try to use shift instructions.
10493 if (SDValue Shift =
10494 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10497 // Use dedicated unpack instructions for masks that match their pattern.
10499 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10502 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10503 // shuffle. However, if we have AVX2 and either inputs are already in place,
10504 // we will be able to shuffle even across lanes the other input in a single
10505 // instruction so skip this pattern.
10506 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10507 isShuffleMaskInputInPlace(1, Mask))))
10508 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10509 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10512 // Otherwise fall back on generic blend lowering.
10513 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10517 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10519 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10520 /// isn't available.
10521 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10522 const X86Subtarget *Subtarget,
10523 SelectionDAG &DAG) {
10525 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10526 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10527 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10528 ArrayRef<int> Mask = SVOp->getMask();
10529 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10531 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10535 // Check for being able to broadcast a single element.
10536 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10537 Mask, Subtarget, DAG))
10540 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10541 // options to efficiently lower the shuffle.
10542 SmallVector<int, 4> RepeatedMask;
10543 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10544 assert(RepeatedMask.size() == 4 &&
10545 "Repeated masks must be half the mask width!");
10547 // Use even/odd duplicate instructions for masks that match their pattern.
10548 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10549 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10550 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10551 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10553 if (isSingleInputShuffleMask(Mask))
10554 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10555 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10557 // Use dedicated unpack instructions for masks that match their pattern.
10559 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10562 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10563 // have already handled any direct blends. We also need to squash the
10564 // repeated mask into a simulated v4f32 mask.
10565 for (int i = 0; i < 4; ++i)
10566 if (RepeatedMask[i] >= 8)
10567 RepeatedMask[i] -= 4;
10568 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10571 // If we have a single input shuffle with different shuffle patterns in the
10572 // two 128-bit lanes use the variable mask to VPERMILPS.
10573 if (isSingleInputShuffleMask(Mask)) {
10574 SDValue VPermMask[8];
10575 for (int i = 0; i < 8; ++i)
10576 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10577 : DAG.getConstant(Mask[i], DL, MVT::i32);
10578 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10579 return DAG.getNode(
10580 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10581 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10583 if (Subtarget->hasAVX2())
10584 return DAG.getNode(
10585 X86ISD::VPERMV, DL, MVT::v8f32,
10586 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10588 // Otherwise, fall back.
10589 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10593 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10595 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10596 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10599 // If we have AVX2 then we always want to lower with a blend because at v8 we
10600 // can fully permute the elements.
10601 if (Subtarget->hasAVX2())
10602 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10605 // Otherwise fall back on generic lowering.
10606 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10609 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10611 /// This routine is only called when we have AVX2 and thus a reasonable
10612 /// instruction set for v8i32 shuffling..
10613 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10614 const X86Subtarget *Subtarget,
10615 SelectionDAG &DAG) {
10617 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10618 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10619 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10620 ArrayRef<int> Mask = SVOp->getMask();
10621 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10622 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10624 // Whenever we can lower this as a zext, that instruction is strictly faster
10625 // than any alternative. It also allows us to fold memory operands into the
10626 // shuffle in many cases.
10627 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10628 Mask, Subtarget, DAG))
10631 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10635 // Check for being able to broadcast a single element.
10636 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10637 Mask, Subtarget, DAG))
10640 // If the shuffle mask is repeated in each 128-bit lane we can use more
10641 // efficient instructions that mirror the shuffles across the two 128-bit
10643 SmallVector<int, 4> RepeatedMask;
10644 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10645 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10646 if (isSingleInputShuffleMask(Mask))
10647 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10648 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10650 // Use dedicated unpack instructions for masks that match their pattern.
10652 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10656 // Try to use shift instructions.
10657 if (SDValue Shift =
10658 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10661 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10662 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10665 // If the shuffle patterns aren't repeated but it is a single input, directly
10666 // generate a cross-lane VPERMD instruction.
10667 if (isSingleInputShuffleMask(Mask)) {
10668 SDValue VPermMask[8];
10669 for (int i = 0; i < 8; ++i)
10670 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10671 : DAG.getConstant(Mask[i], DL, MVT::i32);
10672 return DAG.getNode(
10673 X86ISD::VPERMV, DL, MVT::v8i32,
10674 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10677 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10679 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10680 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10683 // Otherwise fall back on generic blend lowering.
10684 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10688 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10690 /// This routine is only called when we have AVX2 and thus a reasonable
10691 /// instruction set for v16i16 shuffling..
10692 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10693 const X86Subtarget *Subtarget,
10694 SelectionDAG &DAG) {
10696 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10697 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10698 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10699 ArrayRef<int> Mask = SVOp->getMask();
10700 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10701 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10703 // Whenever we can lower this as a zext, that instruction is strictly faster
10704 // than any alternative. It also allows us to fold memory operands into the
10705 // shuffle in many cases.
10706 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10707 Mask, Subtarget, DAG))
10710 // Check for being able to broadcast a single element.
10711 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10712 Mask, Subtarget, DAG))
10715 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10719 // Use dedicated unpack instructions for masks that match their pattern.
10721 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10724 // Try to use shift instructions.
10725 if (SDValue Shift =
10726 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10729 // Try to use byte rotation instructions.
10730 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10731 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10734 if (isSingleInputShuffleMask(Mask)) {
10735 // There are no generalized cross-lane shuffle operations available on i16
10737 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10738 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10741 SmallVector<int, 8> RepeatedMask;
10742 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10743 // As this is a single-input shuffle, the repeated mask should be
10744 // a strictly valid v8i16 mask that we can pass through to the v8i16
10745 // lowering to handle even the v16 case.
10746 return lowerV8I16GeneralSingleInputVectorShuffle(
10747 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10750 SDValue PSHUFBMask[32];
10751 for (int i = 0; i < 16; ++i) {
10752 if (Mask[i] == -1) {
10753 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10757 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10758 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10759 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10760 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10762 return DAG.getBitcast(MVT::v16i16,
10763 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10764 DAG.getBitcast(MVT::v32i8, V1),
10765 DAG.getNode(ISD::BUILD_VECTOR, DL,
10766 MVT::v32i8, PSHUFBMask)));
10769 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10771 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10772 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10775 // Otherwise fall back on generic lowering.
10776 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10779 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10781 /// This routine is only called when we have AVX2 and thus a reasonable
10782 /// instruction set for v32i8 shuffling..
10783 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10784 const X86Subtarget *Subtarget,
10785 SelectionDAG &DAG) {
10787 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10788 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10789 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10790 ArrayRef<int> Mask = SVOp->getMask();
10791 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10792 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10794 // Whenever we can lower this as a zext, that instruction is strictly faster
10795 // than any alternative. It also allows us to fold memory operands into the
10796 // shuffle in many cases.
10797 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10798 Mask, Subtarget, DAG))
10801 // Check for being able to broadcast a single element.
10802 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10803 Mask, Subtarget, DAG))
10806 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10810 // Use dedicated unpack instructions for masks that match their pattern.
10812 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10815 // Try to use shift instructions.
10816 if (SDValue Shift =
10817 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10820 // Try to use byte rotation instructions.
10821 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10822 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10825 if (isSingleInputShuffleMask(Mask)) {
10826 // There are no generalized cross-lane shuffle operations available on i8
10828 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10829 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10832 SDValue PSHUFBMask[32];
10833 for (int i = 0; i < 32; ++i)
10836 ? DAG.getUNDEF(MVT::i8)
10837 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10840 return DAG.getNode(
10841 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10842 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10845 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10847 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10848 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10851 // Otherwise fall back on generic lowering.
10852 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10855 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10857 /// This routine either breaks down the specific type of a 256-bit x86 vector
10858 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10859 /// together based on the available instructions.
10860 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10861 MVT VT, const X86Subtarget *Subtarget,
10862 SelectionDAG &DAG) {
10864 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10865 ArrayRef<int> Mask = SVOp->getMask();
10867 // If we have a single input to the zero element, insert that into V1 if we
10868 // can do so cheaply.
10869 int NumElts = VT.getVectorNumElements();
10870 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10871 return M >= NumElts;
10874 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10875 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10876 DL, VT, V1, V2, Mask, Subtarget, DAG))
10879 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10880 // can check for those subtargets here and avoid much of the subtarget
10881 // querying in the per-vector-type lowering routines. With AVX1 we have
10882 // essentially *zero* ability to manipulate a 256-bit vector with integer
10883 // types. Since we'll use floating point types there eventually, just
10884 // immediately cast everything to a float and operate entirely in that domain.
10885 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10886 int ElementBits = VT.getScalarSizeInBits();
10887 if (ElementBits < 32)
10888 // No floating point type available, decompose into 128-bit vectors.
10889 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10891 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10892 VT.getVectorNumElements());
10893 V1 = DAG.getBitcast(FpVT, V1);
10894 V2 = DAG.getBitcast(FpVT, V2);
10895 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10898 switch (VT.SimpleTy) {
10900 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10902 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10904 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10906 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10908 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10910 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10913 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10917 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10918 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10919 ArrayRef<int> Mask,
10920 SDValue V1, SDValue V2,
10921 SelectionDAG &DAG) {
10922 assert(VT.getScalarSizeInBits() == 64 &&
10923 "Unexpected element type size for 128bit shuffle.");
10925 // To handle 256 bit vector requires VLX and most probably
10926 // function lowerV2X128VectorShuffle() is better solution.
10927 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10929 SmallVector<int, 4> WidenedMask;
10930 if (!canWidenShuffleElements(Mask, WidenedMask))
10933 // Form a 128-bit permutation.
10934 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10935 // bits defined by a vshuf64x2 instruction's immediate control byte.
10936 unsigned PermMask = 0, Imm = 0;
10937 unsigned ControlBitsNum = WidenedMask.size() / 2;
10939 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10940 if (WidenedMask[i] == SM_SentinelZero)
10943 // Use first element in place of undef mask.
10944 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10945 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10948 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10949 DAG.getConstant(PermMask, DL, MVT::i8));
10952 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10953 ArrayRef<int> Mask, SDValue V1,
10954 SDValue V2, SelectionDAG &DAG) {
10956 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10958 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10959 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10961 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10962 if (isSingleInputShuffleMask(Mask))
10963 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10965 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10968 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10969 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10970 const X86Subtarget *Subtarget,
10971 SelectionDAG &DAG) {
10973 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10974 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10975 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10976 ArrayRef<int> Mask = SVOp->getMask();
10977 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10979 if (SDValue Shuf128 =
10980 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10983 if (SDValue Unpck =
10984 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10987 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10990 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10991 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10992 const X86Subtarget *Subtarget,
10993 SelectionDAG &DAG) {
10995 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10996 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10997 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10998 ArrayRef<int> Mask = SVOp->getMask();
10999 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11001 if (SDValue Unpck =
11002 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
11005 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
11008 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11009 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11010 const X86Subtarget *Subtarget,
11011 SelectionDAG &DAG) {
11013 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11014 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11015 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11016 ArrayRef<int> Mask = SVOp->getMask();
11017 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11019 if (SDValue Shuf128 =
11020 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
11023 if (SDValue Unpck =
11024 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
11027 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
11030 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11031 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11032 const X86Subtarget *Subtarget,
11033 SelectionDAG &DAG) {
11035 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11036 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11037 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11038 ArrayRef<int> Mask = SVOp->getMask();
11039 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11041 if (SDValue Unpck =
11042 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
11045 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
11048 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11049 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11050 const X86Subtarget *Subtarget,
11051 SelectionDAG &DAG) {
11053 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11054 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11055 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11056 ArrayRef<int> Mask = SVOp->getMask();
11057 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11058 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11060 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
11063 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11064 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11065 const X86Subtarget *Subtarget,
11066 SelectionDAG &DAG) {
11068 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11069 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11070 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11071 ArrayRef<int> Mask = SVOp->getMask();
11072 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11073 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11075 // FIXME: Implement direct support for this type!
11076 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11079 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11081 /// This routine either breaks down the specific type of a 512-bit x86 vector
11082 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11083 /// together based on the available instructions.
11084 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11085 MVT VT, const X86Subtarget *Subtarget,
11086 SelectionDAG &DAG) {
11088 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11089 ArrayRef<int> Mask = SVOp->getMask();
11090 assert(Subtarget->hasAVX512() &&
11091 "Cannot lower 512-bit vectors w/ basic ISA!");
11093 // Check for being able to broadcast a single element.
11094 if (SDValue Broadcast =
11095 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
11098 // Dispatch to each element type for lowering. If we don't have supprot for
11099 // specific element type shuffles at 512 bits, immediately split them and
11100 // lower them. Each lowering routine of a given type is allowed to assume that
11101 // the requisite ISA extensions for that element type are available.
11102 switch (VT.SimpleTy) {
11104 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11106 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11108 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11110 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11112 if (Subtarget->hasBWI())
11113 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11116 if (Subtarget->hasBWI())
11117 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11121 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11124 // Otherwise fall back on splitting.
11125 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11128 // Lower vXi1 vector shuffles.
11129 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11130 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11131 // vector, shuffle and then truncate it back.
11132 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11133 MVT VT, const X86Subtarget *Subtarget,
11134 SelectionDAG &DAG) {
11136 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11137 ArrayRef<int> Mask = SVOp->getMask();
11138 assert(Subtarget->hasAVX512() &&
11139 "Cannot lower 512-bit vectors w/o basic ISA!");
11141 switch (VT.SimpleTy) {
11143 llvm_unreachable("Expected a vector of i1 elements");
11145 ExtVT = MVT::v2i64;
11148 ExtVT = MVT::v4i32;
11151 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11154 ExtVT = MVT::v16i32;
11157 ExtVT = MVT::v32i16;
11160 ExtVT = MVT::v64i8;
11164 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11165 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11166 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11167 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11169 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11172 V2 = DAG.getUNDEF(ExtVT);
11173 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11174 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11175 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11176 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11178 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11179 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11180 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11182 /// \brief Top-level lowering for x86 vector shuffles.
11184 /// This handles decomposition, canonicalization, and lowering of all x86
11185 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11186 /// above in helper routines. The canonicalization attempts to widen shuffles
11187 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11188 /// s.t. only one of the two inputs needs to be tested, etc.
11189 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11190 SelectionDAG &DAG) {
11191 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11192 ArrayRef<int> Mask = SVOp->getMask();
11193 SDValue V1 = Op.getOperand(0);
11194 SDValue V2 = Op.getOperand(1);
11195 MVT VT = Op.getSimpleValueType();
11196 int NumElements = VT.getVectorNumElements();
11198 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11200 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11201 "Can't lower MMX shuffles");
11203 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11204 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11205 if (V1IsUndef && V2IsUndef)
11206 return DAG.getUNDEF(VT);
11208 // When we create a shuffle node we put the UNDEF node to second operand,
11209 // but in some cases the first operand may be transformed to UNDEF.
11210 // In this case we should just commute the node.
11212 return DAG.getCommutedVectorShuffle(*SVOp);
11214 // Check for non-undef masks pointing at an undef vector and make the masks
11215 // undef as well. This makes it easier to match the shuffle based solely on
11219 if (M >= NumElements) {
11220 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11221 for (int &M : NewMask)
11222 if (M >= NumElements)
11224 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11227 // We actually see shuffles that are entirely re-arrangements of a set of
11228 // zero inputs. This mostly happens while decomposing complex shuffles into
11229 // simple ones. Directly lower these as a buildvector of zeros.
11230 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11231 if (Zeroable.all())
11232 return getZeroVector(VT, Subtarget, DAG, dl);
11234 // Try to collapse shuffles into using a vector type with fewer elements but
11235 // wider element types. We cap this to not form integers or floating point
11236 // elements wider than 64 bits, but it might be interesting to form i128
11237 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11238 SmallVector<int, 16> WidenedMask;
11239 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11240 canWidenShuffleElements(Mask, WidenedMask)) {
11241 MVT NewEltVT = VT.isFloatingPoint()
11242 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11243 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11244 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11245 // Make sure that the new vector type is legal. For example, v2f64 isn't
11247 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11248 V1 = DAG.getBitcast(NewVT, V1);
11249 V2 = DAG.getBitcast(NewVT, V2);
11250 return DAG.getBitcast(
11251 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11255 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11256 for (int M : SVOp->getMask())
11258 ++NumUndefElements;
11259 else if (M < NumElements)
11264 // Commute the shuffle as needed such that more elements come from V1 than
11265 // V2. This allows us to match the shuffle pattern strictly on how many
11266 // elements come from V1 without handling the symmetric cases.
11267 if (NumV2Elements > NumV1Elements)
11268 return DAG.getCommutedVectorShuffle(*SVOp);
11270 // When the number of V1 and V2 elements are the same, try to minimize the
11271 // number of uses of V2 in the low half of the vector. When that is tied,
11272 // ensure that the sum of indices for V1 is equal to or lower than the sum
11273 // indices for V2. When those are equal, try to ensure that the number of odd
11274 // indices for V1 is lower than the number of odd indices for V2.
11275 if (NumV1Elements == NumV2Elements) {
11276 int LowV1Elements = 0, LowV2Elements = 0;
11277 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11278 if (M >= NumElements)
11282 if (LowV2Elements > LowV1Elements) {
11283 return DAG.getCommutedVectorShuffle(*SVOp);
11284 } else if (LowV2Elements == LowV1Elements) {
11285 int SumV1Indices = 0, SumV2Indices = 0;
11286 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11287 if (SVOp->getMask()[i] >= NumElements)
11289 else if (SVOp->getMask()[i] >= 0)
11291 if (SumV2Indices < SumV1Indices) {
11292 return DAG.getCommutedVectorShuffle(*SVOp);
11293 } else if (SumV2Indices == SumV1Indices) {
11294 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11295 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11296 if (SVOp->getMask()[i] >= NumElements)
11297 NumV2OddIndices += i % 2;
11298 else if (SVOp->getMask()[i] >= 0)
11299 NumV1OddIndices += i % 2;
11300 if (NumV2OddIndices < NumV1OddIndices)
11301 return DAG.getCommutedVectorShuffle(*SVOp);
11306 // For each vector width, delegate to a specialized lowering routine.
11307 if (VT.is128BitVector())
11308 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11310 if (VT.is256BitVector())
11311 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11313 if (VT.is512BitVector())
11314 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11317 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11318 llvm_unreachable("Unimplemented!");
11321 // This function assumes its argument is a BUILD_VECTOR of constants or
11322 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11324 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11325 unsigned &MaskValue) {
11327 unsigned NumElems = BuildVector->getNumOperands();
11329 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11330 // We don't handle the >2 lanes case right now.
11331 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11335 unsigned NumElemsInLane = NumElems / NumLanes;
11337 // Blend for v16i16 should be symmetric for the both lanes.
11338 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11339 SDValue EltCond = BuildVector->getOperand(i);
11340 SDValue SndLaneEltCond =
11341 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11343 int Lane1Cond = -1, Lane2Cond = -1;
11344 if (isa<ConstantSDNode>(EltCond))
11345 Lane1Cond = !isNullConstant(EltCond);
11346 if (isa<ConstantSDNode>(SndLaneEltCond))
11347 Lane2Cond = !isNullConstant(SndLaneEltCond);
11349 unsigned LaneMask = 0;
11350 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11351 // Lane1Cond != 0, means we want the first argument.
11352 // Lane1Cond == 0, means we want the second argument.
11353 // The encoding of this argument is 0 for the first argument, 1
11354 // for the second. Therefore, invert the condition.
11355 LaneMask = !Lane1Cond << i;
11356 else if (Lane1Cond < 0)
11357 LaneMask = !Lane2Cond << i;
11361 MaskValue |= LaneMask;
11363 MaskValue |= LaneMask << NumElemsInLane;
11368 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11369 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11370 const X86Subtarget *Subtarget,
11371 SelectionDAG &DAG) {
11372 SDValue Cond = Op.getOperand(0);
11373 SDValue LHS = Op.getOperand(1);
11374 SDValue RHS = Op.getOperand(2);
11376 MVT VT = Op.getSimpleValueType();
11378 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11380 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11382 // Only non-legal VSELECTs reach this lowering, convert those into generic
11383 // shuffles and re-use the shuffle lowering path for blends.
11384 SmallVector<int, 32> Mask;
11385 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11386 SDValue CondElt = CondBV->getOperand(i);
11388 isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
11391 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11394 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11395 // A vselect where all conditions and data are constants can be optimized into
11396 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11397 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11398 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11399 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11402 // Try to lower this to a blend-style vector shuffle. This can handle all
11403 // constant condition cases.
11404 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11407 // Variable blends are only legal from SSE4.1 onward.
11408 if (!Subtarget->hasSSE41())
11411 // Only some types will be legal on some subtargets. If we can emit a legal
11412 // VSELECT-matching blend, return Op, and but if we need to expand, return
11414 switch (Op.getSimpleValueType().SimpleTy) {
11416 // Most of the vector types have blends past SSE4.1.
11420 // The byte blends for AVX vectors were introduced only in AVX2.
11421 if (Subtarget->hasAVX2())
11428 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11429 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11432 // FIXME: We should custom lower this by fixing the condition and using i8
11438 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11439 MVT VT = Op.getSimpleValueType();
11442 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11445 if (VT.getSizeInBits() == 8) {
11446 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11447 Op.getOperand(0), Op.getOperand(1));
11448 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11449 DAG.getValueType(VT));
11450 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11453 if (VT.getSizeInBits() == 16) {
11454 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11455 if (isNullConstant(Op.getOperand(1)))
11456 return DAG.getNode(
11457 ISD::TRUNCATE, dl, MVT::i16,
11458 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11459 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11460 Op.getOperand(1)));
11461 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11462 Op.getOperand(0), Op.getOperand(1));
11463 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11464 DAG.getValueType(VT));
11465 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11468 if (VT == MVT::f32) {
11469 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11470 // the result back to FR32 register. It's only worth matching if the
11471 // result has a single use which is a store or a bitcast to i32. And in
11472 // the case of a store, it's not worth it if the index is a constant 0,
11473 // because a MOVSSmr can be used instead, which is smaller and faster.
11474 if (!Op.hasOneUse())
11476 SDNode *User = *Op.getNode()->use_begin();
11477 if ((User->getOpcode() != ISD::STORE ||
11478 isNullConstant(Op.getOperand(1))) &&
11479 (User->getOpcode() != ISD::BITCAST ||
11480 User->getValueType(0) != MVT::i32))
11482 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11483 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11485 return DAG.getBitcast(MVT::f32, Extract);
11488 if (VT == MVT::i32 || VT == MVT::i64) {
11489 // ExtractPS/pextrq works with constant index.
11490 if (isa<ConstantSDNode>(Op.getOperand(1)))
11496 /// Extract one bit from mask vector, like v16i1 or v8i1.
11497 /// AVX-512 feature.
11499 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11500 SDValue Vec = Op.getOperand(0);
11502 MVT VecVT = Vec.getSimpleValueType();
11503 SDValue Idx = Op.getOperand(1);
11504 MVT EltVT = Op.getSimpleValueType();
11506 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11507 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11508 "Unexpected vector type in ExtractBitFromMaskVector");
11510 // variable index can't be handled in mask registers,
11511 // extend vector to VR512
11512 if (!isa<ConstantSDNode>(Idx)) {
11513 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11514 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11515 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11516 ExtVT.getVectorElementType(), Ext, Idx);
11517 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11520 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11521 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11522 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11523 rc = getRegClassFor(MVT::v16i1);
11524 unsigned MaxSift = rc->getSize()*8 - 1;
11525 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11526 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11527 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11528 DAG.getConstant(MaxSift, dl, MVT::i8));
11529 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11530 DAG.getIntPtrConstant(0, dl));
11534 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11535 SelectionDAG &DAG) const {
11537 SDValue Vec = Op.getOperand(0);
11538 MVT VecVT = Vec.getSimpleValueType();
11539 SDValue Idx = Op.getOperand(1);
11541 if (Op.getSimpleValueType() == MVT::i1)
11542 return ExtractBitFromMaskVector(Op, DAG);
11544 if (!isa<ConstantSDNode>(Idx)) {
11545 if (VecVT.is512BitVector() ||
11546 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11547 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11550 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11551 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11552 MaskEltVT.getSizeInBits());
11554 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11555 auto PtrVT = getPointerTy(DAG.getDataLayout());
11556 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11557 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11558 DAG.getConstant(0, dl, PtrVT));
11559 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11560 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11561 DAG.getConstant(0, dl, PtrVT));
11566 // If this is a 256-bit vector result, first extract the 128-bit vector and
11567 // then extract the element from the 128-bit vector.
11568 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11570 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11571 // Get the 128-bit vector.
11572 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11573 MVT EltVT = VecVT.getVectorElementType();
11575 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11576 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11578 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11579 // this can be done with a mask.
11580 IdxVal &= ElemsPerChunk - 1;
11581 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11582 DAG.getConstant(IdxVal, dl, MVT::i32));
11585 assert(VecVT.is128BitVector() && "Unexpected vector length");
11587 if (Subtarget->hasSSE41())
11588 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11591 MVT VT = Op.getSimpleValueType();
11592 // TODO: handle v16i8.
11593 if (VT.getSizeInBits() == 16) {
11594 SDValue Vec = Op.getOperand(0);
11595 if (isNullConstant(Op.getOperand(1)))
11596 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11597 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11598 DAG.getBitcast(MVT::v4i32, Vec),
11599 Op.getOperand(1)));
11600 // Transform it so it match pextrw which produces a 32-bit result.
11601 MVT EltVT = MVT::i32;
11602 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11603 Op.getOperand(0), Op.getOperand(1));
11604 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11605 DAG.getValueType(VT));
11606 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11609 if (VT.getSizeInBits() == 32) {
11610 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11614 // SHUFPS the element to the lowest double word, then movss.
11615 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11616 MVT VVT = Op.getOperand(0).getSimpleValueType();
11617 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11618 DAG.getUNDEF(VVT), Mask);
11619 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11620 DAG.getIntPtrConstant(0, dl));
11623 if (VT.getSizeInBits() == 64) {
11624 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11625 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11626 // to match extract_elt for f64.
11627 if (isNullConstant(Op.getOperand(1)))
11630 // UNPCKHPD the element to the lowest double word, then movsd.
11631 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11632 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11633 int Mask[2] = { 1, -1 };
11634 MVT VVT = Op.getOperand(0).getSimpleValueType();
11635 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11636 DAG.getUNDEF(VVT), Mask);
11637 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11638 DAG.getIntPtrConstant(0, dl));
11644 /// Insert one bit to mask vector, like v16i1 or v8i1.
11645 /// AVX-512 feature.
11647 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11649 SDValue Vec = Op.getOperand(0);
11650 SDValue Elt = Op.getOperand(1);
11651 SDValue Idx = Op.getOperand(2);
11652 MVT VecVT = Vec.getSimpleValueType();
11654 if (!isa<ConstantSDNode>(Idx)) {
11655 // Non constant index. Extend source and destination,
11656 // insert element and then truncate the result.
11657 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11658 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11659 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11660 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11661 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11662 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11665 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11666 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11668 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11669 DAG.getConstant(IdxVal, dl, MVT::i8));
11670 if (Vec.getOpcode() == ISD::UNDEF)
11672 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11675 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11676 SelectionDAG &DAG) const {
11677 MVT VT = Op.getSimpleValueType();
11678 MVT EltVT = VT.getVectorElementType();
11680 if (EltVT == MVT::i1)
11681 return InsertBitToMaskVector(Op, DAG);
11684 SDValue N0 = Op.getOperand(0);
11685 SDValue N1 = Op.getOperand(1);
11686 SDValue N2 = Op.getOperand(2);
11687 if (!isa<ConstantSDNode>(N2))
11689 auto *N2C = cast<ConstantSDNode>(N2);
11690 unsigned IdxVal = N2C->getZExtValue();
11692 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11693 // into that, and then insert the subvector back into the result.
11694 if (VT.is256BitVector() || VT.is512BitVector()) {
11695 // With a 256-bit vector, we can insert into the zero element efficiently
11696 // using a blend if we have AVX or AVX2 and the right data type.
11697 if (VT.is256BitVector() && IdxVal == 0) {
11698 // TODO: It is worthwhile to cast integer to floating point and back
11699 // and incur a domain crossing penalty if that's what we'll end up
11700 // doing anyway after extracting to a 128-bit vector.
11701 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11702 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11703 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11704 N2 = DAG.getIntPtrConstant(1, dl);
11705 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11709 // Get the desired 128-bit vector chunk.
11710 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11712 // Insert the element into the desired chunk.
11713 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11714 assert(isPowerOf2_32(NumEltsIn128));
11715 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11716 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11718 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11719 DAG.getConstant(IdxIn128, dl, MVT::i32));
11721 // Insert the changed part back into the bigger vector
11722 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11724 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11726 if (Subtarget->hasSSE41()) {
11727 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11729 if (VT == MVT::v8i16) {
11730 Opc = X86ISD::PINSRW;
11732 assert(VT == MVT::v16i8);
11733 Opc = X86ISD::PINSRB;
11736 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11738 if (N1.getValueType() != MVT::i32)
11739 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11740 if (N2.getValueType() != MVT::i32)
11741 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11742 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11745 if (EltVT == MVT::f32) {
11746 // Bits [7:6] of the constant are the source select. This will always be
11747 // zero here. The DAG Combiner may combine an extract_elt index into
11748 // these bits. For example (insert (extract, 3), 2) could be matched by
11749 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11750 // Bits [5:4] of the constant are the destination select. This is the
11751 // value of the incoming immediate.
11752 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11753 // combine either bitwise AND or insert of float 0.0 to set these bits.
11755 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11756 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11757 // If this is an insertion of 32-bits into the low 32-bits of
11758 // a vector, we prefer to generate a blend with immediate rather
11759 // than an insertps. Blends are simpler operations in hardware and so
11760 // will always have equal or better performance than insertps.
11761 // But if optimizing for size and there's a load folding opportunity,
11762 // generate insertps because blendps does not have a 32-bit memory
11764 N2 = DAG.getIntPtrConstant(1, dl);
11765 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11766 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11768 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11769 // Create this as a scalar to vector..
11770 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11771 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11774 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11775 // PINSR* works with constant index.
11780 if (EltVT == MVT::i8)
11783 if (EltVT.getSizeInBits() == 16) {
11784 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11785 // as its second argument.
11786 if (N1.getValueType() != MVT::i32)
11787 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11788 if (N2.getValueType() != MVT::i32)
11789 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11790 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11795 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11797 MVT OpVT = Op.getSimpleValueType();
11799 // If this is a 256-bit vector result, first insert into a 128-bit
11800 // vector and then insert into the 256-bit vector.
11801 if (!OpVT.is128BitVector()) {
11802 // Insert into a 128-bit vector.
11803 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11804 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11805 OpVT.getVectorNumElements() / SizeFactor);
11807 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11809 // Insert the 128-bit vector.
11810 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11813 if (OpVT == MVT::v1i64 &&
11814 Op.getOperand(0).getValueType() == MVT::i64)
11815 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11817 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11818 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11819 return DAG.getBitcast(
11820 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11823 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11824 // a simple subregister reference or explicit instructions to grab
11825 // upper bits of a vector.
11826 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11827 SelectionDAG &DAG) {
11829 SDValue In = Op.getOperand(0);
11830 SDValue Idx = Op.getOperand(1);
11831 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11832 MVT ResVT = Op.getSimpleValueType();
11833 MVT InVT = In.getSimpleValueType();
11835 if (Subtarget->hasFp256()) {
11836 if (ResVT.is128BitVector() &&
11837 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11838 isa<ConstantSDNode>(Idx)) {
11839 return Extract128BitVector(In, IdxVal, DAG, dl);
11841 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11842 isa<ConstantSDNode>(Idx)) {
11843 return Extract256BitVector(In, IdxVal, DAG, dl);
11849 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11850 // simple superregister reference or explicit instructions to insert
11851 // the upper bits of a vector.
11852 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11853 SelectionDAG &DAG) {
11854 if (!Subtarget->hasAVX())
11858 SDValue Vec = Op.getOperand(0);
11859 SDValue SubVec = Op.getOperand(1);
11860 SDValue Idx = Op.getOperand(2);
11862 if (!isa<ConstantSDNode>(Idx))
11865 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11866 MVT OpVT = Op.getSimpleValueType();
11867 MVT SubVecVT = SubVec.getSimpleValueType();
11869 // Fold two 16-byte subvector loads into one 32-byte load:
11870 // (insert_subvector (insert_subvector undef, (load addr), 0),
11871 // (load addr + 16), Elts/2)
11873 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11874 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11875 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11876 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11877 if (Idx2 && Idx2->getZExtValue() == 0) {
11878 SDValue SubVec2 = Vec.getOperand(1);
11879 // If needed, look through a bitcast to get to the load.
11880 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11881 SubVec2 = SubVec2.getOperand(0);
11883 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11885 unsigned Alignment = FirstLd->getAlignment();
11886 unsigned AS = FirstLd->getAddressSpace();
11887 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11888 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11889 OpVT, AS, Alignment, &Fast) && Fast) {
11890 SDValue Ops[] = { SubVec2, SubVec };
11891 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11898 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11899 SubVecVT.is128BitVector())
11900 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11902 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11903 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11905 if (OpVT.getVectorElementType() == MVT::i1)
11906 return Insert1BitVector(Op, DAG);
11911 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11912 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11913 // one of the above mentioned nodes. It has to be wrapped because otherwise
11914 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11915 // be used to form addressing mode. These wrapped nodes will be selected
11918 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11919 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11921 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11922 // global base reg.
11923 unsigned char OpFlag = 0;
11924 unsigned WrapperKind = X86ISD::Wrapper;
11925 CodeModel::Model M = DAG.getTarget().getCodeModel();
11927 if (Subtarget->isPICStyleRIPRel() &&
11928 (M == CodeModel::Small || M == CodeModel::Kernel))
11929 WrapperKind = X86ISD::WrapperRIP;
11930 else if (Subtarget->isPICStyleGOT())
11931 OpFlag = X86II::MO_GOTOFF;
11932 else if (Subtarget->isPICStyleStubPIC())
11933 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11935 auto PtrVT = getPointerTy(DAG.getDataLayout());
11936 SDValue Result = DAG.getTargetConstantPool(
11937 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11939 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11940 // With PIC, the address is actually $g + Offset.
11943 DAG.getNode(ISD::ADD, DL, PtrVT,
11944 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11950 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11951 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11953 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11954 // global base reg.
11955 unsigned char OpFlag = 0;
11956 unsigned WrapperKind = X86ISD::Wrapper;
11957 CodeModel::Model M = DAG.getTarget().getCodeModel();
11959 if (Subtarget->isPICStyleRIPRel() &&
11960 (M == CodeModel::Small || M == CodeModel::Kernel))
11961 WrapperKind = X86ISD::WrapperRIP;
11962 else if (Subtarget->isPICStyleGOT())
11963 OpFlag = X86II::MO_GOTOFF;
11964 else if (Subtarget->isPICStyleStubPIC())
11965 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11967 auto PtrVT = getPointerTy(DAG.getDataLayout());
11968 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11970 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11972 // With PIC, the address is actually $g + Offset.
11975 DAG.getNode(ISD::ADD, DL, PtrVT,
11976 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11982 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11983 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11985 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11986 // global base reg.
11987 unsigned char OpFlag = 0;
11988 unsigned WrapperKind = X86ISD::Wrapper;
11989 CodeModel::Model M = DAG.getTarget().getCodeModel();
11991 if (Subtarget->isPICStyleRIPRel() &&
11992 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11993 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11994 OpFlag = X86II::MO_GOTPCREL;
11995 WrapperKind = X86ISD::WrapperRIP;
11996 } else if (Subtarget->isPICStyleGOT()) {
11997 OpFlag = X86II::MO_GOT;
11998 } else if (Subtarget->isPICStyleStubPIC()) {
11999 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
12000 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12001 OpFlag = X86II::MO_DARWIN_NONLAZY;
12004 auto PtrVT = getPointerTy(DAG.getDataLayout());
12005 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
12008 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12010 // With PIC, the address is actually $g + Offset.
12011 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12012 !Subtarget->is64Bit()) {
12014 DAG.getNode(ISD::ADD, DL, PtrVT,
12015 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12018 // For symbols that require a load from a stub to get the address, emit the
12020 if (isGlobalStubReference(OpFlag))
12021 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
12022 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12023 false, false, false, 0);
12029 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12030 // Create the TargetBlockAddressAddress node.
12031 unsigned char OpFlags =
12032 Subtarget->ClassifyBlockAddressReference();
12033 CodeModel::Model M = DAG.getTarget().getCodeModel();
12034 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12035 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12037 auto PtrVT = getPointerTy(DAG.getDataLayout());
12038 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
12040 if (Subtarget->isPICStyleRIPRel() &&
12041 (M == CodeModel::Small || M == CodeModel::Kernel))
12042 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12044 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12046 // With PIC, the address is actually $g + Offset.
12047 if (isGlobalRelativeToPICBase(OpFlags)) {
12048 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12049 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12056 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12057 int64_t Offset, SelectionDAG &DAG) const {
12058 // Create the TargetGlobalAddress node, folding in the constant
12059 // offset if it is legal.
12060 unsigned char OpFlags =
12061 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12062 CodeModel::Model M = DAG.getTarget().getCodeModel();
12063 auto PtrVT = getPointerTy(DAG.getDataLayout());
12065 if (OpFlags == X86II::MO_NO_FLAG &&
12066 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12067 // A direct static reference to a global.
12068 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
12071 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
12074 if (Subtarget->isPICStyleRIPRel() &&
12075 (M == CodeModel::Small || M == CodeModel::Kernel))
12076 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12078 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12080 // With PIC, the address is actually $g + Offset.
12081 if (isGlobalRelativeToPICBase(OpFlags)) {
12082 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12083 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12086 // For globals that require a load from a stub to get the address, emit the
12088 if (isGlobalStubReference(OpFlags))
12089 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12090 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12091 false, false, false, 0);
12093 // If there was a non-zero offset that we didn't fold, create an explicit
12094 // addition for it.
12096 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12097 DAG.getConstant(Offset, dl, PtrVT));
12103 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12104 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12105 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12106 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12110 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12111 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12112 unsigned char OperandFlags, bool LocalDynamic = false) {
12113 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12114 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12116 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12117 GA->getValueType(0),
12121 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12125 SDValue Ops[] = { Chain, TGA, *InFlag };
12126 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12128 SDValue Ops[] = { Chain, TGA };
12129 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12132 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12133 MFI->setAdjustsStack(true);
12134 MFI->setHasCalls(true);
12136 SDValue Flag = Chain.getValue(1);
12137 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12140 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12142 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12145 SDLoc dl(GA); // ? function entry point might be better
12146 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12147 DAG.getNode(X86ISD::GlobalBaseReg,
12148 SDLoc(), PtrVT), InFlag);
12149 InFlag = Chain.getValue(1);
12151 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12154 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12156 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12158 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12159 X86::RAX, X86II::MO_TLSGD);
12162 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12168 // Get the start address of the TLS block for this module.
12169 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12170 .getInfo<X86MachineFunctionInfo>();
12171 MFI->incNumLocalDynamicTLSAccesses();
12175 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12176 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12179 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12180 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12181 InFlag = Chain.getValue(1);
12182 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12183 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12186 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12190 unsigned char OperandFlags = X86II::MO_DTPOFF;
12191 unsigned WrapperKind = X86ISD::Wrapper;
12192 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12193 GA->getValueType(0),
12194 GA->getOffset(), OperandFlags);
12195 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12197 // Add x@dtpoff with the base.
12198 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12201 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12202 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12203 const EVT PtrVT, TLSModel::Model model,
12204 bool is64Bit, bool isPIC) {
12207 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12208 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12209 is64Bit ? 257 : 256));
12211 SDValue ThreadPointer =
12212 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12213 MachinePointerInfo(Ptr), false, false, false, 0);
12215 unsigned char OperandFlags = 0;
12216 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12218 unsigned WrapperKind = X86ISD::Wrapper;
12219 if (model == TLSModel::LocalExec) {
12220 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12221 } else if (model == TLSModel::InitialExec) {
12223 OperandFlags = X86II::MO_GOTTPOFF;
12224 WrapperKind = X86ISD::WrapperRIP;
12226 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12229 llvm_unreachable("Unexpected model");
12232 // emit "addl x@ntpoff,%eax" (local exec)
12233 // or "addl x@indntpoff,%eax" (initial exec)
12234 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12236 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12237 GA->getOffset(), OperandFlags);
12238 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12240 if (model == TLSModel::InitialExec) {
12241 if (isPIC && !is64Bit) {
12242 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12243 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12247 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12248 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12249 false, false, false, 0);
12252 // The address of the thread local variable is the add of the thread
12253 // pointer with the offset of the variable.
12254 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12258 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12260 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12262 // Cygwin uses emutls.
12263 // FIXME: It may be EmulatedTLS-generic also for X86-Android.
12264 if (Subtarget->isTargetWindowsCygwin())
12265 return LowerToTLSEmulatedModel(GA, DAG);
12267 const GlobalValue *GV = GA->getGlobal();
12268 auto PtrVT = getPointerTy(DAG.getDataLayout());
12270 if (Subtarget->isTargetELF()) {
12271 if (DAG.getTarget().Options.EmulatedTLS)
12272 return LowerToTLSEmulatedModel(GA, DAG);
12273 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12275 case TLSModel::GeneralDynamic:
12276 if (Subtarget->is64Bit())
12277 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12278 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12279 case TLSModel::LocalDynamic:
12280 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12281 Subtarget->is64Bit());
12282 case TLSModel::InitialExec:
12283 case TLSModel::LocalExec:
12284 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12285 DAG.getTarget().getRelocationModel() ==
12288 llvm_unreachable("Unknown TLS model.");
12291 if (Subtarget->isTargetDarwin()) {
12292 // Darwin only has one model of TLS. Lower to that.
12293 unsigned char OpFlag = 0;
12294 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12295 X86ISD::WrapperRIP : X86ISD::Wrapper;
12297 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12298 // global base reg.
12299 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12300 !Subtarget->is64Bit();
12302 OpFlag = X86II::MO_TLVP_PIC_BASE;
12304 OpFlag = X86II::MO_TLVP;
12306 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12307 GA->getValueType(0),
12308 GA->getOffset(), OpFlag);
12309 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12311 // With PIC32, the address is actually $g + Offset.
12313 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12314 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12317 // Lowering the machine isd will make sure everything is in the right
12319 SDValue Chain = DAG.getEntryNode();
12320 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12321 SDValue Args[] = { Chain, Offset };
12322 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12324 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12325 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12326 MFI->setAdjustsStack(true);
12328 // And our return value (tls address) is in the standard call return value
12330 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12331 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12334 if (Subtarget->isTargetKnownWindowsMSVC() ||
12335 Subtarget->isTargetWindowsGNU()) {
12336 // Just use the implicit TLS architecture
12337 // Need to generate someting similar to:
12338 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12340 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12341 // mov rcx, qword [rdx+rcx*8]
12342 // mov eax, .tls$:tlsvar
12343 // [rax+rcx] contains the address
12344 // Windows 64bit: gs:0x58
12345 // Windows 32bit: fs:__tls_array
12348 SDValue Chain = DAG.getEntryNode();
12350 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12351 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12352 // use its literal value of 0x2C.
12353 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12354 ? Type::getInt8PtrTy(*DAG.getContext(),
12356 : Type::getInt32PtrTy(*DAG.getContext(),
12359 SDValue TlsArray = Subtarget->is64Bit()
12360 ? DAG.getIntPtrConstant(0x58, dl)
12361 : (Subtarget->isTargetWindowsGNU()
12362 ? DAG.getIntPtrConstant(0x2C, dl)
12363 : DAG.getExternalSymbol("_tls_array", PtrVT));
12365 SDValue ThreadPointer =
12366 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12370 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12371 res = ThreadPointer;
12373 // Load the _tls_index variable
12374 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12375 if (Subtarget->is64Bit())
12376 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12377 MachinePointerInfo(), MVT::i32, false, false,
12380 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12383 auto &DL = DAG.getDataLayout();
12385 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12386 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12388 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12391 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12394 // Get the offset of start of .tls section
12395 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12396 GA->getValueType(0),
12397 GA->getOffset(), X86II::MO_SECREL);
12398 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12400 // The address of the thread local variable is the add of the thread
12401 // pointer with the offset of the variable.
12402 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12405 llvm_unreachable("TLS not implemented for this target.");
12408 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12409 /// and take a 2 x i32 value to shift plus a shift amount.
12410 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12411 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12412 MVT VT = Op.getSimpleValueType();
12413 unsigned VTBits = VT.getSizeInBits();
12415 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12416 SDValue ShOpLo = Op.getOperand(0);
12417 SDValue ShOpHi = Op.getOperand(1);
12418 SDValue ShAmt = Op.getOperand(2);
12419 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12420 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12422 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12423 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12424 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12425 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12426 : DAG.getConstant(0, dl, VT);
12428 SDValue Tmp2, Tmp3;
12429 if (Op.getOpcode() == ISD::SHL_PARTS) {
12430 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12431 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12433 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12434 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12437 // If the shift amount is larger or equal than the width of a part we can't
12438 // rely on the results of shld/shrd. Insert a test and select the appropriate
12439 // values for large shift amounts.
12440 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12441 DAG.getConstant(VTBits, dl, MVT::i8));
12442 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12443 AndNode, DAG.getConstant(0, dl, MVT::i8));
12446 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12447 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12448 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12450 if (Op.getOpcode() == ISD::SHL_PARTS) {
12451 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12452 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12454 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12455 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12458 SDValue Ops[2] = { Lo, Hi };
12459 return DAG.getMergeValues(Ops, dl);
12462 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12463 SelectionDAG &DAG) const {
12464 SDValue Src = Op.getOperand(0);
12465 MVT SrcVT = Src.getSimpleValueType();
12466 MVT VT = Op.getSimpleValueType();
12469 if (SrcVT.isVector()) {
12470 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12471 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12472 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12473 DAG.getUNDEF(SrcVT)));
12475 if (SrcVT.getVectorElementType() == MVT::i1) {
12476 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12477 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12478 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12483 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12484 "Unknown SINT_TO_FP to lower!");
12486 // These are really Legal; return the operand so the caller accepts it as
12488 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12490 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12491 Subtarget->is64Bit()) {
12495 unsigned Size = SrcVT.getSizeInBits()/8;
12496 MachineFunction &MF = DAG.getMachineFunction();
12497 auto PtrVT = getPointerTy(MF.getDataLayout());
12498 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12499 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12500 SDValue Chain = DAG.getStore(
12501 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12502 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12504 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12507 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12509 SelectionDAG &DAG) const {
12513 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12515 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12517 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12519 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12521 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12522 MachineMemOperand *MMO;
12524 int SSFI = FI->getIndex();
12525 MMO = DAG.getMachineFunction().getMachineMemOperand(
12526 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12527 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12529 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12530 StackSlot = StackSlot.getOperand(1);
12532 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12533 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12535 Tys, Ops, SrcVT, MMO);
12538 Chain = Result.getValue(1);
12539 SDValue InFlag = Result.getValue(2);
12541 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12542 // shouldn't be necessary except that RFP cannot be live across
12543 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12544 MachineFunction &MF = DAG.getMachineFunction();
12545 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12546 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12547 auto PtrVT = getPointerTy(MF.getDataLayout());
12548 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12549 Tys = DAG.getVTList(MVT::Other);
12551 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12553 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12554 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12555 MachineMemOperand::MOStore, SSFISize, SSFISize);
12557 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12558 Ops, Op.getValueType(), MMO);
12559 Result = DAG.getLoad(
12560 Op.getValueType(), DL, Chain, StackSlot,
12561 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12562 false, false, false, 0);
12568 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12569 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12570 SelectionDAG &DAG) const {
12571 // This algorithm is not obvious. Here it is what we're trying to output:
12574 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12575 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12577 haddpd %xmm0, %xmm0
12579 pshufd $0x4e, %xmm0, %xmm1
12585 LLVMContext *Context = DAG.getContext();
12587 // Build some magic constants.
12588 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12589 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12590 auto PtrVT = getPointerTy(DAG.getDataLayout());
12591 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12593 SmallVector<Constant*,2> CV1;
12595 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12596 APInt(64, 0x4330000000000000ULL))));
12598 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12599 APInt(64, 0x4530000000000000ULL))));
12600 Constant *C1 = ConstantVector::get(CV1);
12601 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12603 // Load the 64-bit value into an XMM register.
12604 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12607 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12608 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12609 false, false, false, 16);
12611 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12614 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12615 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12616 false, false, false, 16);
12617 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12618 // TODO: Are there any fast-math-flags to propagate here?
12619 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12622 if (Subtarget->hasSSE3()) {
12623 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12624 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12626 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12627 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12629 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12630 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12633 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12634 DAG.getIntPtrConstant(0, dl));
12637 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12638 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12639 SelectionDAG &DAG) const {
12641 // FP constant to bias correct the final result.
12642 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12645 // Load the 32-bit value into an XMM register.
12646 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12649 // Zero out the upper parts of the register.
12650 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12652 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12653 DAG.getBitcast(MVT::v2f64, Load),
12654 DAG.getIntPtrConstant(0, dl));
12656 // Or the load with the bias.
12657 SDValue Or = DAG.getNode(
12658 ISD::OR, dl, MVT::v2i64,
12659 DAG.getBitcast(MVT::v2i64,
12660 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12661 DAG.getBitcast(MVT::v2i64,
12662 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12664 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12665 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12667 // Subtract the bias.
12668 // TODO: Are there any fast-math-flags to propagate here?
12669 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12671 // Handle final rounding.
12672 MVT DestVT = Op.getSimpleValueType();
12674 if (DestVT.bitsLT(MVT::f64))
12675 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12676 DAG.getIntPtrConstant(0, dl));
12677 if (DestVT.bitsGT(MVT::f64))
12678 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12680 // Handle final rounding.
12684 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12685 const X86Subtarget &Subtarget) {
12686 // The algorithm is the following:
12687 // #ifdef __SSE4_1__
12688 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12689 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12690 // (uint4) 0x53000000, 0xaa);
12692 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12693 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12695 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12696 // return (float4) lo + fhi;
12698 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12699 // reassociate the two FADDs, and if we do that, the algorithm fails
12700 // spectacularly (PR24512).
12701 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12702 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12703 // there's also the MachineCombiner reassociations happening on Machine IR.
12704 if (DAG.getTarget().Options.UnsafeFPMath)
12708 SDValue V = Op->getOperand(0);
12709 MVT VecIntVT = V.getSimpleValueType();
12710 bool Is128 = VecIntVT == MVT::v4i32;
12711 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12712 // If we convert to something else than the supported type, e.g., to v4f64,
12714 if (VecFloatVT != Op->getSimpleValueType(0))
12717 unsigned NumElts = VecIntVT.getVectorNumElements();
12718 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12719 "Unsupported custom type");
12720 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12722 // In the #idef/#else code, we have in common:
12723 // - The vector of constants:
12729 // Create the splat vector for 0x4b000000.
12730 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12731 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12732 CstLow, CstLow, CstLow, CstLow};
12733 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12734 makeArrayRef(&CstLowArray[0], NumElts));
12735 // Create the splat vector for 0x53000000.
12736 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12737 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12738 CstHigh, CstHigh, CstHigh, CstHigh};
12739 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12740 makeArrayRef(&CstHighArray[0], NumElts));
12742 // Create the right shift.
12743 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12744 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12745 CstShift, CstShift, CstShift, CstShift};
12746 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12747 makeArrayRef(&CstShiftArray[0], NumElts));
12748 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12751 if (Subtarget.hasSSE41()) {
12752 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12753 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12754 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12755 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12756 // Low will be bitcasted right away, so do not bother bitcasting back to its
12758 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12759 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12760 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12761 // (uint4) 0x53000000, 0xaa);
12762 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12763 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12764 // High will be bitcasted right away, so do not bother bitcasting back to
12765 // its original type.
12766 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12767 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12769 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12770 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12771 CstMask, CstMask, CstMask);
12772 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12773 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12774 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12776 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12777 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12780 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12781 SDValue CstFAdd = DAG.getConstantFP(
12782 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12783 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12784 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12785 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12786 makeArrayRef(&CstFAddArray[0], NumElts));
12788 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12789 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12790 // TODO: Are there any fast-math-flags to propagate here?
12792 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12793 // return (float4) lo + fhi;
12794 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12795 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12798 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12799 SelectionDAG &DAG) const {
12800 SDValue N0 = Op.getOperand(0);
12801 MVT SVT = N0.getSimpleValueType();
12804 switch (SVT.SimpleTy) {
12806 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12811 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12812 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12813 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12817 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12820 assert(Subtarget->hasAVX512());
12821 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12822 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12826 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12827 SelectionDAG &DAG) const {
12828 SDValue N0 = Op.getOperand(0);
12830 auto PtrVT = getPointerTy(DAG.getDataLayout());
12832 if (Op.getSimpleValueType().isVector())
12833 return lowerUINT_TO_FP_vec(Op, DAG);
12835 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12836 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12837 // the optimization here.
12838 if (DAG.SignBitIsZero(N0))
12839 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12841 MVT SrcVT = N0.getSimpleValueType();
12842 MVT DstVT = Op.getSimpleValueType();
12844 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12845 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12846 // Conversions from unsigned i32 to f32/f64 are legal,
12847 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12851 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12852 return LowerUINT_TO_FP_i64(Op, DAG);
12853 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12854 return LowerUINT_TO_FP_i32(Op, DAG);
12855 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12858 // Make a 64-bit buffer, and use it to build an FILD.
12859 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12860 if (SrcVT == MVT::i32) {
12861 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12862 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12863 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12864 StackSlot, MachinePointerInfo(),
12866 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12867 OffsetSlot, MachinePointerInfo(),
12869 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12873 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12874 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12875 StackSlot, MachinePointerInfo(),
12877 // For i64 source, we need to add the appropriate power of 2 if the input
12878 // was negative. This is the same as the optimization in
12879 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12880 // we must be careful to do the computation in x87 extended precision, not
12881 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12882 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12883 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12884 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12885 MachineMemOperand::MOLoad, 8, 8);
12887 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12888 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12889 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12892 APInt FF(32, 0x5F800000ULL);
12894 // Check whether the sign bit is set.
12895 SDValue SignSet = DAG.getSetCC(
12896 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12897 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12899 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12900 SDValue FudgePtr = DAG.getConstantPool(
12901 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12903 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12904 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12905 SDValue Four = DAG.getIntPtrConstant(4, dl);
12906 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12908 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12910 // Load the value out, extending it from f32 to f80.
12911 // FIXME: Avoid the extend by constructing the right constant pool?
12912 SDValue Fudge = DAG.getExtLoad(
12913 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12914 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12915 false, false, false, 4);
12916 // Extend everything to 80 bits to force it to be done on x87.
12917 // TODO: Are there any fast-math-flags to propagate here?
12918 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12919 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12920 DAG.getIntPtrConstant(0, dl));
12923 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12924 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12925 // just return an <SDValue(), SDValue()> pair.
12926 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12927 // to i16, i32 or i64, and we lower it to a legal sequence.
12928 // If lowered to the final integer result we return a <result, SDValue()> pair.
12929 // Otherwise we lower it to a sequence ending with a FIST, return a
12930 // <FIST, StackSlot> pair, and the caller is responsible for loading
12931 // the final integer result from StackSlot.
12932 std::pair<SDValue,SDValue>
12933 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12934 bool IsSigned, bool IsReplace) const {
12937 EVT DstTy = Op.getValueType();
12938 EVT TheVT = Op.getOperand(0).getValueType();
12939 auto PtrVT = getPointerTy(DAG.getDataLayout());
12941 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12942 // f16 must be promoted before using the lowering in this routine.
12943 // fp128 does not use this lowering.
12944 return std::make_pair(SDValue(), SDValue());
12947 // If using FIST to compute an unsigned i64, we'll need some fixup
12948 // to handle values above the maximum signed i64. A FIST is always
12949 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12950 bool UnsignedFixup = !IsSigned &&
12951 DstTy == MVT::i64 &&
12952 (!Subtarget->is64Bit() ||
12953 !isScalarFPTypeInSSEReg(TheVT));
12955 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12956 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12957 // The low 32 bits of the fist result will have the correct uint32 result.
12958 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12962 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12963 DstTy.getSimpleVT() >= MVT::i16 &&
12964 "Unknown FP_TO_INT to lower!");
12966 // These are really Legal.
12967 if (DstTy == MVT::i32 &&
12968 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12969 return std::make_pair(SDValue(), SDValue());
12970 if (Subtarget->is64Bit() &&
12971 DstTy == MVT::i64 &&
12972 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12973 return std::make_pair(SDValue(), SDValue());
12975 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12977 MachineFunction &MF = DAG.getMachineFunction();
12978 unsigned MemSize = DstTy.getSizeInBits()/8;
12979 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12980 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12983 switch (DstTy.getSimpleVT().SimpleTy) {
12984 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12985 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12986 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12987 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12990 SDValue Chain = DAG.getEntryNode();
12991 SDValue Value = Op.getOperand(0);
12992 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12994 if (UnsignedFixup) {
12996 // Conversion to unsigned i64 is implemented with a select,
12997 // depending on whether the source value fits in the range
12998 // of a signed i64. Let Thresh be the FP equivalent of
12999 // 0x8000000000000000ULL.
13001 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
13002 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
13003 // Fist-to-mem64 FistSrc
13004 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
13005 // to XOR'ing the high 32 bits with Adjust.
13007 // Being a power of 2, Thresh is exactly representable in all FP formats.
13008 // For X87 we'd like to use the smallest FP type for this constant, but
13009 // for DAG type consistency we have to match the FP operand type.
13011 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
13012 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
13013 bool LosesInfo = false;
13014 if (TheVT == MVT::f64)
13015 // The rounding mode is irrelevant as the conversion should be exact.
13016 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
13018 else if (TheVT == MVT::f80)
13019 Status = Thresh.convert(APFloat::x87DoubleExtended,
13020 APFloat::rmNearestTiesToEven, &LosesInfo);
13022 assert(Status == APFloat::opOK && !LosesInfo &&
13023 "FP conversion should have been exact");
13025 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
13027 SDValue Cmp = DAG.getSetCC(DL,
13028 getSetCCResultType(DAG.getDataLayout(),
13029 *DAG.getContext(), TheVT),
13030 Value, ThreshVal, ISD::SETLT);
13031 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
13032 DAG.getConstant(0, DL, MVT::i32),
13033 DAG.getConstant(0x80000000, DL, MVT::i32));
13034 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
13035 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
13036 *DAG.getContext(), TheVT),
13037 Value, ThreshVal, ISD::SETLT);
13038 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
13041 // FIXME This causes a redundant load/store if the SSE-class value is already
13042 // in memory, such as if it is on the callstack.
13043 if (isScalarFPTypeInSSEReg(TheVT)) {
13044 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13045 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13046 MachinePointerInfo::getFixedStack(MF, SSFI), false,
13048 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13050 Chain, StackSlot, DAG.getValueType(TheVT)
13053 MachineMemOperand *MMO =
13054 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13055 MachineMemOperand::MOLoad, MemSize, MemSize);
13056 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13057 Chain = Value.getValue(1);
13058 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13059 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13062 MachineMemOperand *MMO =
13063 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13064 MachineMemOperand::MOStore, MemSize, MemSize);
13066 if (UnsignedFixup) {
13068 // Insert the FIST, load its result as two i32's,
13069 // and XOR the high i32 with Adjust.
13071 SDValue FistOps[] = { Chain, Value, StackSlot };
13072 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13073 FistOps, DstTy, MMO);
13075 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
13076 MachinePointerInfo(),
13077 false, false, false, 0);
13078 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
13079 DAG.getConstant(4, DL, PtrVT));
13081 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
13082 MachinePointerInfo(),
13083 false, false, false, 0);
13084 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13086 if (Subtarget->is64Bit()) {
13087 // Join High32 and Low32 into a 64-bit result.
13088 // (High32 << 32) | Low32
13089 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13090 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13091 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13092 DAG.getConstant(32, DL, MVT::i8));
13093 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13094 return std::make_pair(Result, SDValue());
13097 SDValue ResultOps[] = { Low32, High32 };
13099 SDValue pair = IsReplace
13100 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13101 : DAG.getMergeValues(ResultOps, DL);
13102 return std::make_pair(pair, SDValue());
13104 // Build the FP_TO_INT*_IN_MEM
13105 SDValue Ops[] = { Chain, Value, StackSlot };
13106 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13108 return std::make_pair(FIST, StackSlot);
13112 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13113 const X86Subtarget *Subtarget) {
13114 MVT VT = Op->getSimpleValueType(0);
13115 SDValue In = Op->getOperand(0);
13116 MVT InVT = In.getSimpleValueType();
13119 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13120 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13122 // Optimize vectors in AVX mode:
13125 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13126 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13127 // Concat upper and lower parts.
13130 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13131 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13132 // Concat upper and lower parts.
13135 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13136 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13137 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13140 if (Subtarget->hasInt256())
13141 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13143 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13144 SDValue Undef = DAG.getUNDEF(InVT);
13145 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13146 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13147 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13149 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13150 VT.getVectorNumElements()/2);
13152 OpLo = DAG.getBitcast(HVT, OpLo);
13153 OpHi = DAG.getBitcast(HVT, OpHi);
13155 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13158 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13159 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13160 MVT VT = Op->getSimpleValueType(0);
13161 SDValue In = Op->getOperand(0);
13162 MVT InVT = In.getSimpleValueType();
13164 unsigned int NumElts = VT.getVectorNumElements();
13165 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13168 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13169 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13171 assert(InVT.getVectorElementType() == MVT::i1);
13172 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13174 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13176 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13178 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13179 if (VT.is512BitVector())
13181 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13184 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13185 SelectionDAG &DAG) {
13186 if (Subtarget->hasFp256())
13187 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13193 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13194 SelectionDAG &DAG) {
13196 MVT VT = Op.getSimpleValueType();
13197 SDValue In = Op.getOperand(0);
13198 MVT SVT = In.getSimpleValueType();
13200 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13201 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13203 if (Subtarget->hasFp256())
13204 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13207 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13208 VT.getVectorNumElements() != SVT.getVectorNumElements());
13212 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13214 MVT VT = Op.getSimpleValueType();
13215 SDValue In = Op.getOperand(0);
13216 MVT InVT = In.getSimpleValueType();
13218 if (VT == MVT::i1) {
13219 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13220 "Invalid scalar TRUNCATE operation");
13221 if (InVT.getSizeInBits() >= 32)
13223 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13224 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13226 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13227 "Invalid TRUNCATE operation");
13229 // move vector to mask - truncate solution for SKX
13230 if (VT.getVectorElementType() == MVT::i1) {
13231 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13232 Subtarget->hasBWI())
13233 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13234 if ((InVT.is256BitVector() || InVT.is128BitVector())
13235 && InVT.getScalarSizeInBits() <= 16 &&
13236 Subtarget->hasBWI() && Subtarget->hasVLX())
13237 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13238 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13239 Subtarget->hasDQI())
13240 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13241 if ((InVT.is256BitVector() || InVT.is128BitVector())
13242 && InVT.getScalarSizeInBits() >= 32 &&
13243 Subtarget->hasDQI() && Subtarget->hasVLX())
13244 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13247 if (VT.getVectorElementType() == MVT::i1) {
13248 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13249 unsigned NumElts = InVT.getVectorNumElements();
13250 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13251 if (InVT.getSizeInBits() < 512) {
13252 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13253 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13258 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13259 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13260 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13263 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13264 if (Subtarget->hasAVX512()) {
13265 // word to byte only under BWI
13266 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13267 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13268 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13269 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13271 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13272 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13273 if (Subtarget->hasInt256()) {
13274 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13275 In = DAG.getBitcast(MVT::v8i32, In);
13276 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13278 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13279 DAG.getIntPtrConstant(0, DL));
13282 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13283 DAG.getIntPtrConstant(0, DL));
13284 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13285 DAG.getIntPtrConstant(2, DL));
13286 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13287 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13288 static const int ShufMask[] = {0, 2, 4, 6};
13289 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13292 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13293 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13294 if (Subtarget->hasInt256()) {
13295 In = DAG.getBitcast(MVT::v32i8, In);
13297 SmallVector<SDValue,32> pshufbMask;
13298 for (unsigned i = 0; i < 2; ++i) {
13299 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13300 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13301 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13302 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13303 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13304 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13305 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13306 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13307 for (unsigned j = 0; j < 8; ++j)
13308 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13310 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13311 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13312 In = DAG.getBitcast(MVT::v4i64, In);
13314 static const int ShufMask[] = {0, 2, -1, -1};
13315 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13317 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13318 DAG.getIntPtrConstant(0, DL));
13319 return DAG.getBitcast(VT, In);
13322 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13323 DAG.getIntPtrConstant(0, DL));
13325 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13326 DAG.getIntPtrConstant(4, DL));
13328 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13329 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13331 // The PSHUFB mask:
13332 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13333 -1, -1, -1, -1, -1, -1, -1, -1};
13335 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13336 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13337 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13339 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13340 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13342 // The MOVLHPS Mask:
13343 static const int ShufMask2[] = {0, 1, 4, 5};
13344 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13345 return DAG.getBitcast(MVT::v8i16, res);
13348 // Handle truncation of V256 to V128 using shuffles.
13349 if (!VT.is128BitVector() || !InVT.is256BitVector())
13352 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13354 unsigned NumElems = VT.getVectorNumElements();
13355 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13357 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13358 // Prepare truncation shuffle mask
13359 for (unsigned i = 0; i != NumElems; ++i)
13360 MaskVec[i] = i * 2;
13361 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13362 DAG.getUNDEF(NVT), &MaskVec[0]);
13363 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13364 DAG.getIntPtrConstant(0, DL));
13367 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13368 SelectionDAG &DAG) const {
13369 assert(!Op.getSimpleValueType().isVector());
13371 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13372 /*IsSigned=*/ true, /*IsReplace=*/ false);
13373 SDValue FIST = Vals.first, StackSlot = Vals.second;
13374 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13375 if (!FIST.getNode())
13378 if (StackSlot.getNode())
13379 // Load the result.
13380 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13381 FIST, StackSlot, MachinePointerInfo(),
13382 false, false, false, 0);
13384 // The node is the result.
13388 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13389 SelectionDAG &DAG) const {
13390 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13391 /*IsSigned=*/ false, /*IsReplace=*/ false);
13392 SDValue FIST = Vals.first, StackSlot = Vals.second;
13393 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13394 if (!FIST.getNode())
13397 if (StackSlot.getNode())
13398 // Load the result.
13399 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13400 FIST, StackSlot, MachinePointerInfo(),
13401 false, false, false, 0);
13403 // The node is the result.
13407 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13409 MVT VT = Op.getSimpleValueType();
13410 SDValue In = Op.getOperand(0);
13411 MVT SVT = In.getSimpleValueType();
13413 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13415 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13416 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13417 In, DAG.getUNDEF(SVT)));
13420 /// The only differences between FABS and FNEG are the mask and the logic op.
13421 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13422 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13423 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13424 "Wrong opcode for lowering FABS or FNEG.");
13426 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13428 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13429 // into an FNABS. We'll lower the FABS after that if it is still in use.
13431 for (SDNode *User : Op->uses())
13432 if (User->getOpcode() == ISD::FNEG)
13436 MVT VT = Op.getSimpleValueType();
13438 bool IsF128 = (VT == MVT::f128);
13440 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13441 // decide if we should generate a 16-byte constant mask when we only need 4 or
13442 // 8 bytes for the scalar case.
13448 if (VT.isVector()) {
13450 EltVT = VT.getVectorElementType();
13451 NumElts = VT.getVectorNumElements();
13452 } else if (IsF128) {
13453 // SSE instructions are used for optimized f128 logical operations.
13454 LogicVT = MVT::f128;
13458 // There are no scalar bitwise logical SSE/AVX instructions, so we
13459 // generate a 16-byte vector constant and logic op even for the scalar case.
13460 // Using a 16-byte mask allows folding the load of the mask with
13461 // the logic op, so it can save (~4 bytes) on code size.
13462 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13464 NumElts = (VT == MVT::f64) ? 2 : 4;
13467 unsigned EltBits = EltVT.getSizeInBits();
13468 LLVMContext *Context = DAG.getContext();
13469 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13471 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13472 Constant *C = ConstantInt::get(*Context, MaskElt);
13473 C = ConstantVector::getSplat(NumElts, C);
13474 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13475 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13476 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13478 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13479 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13480 false, false, false, Alignment);
13482 SDValue Op0 = Op.getOperand(0);
13483 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13485 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13486 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13488 if (VT.isVector() || IsF128)
13489 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13491 // For the scalar case extend to a 128-bit vector, perform the logic op,
13492 // and extract the scalar result back out.
13493 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13494 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13495 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13496 DAG.getIntPtrConstant(0, dl));
13499 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13500 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13501 LLVMContext *Context = DAG.getContext();
13502 SDValue Op0 = Op.getOperand(0);
13503 SDValue Op1 = Op.getOperand(1);
13505 MVT VT = Op.getSimpleValueType();
13506 MVT SrcVT = Op1.getSimpleValueType();
13507 bool IsF128 = (VT == MVT::f128);
13509 // If second operand is smaller, extend it first.
13510 if (SrcVT.bitsLT(VT)) {
13511 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13514 // And if it is bigger, shrink it first.
13515 if (SrcVT.bitsGT(VT)) {
13516 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13520 // At this point the operands and the result should have the same
13521 // type, and that won't be f80 since that is not custom lowered.
13522 assert((VT == MVT::f64 || VT == MVT::f32 || IsF128) &&
13523 "Unexpected type in LowerFCOPYSIGN");
13525 const fltSemantics &Sem =
13526 VT == MVT::f64 ? APFloat::IEEEdouble :
13527 (IsF128 ? APFloat::IEEEquad : APFloat::IEEEsingle);
13528 const unsigned SizeInBits = VT.getSizeInBits();
13530 SmallVector<Constant *, 4> CV(
13531 VT == MVT::f64 ? 2 : (IsF128 ? 1 : 4),
13532 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13534 // First, clear all bits but the sign bit from the second operand (sign).
13535 CV[0] = ConstantFP::get(*Context,
13536 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13537 Constant *C = ConstantVector::get(CV);
13538 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13539 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13541 // Perform all logic operations as 16-byte vectors because there are no
13542 // scalar FP logic instructions in SSE. This allows load folding of the
13543 // constants into the logic instructions.
13544 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : (IsF128 ? MVT::f128 : MVT::v4f32);
13546 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13547 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13548 false, false, false, 16);
13550 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13551 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13553 // Next, clear the sign bit from the first operand (magnitude).
13554 // If it's a constant, we can clear it here.
13555 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13556 APFloat APF = Op0CN->getValueAPF();
13557 // If the magnitude is a positive zero, the sign bit alone is enough.
13558 if (APF.isPosZero())
13559 return IsF128 ? SignBit :
13560 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13561 DAG.getIntPtrConstant(0, dl));
13563 CV[0] = ConstantFP::get(*Context, APF);
13565 CV[0] = ConstantFP::get(
13567 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13569 C = ConstantVector::get(CV);
13570 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13572 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13573 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13574 false, false, false, 16);
13575 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13576 if (!isa<ConstantFPSDNode>(Op0)) {
13578 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13579 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13581 // OR the magnitude value with the sign bit.
13582 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13583 return IsF128 ? Val :
13584 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13585 DAG.getIntPtrConstant(0, dl));
13588 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13589 SDValue N0 = Op.getOperand(0);
13591 MVT VT = Op.getSimpleValueType();
13593 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13594 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13595 DAG.getConstant(1, dl, VT));
13596 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13599 // Check whether an OR'd tree is PTEST-able.
13600 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13601 SelectionDAG &DAG) {
13602 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13604 if (!Subtarget->hasSSE41())
13607 if (!Op->hasOneUse())
13610 SDNode *N = Op.getNode();
13613 SmallVector<SDValue, 8> Opnds;
13614 DenseMap<SDValue, unsigned> VecInMap;
13615 SmallVector<SDValue, 8> VecIns;
13616 EVT VT = MVT::Other;
13618 // Recognize a special case where a vector is casted into wide integer to
13620 Opnds.push_back(N->getOperand(0));
13621 Opnds.push_back(N->getOperand(1));
13623 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13624 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13625 // BFS traverse all OR'd operands.
13626 if (I->getOpcode() == ISD::OR) {
13627 Opnds.push_back(I->getOperand(0));
13628 Opnds.push_back(I->getOperand(1));
13629 // Re-evaluate the number of nodes to be traversed.
13630 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13634 // Quit if a non-EXTRACT_VECTOR_ELT
13635 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13638 // Quit if without a constant index.
13639 SDValue Idx = I->getOperand(1);
13640 if (!isa<ConstantSDNode>(Idx))
13643 SDValue ExtractedFromVec = I->getOperand(0);
13644 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13645 if (M == VecInMap.end()) {
13646 VT = ExtractedFromVec.getValueType();
13647 // Quit if not 128/256-bit vector.
13648 if (!VT.is128BitVector() && !VT.is256BitVector())
13650 // Quit if not the same type.
13651 if (VecInMap.begin() != VecInMap.end() &&
13652 VT != VecInMap.begin()->first.getValueType())
13654 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13655 VecIns.push_back(ExtractedFromVec);
13657 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13660 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13661 "Not extracted from 128-/256-bit vector.");
13663 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13665 for (DenseMap<SDValue, unsigned>::const_iterator
13666 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13667 // Quit if not all elements are used.
13668 if (I->second != FullMask)
13672 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13674 // Cast all vectors into TestVT for PTEST.
13675 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13676 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13678 // If more than one full vectors are evaluated, OR them first before PTEST.
13679 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13680 // Each iteration will OR 2 nodes and append the result until there is only
13681 // 1 node left, i.e. the final OR'd value of all vectors.
13682 SDValue LHS = VecIns[Slot];
13683 SDValue RHS = VecIns[Slot + 1];
13684 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13687 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13688 VecIns.back(), VecIns.back());
13691 /// \brief return true if \c Op has a use that doesn't just read flags.
13692 static bool hasNonFlagsUse(SDValue Op) {
13693 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13695 SDNode *User = *UI;
13696 unsigned UOpNo = UI.getOperandNo();
13697 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13698 // Look pass truncate.
13699 UOpNo = User->use_begin().getOperandNo();
13700 User = *User->use_begin();
13703 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13704 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13710 /// Emit nodes that will be selected as "test Op0,Op0", or something
13712 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13713 SelectionDAG &DAG) const {
13714 if (Op.getValueType() == MVT::i1) {
13715 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13716 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13717 DAG.getConstant(0, dl, MVT::i8));
13719 // CF and OF aren't always set the way we want. Determine which
13720 // of these we need.
13721 bool NeedCF = false;
13722 bool NeedOF = false;
13725 case X86::COND_A: case X86::COND_AE:
13726 case X86::COND_B: case X86::COND_BE:
13729 case X86::COND_G: case X86::COND_GE:
13730 case X86::COND_L: case X86::COND_LE:
13731 case X86::COND_O: case X86::COND_NO: {
13732 // Check if we really need to set the
13733 // Overflow flag. If NoSignedWrap is present
13734 // that is not actually needed.
13735 switch (Op->getOpcode()) {
13740 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13741 if (BinNode->Flags.hasNoSignedWrap())
13751 // See if we can use the EFLAGS value from the operand instead of
13752 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13753 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13754 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13755 // Emit a CMP with 0, which is the TEST pattern.
13756 //if (Op.getValueType() == MVT::i1)
13757 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13758 // DAG.getConstant(0, MVT::i1));
13759 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13760 DAG.getConstant(0, dl, Op.getValueType()));
13762 unsigned Opcode = 0;
13763 unsigned NumOperands = 0;
13765 // Truncate operations may prevent the merge of the SETCC instruction
13766 // and the arithmetic instruction before it. Attempt to truncate the operands
13767 // of the arithmetic instruction and use a reduced bit-width instruction.
13768 bool NeedTruncation = false;
13769 SDValue ArithOp = Op;
13770 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13771 SDValue Arith = Op->getOperand(0);
13772 // Both the trunc and the arithmetic op need to have one user each.
13773 if (Arith->hasOneUse())
13774 switch (Arith.getOpcode()) {
13781 NeedTruncation = true;
13787 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13788 // which may be the result of a CAST. We use the variable 'Op', which is the
13789 // non-casted variable when we check for possible users.
13790 switch (ArithOp.getOpcode()) {
13792 // Due to an isel shortcoming, be conservative if this add is likely to be
13793 // selected as part of a load-modify-store instruction. When the root node
13794 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13795 // uses of other nodes in the match, such as the ADD in this case. This
13796 // leads to the ADD being left around and reselected, with the result being
13797 // two adds in the output. Alas, even if none our users are stores, that
13798 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13799 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13800 // climbing the DAG back to the root, and it doesn't seem to be worth the
13802 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13803 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13804 if (UI->getOpcode() != ISD::CopyToReg &&
13805 UI->getOpcode() != ISD::SETCC &&
13806 UI->getOpcode() != ISD::STORE)
13809 if (ConstantSDNode *C =
13810 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13811 // An add of one will be selected as an INC.
13812 if (C->isOne() && !Subtarget->slowIncDec()) {
13813 Opcode = X86ISD::INC;
13818 // An add of negative one (subtract of one) will be selected as a DEC.
13819 if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
13820 Opcode = X86ISD::DEC;
13826 // Otherwise use a regular EFLAGS-setting add.
13827 Opcode = X86ISD::ADD;
13832 // If we have a constant logical shift that's only used in a comparison
13833 // against zero turn it into an equivalent AND. This allows turning it into
13834 // a TEST instruction later.
13835 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13836 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13837 EVT VT = Op.getValueType();
13838 unsigned BitWidth = VT.getSizeInBits();
13839 unsigned ShAmt = Op->getConstantOperandVal(1);
13840 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13842 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13843 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13844 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13845 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13847 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13848 DAG.getConstant(Mask, dl, VT));
13849 DAG.ReplaceAllUsesWith(Op, New);
13855 // If the primary and result isn't used, don't bother using X86ISD::AND,
13856 // because a TEST instruction will be better.
13857 if (!hasNonFlagsUse(Op))
13863 // Due to the ISEL shortcoming noted above, be conservative if this op is
13864 // likely to be selected as part of a load-modify-store instruction.
13865 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13866 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13867 if (UI->getOpcode() == ISD::STORE)
13870 // Otherwise use a regular EFLAGS-setting instruction.
13871 switch (ArithOp.getOpcode()) {
13872 default: llvm_unreachable("unexpected operator!");
13873 case ISD::SUB: Opcode = X86ISD::SUB; break;
13874 case ISD::XOR: Opcode = X86ISD::XOR; break;
13875 case ISD::AND: Opcode = X86ISD::AND; break;
13877 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13878 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13879 if (EFLAGS.getNode())
13882 Opcode = X86ISD::OR;
13896 return SDValue(Op.getNode(), 1);
13902 // If we found that truncation is beneficial, perform the truncation and
13904 if (NeedTruncation) {
13905 EVT VT = Op.getValueType();
13906 SDValue WideVal = Op->getOperand(0);
13907 EVT WideVT = WideVal.getValueType();
13908 unsigned ConvertedOp = 0;
13909 // Use a target machine opcode to prevent further DAGCombine
13910 // optimizations that may separate the arithmetic operations
13911 // from the setcc node.
13912 switch (WideVal.getOpcode()) {
13914 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13915 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13916 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13917 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13918 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13922 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13923 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13924 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13925 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13926 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13932 // Emit a CMP with 0, which is the TEST pattern.
13933 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13934 DAG.getConstant(0, dl, Op.getValueType()));
13936 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13937 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13939 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13940 DAG.ReplaceAllUsesWith(Op, New);
13941 return SDValue(New.getNode(), 1);
13944 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13946 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13947 SDLoc dl, SelectionDAG &DAG) const {
13948 if (isNullConstant(Op1))
13949 return EmitTest(Op0, X86CC, dl, DAG);
13951 assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
13952 "Unexpected comparison operation for MVT::i1 operands");
13954 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13955 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13956 // Do the comparison at i32 if it's smaller, besides the Atom case.
13957 // This avoids subregister aliasing issues. Keep the smaller reference
13958 // if we're optimizing for size, however, as that'll allow better folding
13959 // of memory operations.
13960 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13961 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13962 !Subtarget->isAtom()) {
13963 unsigned ExtendOp =
13964 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13965 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13966 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13968 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13969 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13970 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13972 return SDValue(Sub.getNode(), 1);
13974 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13977 /// Convert a comparison if required by the subtarget.
13978 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13979 SelectionDAG &DAG) const {
13980 // If the subtarget does not support the FUCOMI instruction, floating-point
13981 // comparisons have to be converted.
13982 if (Subtarget->hasCMov() ||
13983 Cmp.getOpcode() != X86ISD::CMP ||
13984 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13985 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13988 // The instruction selector will select an FUCOM instruction instead of
13989 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13990 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13991 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13993 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13994 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13995 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13996 DAG.getConstant(8, dl, MVT::i8));
13997 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13999 // Some 64-bit targets lack SAHF support, but they do support FCOMI.
14000 assert(Subtarget->hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?");
14001 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
14004 /// The minimum architected relative accuracy is 2^-12. We need one
14005 /// Newton-Raphson step to have a good float result (24 bits of precision).
14006 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
14007 DAGCombinerInfo &DCI,
14008 unsigned &RefinementSteps,
14009 bool &UseOneConstNR) const {
14010 EVT VT = Op.getValueType();
14011 const char *RecipOp;
14013 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
14014 // TODO: Add support for AVX512 (v16f32).
14015 // It is likely not profitable to do this for f64 because a double-precision
14016 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
14017 // instructions: convert to single, rsqrtss, convert back to double, refine
14018 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
14019 // along with FMA, this could be a throughput win.
14020 if (VT == MVT::f32 && Subtarget->hasSSE1())
14022 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14023 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14024 RecipOp = "vec-sqrtf";
14028 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14029 if (!Recips.isEnabled(RecipOp))
14032 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14033 UseOneConstNR = false;
14034 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
14037 /// The minimum architected relative accuracy is 2^-12. We need one
14038 /// Newton-Raphson step to have a good float result (24 bits of precision).
14039 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
14040 DAGCombinerInfo &DCI,
14041 unsigned &RefinementSteps) const {
14042 EVT VT = Op.getValueType();
14043 const char *RecipOp;
14045 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
14046 // TODO: Add support for AVX512 (v16f32).
14047 // It is likely not profitable to do this for f64 because a double-precision
14048 // reciprocal estimate with refinement on x86 prior to FMA requires
14049 // 15 instructions: convert to single, rcpss, convert back to double, refine
14050 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
14051 // along with FMA, this could be a throughput win.
14052 if (VT == MVT::f32 && Subtarget->hasSSE1())
14054 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14055 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14056 RecipOp = "vec-divf";
14060 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14061 if (!Recips.isEnabled(RecipOp))
14064 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14065 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
14068 /// If we have at least two divisions that use the same divisor, convert to
14069 /// multplication by a reciprocal. This may need to be adjusted for a given
14070 /// CPU if a division's cost is not at least twice the cost of a multiplication.
14071 /// This is because we still need one division to calculate the reciprocal and
14072 /// then we need two multiplies by that reciprocal as replacements for the
14073 /// original divisions.
14074 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
14078 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
14079 /// if it's possible.
14080 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
14081 SDLoc dl, SelectionDAG &DAG) const {
14082 SDValue Op0 = And.getOperand(0);
14083 SDValue Op1 = And.getOperand(1);
14084 if (Op0.getOpcode() == ISD::TRUNCATE)
14085 Op0 = Op0.getOperand(0);
14086 if (Op1.getOpcode() == ISD::TRUNCATE)
14087 Op1 = Op1.getOperand(0);
14090 if (Op1.getOpcode() == ISD::SHL)
14091 std::swap(Op0, Op1);
14092 if (Op0.getOpcode() == ISD::SHL) {
14093 if (isOneConstant(Op0.getOperand(0))) {
14094 // If we looked past a truncate, check that it's only truncating away
14096 unsigned BitWidth = Op0.getValueSizeInBits();
14097 unsigned AndBitWidth = And.getValueSizeInBits();
14098 if (BitWidth > AndBitWidth) {
14100 DAG.computeKnownBits(Op0, Zeros, Ones);
14101 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14105 RHS = Op0.getOperand(1);
14107 } else if (Op1.getOpcode() == ISD::Constant) {
14108 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14109 uint64_t AndRHSVal = AndRHS->getZExtValue();
14110 SDValue AndLHS = Op0;
14112 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14113 LHS = AndLHS.getOperand(0);
14114 RHS = AndLHS.getOperand(1);
14117 // Use BT if the immediate can't be encoded in a TEST instruction.
14118 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14120 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14124 if (LHS.getNode()) {
14125 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14126 // instruction. Since the shift amount is in-range-or-undefined, we know
14127 // that doing a bittest on the i32 value is ok. We extend to i32 because
14128 // the encoding for the i16 version is larger than the i32 version.
14129 // Also promote i16 to i32 for performance / code size reason.
14130 if (LHS.getValueType() == MVT::i8 ||
14131 LHS.getValueType() == MVT::i16)
14132 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14134 // If the operand types disagree, extend the shift amount to match. Since
14135 // BT ignores high bits (like shifts) we can use anyextend.
14136 if (LHS.getValueType() != RHS.getValueType())
14137 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14139 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14140 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14141 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14142 DAG.getConstant(Cond, dl, MVT::i8), BT);
14148 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14150 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14155 // SSE Condition code mapping:
14164 switch (SetCCOpcode) {
14165 default: llvm_unreachable("Unexpected SETCC condition");
14167 case ISD::SETEQ: SSECC = 0; break;
14169 case ISD::SETGT: Swap = true; // Fallthrough
14171 case ISD::SETOLT: SSECC = 1; break;
14173 case ISD::SETGE: Swap = true; // Fallthrough
14175 case ISD::SETOLE: SSECC = 2; break;
14176 case ISD::SETUO: SSECC = 3; break;
14178 case ISD::SETNE: SSECC = 4; break;
14179 case ISD::SETULE: Swap = true; // Fallthrough
14180 case ISD::SETUGE: SSECC = 5; break;
14181 case ISD::SETULT: Swap = true; // Fallthrough
14182 case ISD::SETUGT: SSECC = 6; break;
14183 case ISD::SETO: SSECC = 7; break;
14185 case ISD::SETONE: SSECC = 8; break;
14188 std::swap(Op0, Op1);
14193 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14194 // ones, and then concatenate the result back.
14195 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14196 MVT VT = Op.getSimpleValueType();
14198 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14199 "Unsupported value type for operation");
14201 unsigned NumElems = VT.getVectorNumElements();
14203 SDValue CC = Op.getOperand(2);
14205 // Extract the LHS vectors
14206 SDValue LHS = Op.getOperand(0);
14207 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14208 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14210 // Extract the RHS vectors
14211 SDValue RHS = Op.getOperand(1);
14212 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14213 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14215 // Issue the operation on the smaller types and concatenate the result back
14216 MVT EltVT = VT.getVectorElementType();
14217 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14218 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14219 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14220 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14223 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14224 SDValue Op0 = Op.getOperand(0);
14225 SDValue Op1 = Op.getOperand(1);
14226 SDValue CC = Op.getOperand(2);
14227 MVT VT = Op.getSimpleValueType();
14230 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14231 "Unexpected type for boolean compare operation");
14232 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14233 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14234 DAG.getConstant(-1, dl, VT));
14235 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14236 DAG.getConstant(-1, dl, VT));
14237 switch (SetCCOpcode) {
14238 default: llvm_unreachable("Unexpected SETCC condition");
14240 // (x == y) -> ~(x ^ y)
14241 return DAG.getNode(ISD::XOR, dl, VT,
14242 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14243 DAG.getConstant(-1, dl, VT));
14245 // (x != y) -> (x ^ y)
14246 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14249 // (x > y) -> (x & ~y)
14250 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14253 // (x < y) -> (~x & y)
14254 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14257 // (x <= y) -> (~x | y)
14258 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14261 // (x >=y) -> (x | ~y)
14262 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14266 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14267 const X86Subtarget *Subtarget) {
14268 SDValue Op0 = Op.getOperand(0);
14269 SDValue Op1 = Op.getOperand(1);
14270 SDValue CC = Op.getOperand(2);
14271 MVT VT = Op.getSimpleValueType();
14274 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14275 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14276 "Cannot set masked compare for this operation");
14278 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14280 bool Unsigned = false;
14283 switch (SetCCOpcode) {
14284 default: llvm_unreachable("Unexpected SETCC condition");
14285 case ISD::SETNE: SSECC = 4; break;
14286 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14287 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14288 case ISD::SETLT: Swap = true; //fall-through
14289 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14290 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14291 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14292 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14293 case ISD::SETULE: Unsigned = true; //fall-through
14294 case ISD::SETLE: SSECC = 2; break;
14298 std::swap(Op0, Op1);
14300 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14301 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14302 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14303 DAG.getConstant(SSECC, dl, MVT::i8));
14306 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14307 /// operand \p Op1. If non-trivial (for example because it's not constant)
14308 /// return an empty value.
14309 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14311 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14315 MVT VT = Op1.getSimpleValueType();
14316 MVT EVT = VT.getVectorElementType();
14317 unsigned n = VT.getVectorNumElements();
14318 SmallVector<SDValue, 8> ULTOp1;
14320 for (unsigned i = 0; i < n; ++i) {
14321 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14322 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14325 // Avoid underflow.
14326 APInt Val = Elt->getAPIntValue();
14330 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14333 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14336 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14337 SelectionDAG &DAG) {
14338 SDValue Op0 = Op.getOperand(0);
14339 SDValue Op1 = Op.getOperand(1);
14340 SDValue CC = Op.getOperand(2);
14341 MVT VT = Op.getSimpleValueType();
14342 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14343 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14348 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14349 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14352 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14353 unsigned Opc = X86ISD::CMPP;
14354 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14355 assert(VT.getVectorNumElements() <= 16);
14356 Opc = X86ISD::CMPM;
14358 // In the two special cases we can't handle, emit two comparisons.
14361 unsigned CombineOpc;
14362 if (SetCCOpcode == ISD::SETUEQ) {
14363 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14365 assert(SetCCOpcode == ISD::SETONE);
14366 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14369 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14370 DAG.getConstant(CC0, dl, MVT::i8));
14371 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14372 DAG.getConstant(CC1, dl, MVT::i8));
14373 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14375 // Handle all other FP comparisons here.
14376 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14377 DAG.getConstant(SSECC, dl, MVT::i8));
14380 MVT VTOp0 = Op0.getSimpleValueType();
14381 assert(VTOp0 == Op1.getSimpleValueType() &&
14382 "Expected operands with same type!");
14383 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14384 "Invalid number of packed elements for source and destination!");
14386 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14387 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14388 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14389 // legalizer firstly checks if the first operand in input to the setcc has
14390 // a legal type. If so, then it promotes the return type to that same type.
14391 // Otherwise, the return type is promoted to the 'next legal type' which,
14392 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14394 // We reach this code only if the following two conditions are met:
14395 // 1. Both return type and operand type have been promoted to wider types
14396 // by the type legalizer.
14397 // 2. The original operand type has been promoted to a 256-bit vector.
14399 // Note that condition 2. only applies for AVX targets.
14400 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14401 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14404 // The non-AVX512 code below works under the assumption that source and
14405 // destination types are the same.
14406 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14407 "Value types for source and destination must be the same!");
14409 // Break 256-bit integer vector compare into smaller ones.
14410 if (VT.is256BitVector() && !Subtarget->hasInt256())
14411 return Lower256IntVSETCC(Op, DAG);
14413 MVT OpVT = Op1.getSimpleValueType();
14414 if (OpVT.getVectorElementType() == MVT::i1)
14415 return LowerBoolVSETCC_AVX512(Op, DAG);
14417 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14418 if (Subtarget->hasAVX512()) {
14419 if (Op1.getSimpleValueType().is512BitVector() ||
14420 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14421 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14422 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14424 // In AVX-512 architecture setcc returns mask with i1 elements,
14425 // But there is no compare instruction for i8 and i16 elements in KNL.
14426 // We are not talking about 512-bit operands in this case, these
14427 // types are illegal.
14429 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14430 OpVT.getVectorElementType().getSizeInBits() >= 8))
14431 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14432 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14435 // Lower using XOP integer comparisons.
14436 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14437 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14438 // Translate compare code to XOP PCOM compare mode.
14439 unsigned CmpMode = 0;
14440 switch (SetCCOpcode) {
14441 default: llvm_unreachable("Unexpected SETCC condition");
14443 case ISD::SETLT: CmpMode = 0x00; break;
14445 case ISD::SETLE: CmpMode = 0x01; break;
14447 case ISD::SETGT: CmpMode = 0x02; break;
14449 case ISD::SETGE: CmpMode = 0x03; break;
14450 case ISD::SETEQ: CmpMode = 0x04; break;
14451 case ISD::SETNE: CmpMode = 0x05; break;
14454 // Are we comparing unsigned or signed integers?
14455 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14456 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14458 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14459 DAG.getConstant(CmpMode, dl, MVT::i8));
14462 // We are handling one of the integer comparisons here. Since SSE only has
14463 // GT and EQ comparisons for integer, swapping operands and multiple
14464 // operations may be required for some comparisons.
14466 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14467 bool Subus = false;
14469 switch (SetCCOpcode) {
14470 default: llvm_unreachable("Unexpected SETCC condition");
14471 case ISD::SETNE: Invert = true;
14472 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14473 case ISD::SETLT: Swap = true;
14474 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14475 case ISD::SETGE: Swap = true;
14476 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14477 Invert = true; break;
14478 case ISD::SETULT: Swap = true;
14479 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14480 FlipSigns = true; break;
14481 case ISD::SETUGE: Swap = true;
14482 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14483 FlipSigns = true; Invert = true; break;
14486 // Special case: Use min/max operations for SETULE/SETUGE
14487 MVT VET = VT.getVectorElementType();
14489 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14490 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14493 switch (SetCCOpcode) {
14495 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14496 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14499 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14502 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14503 if (!MinMax && hasSubus) {
14504 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14506 // t = psubus Op0, Op1
14507 // pcmpeq t, <0..0>
14508 switch (SetCCOpcode) {
14510 case ISD::SETULT: {
14511 // If the comparison is against a constant we can turn this into a
14512 // setule. With psubus, setule does not require a swap. This is
14513 // beneficial because the constant in the register is no longer
14514 // destructed as the destination so it can be hoisted out of a loop.
14515 // Only do this pre-AVX since vpcmp* is no longer destructive.
14516 if (Subtarget->hasAVX())
14518 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14519 if (ULEOp1.getNode()) {
14521 Subus = true; Invert = false; Swap = false;
14525 // Psubus is better than flip-sign because it requires no inversion.
14526 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14527 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14531 Opc = X86ISD::SUBUS;
14537 std::swap(Op0, Op1);
14539 // Check that the operation in question is available (most are plain SSE2,
14540 // but PCMPGTQ and PCMPEQQ have different requirements).
14541 if (VT == MVT::v2i64) {
14542 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14543 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14545 // First cast everything to the right type.
14546 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14547 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14549 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14550 // bits of the inputs before performing those operations. The lower
14551 // compare is always unsigned.
14554 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14556 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14557 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14558 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14559 Sign, Zero, Sign, Zero);
14561 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14562 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14564 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14565 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14566 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14568 // Create masks for only the low parts/high parts of the 64 bit integers.
14569 static const int MaskHi[] = { 1, 1, 3, 3 };
14570 static const int MaskLo[] = { 0, 0, 2, 2 };
14571 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14572 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14573 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14575 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14576 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14579 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14581 return DAG.getBitcast(VT, Result);
14584 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14585 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14586 // pcmpeqd + pshufd + pand.
14587 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14589 // First cast everything to the right type.
14590 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14591 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14594 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14596 // Make sure the lower and upper halves are both all-ones.
14597 static const int Mask[] = { 1, 0, 3, 2 };
14598 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14599 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14602 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14604 return DAG.getBitcast(VT, Result);
14608 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14609 // bits of the inputs before performing those operations.
14611 MVT EltVT = VT.getVectorElementType();
14612 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14614 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14615 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14618 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14620 // If the logical-not of the result is required, perform that now.
14622 Result = DAG.getNOT(dl, Result, VT);
14625 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14628 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14629 getZeroVector(VT, Subtarget, DAG, dl));
14634 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14636 MVT VT = Op.getSimpleValueType();
14638 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14640 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14641 && "SetCC type must be 8-bit or 1-bit integer");
14642 SDValue Op0 = Op.getOperand(0);
14643 SDValue Op1 = Op.getOperand(1);
14645 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14647 // Optimize to BT if possible.
14648 // Lower (X & (1 << N)) == 0 to BT(X, N).
14649 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14650 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14651 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14652 isNullConstant(Op1) &&
14653 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14654 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14656 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14661 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14663 if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
14664 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14666 // If the input is a setcc, then reuse the input setcc or use a new one with
14667 // the inverted condition.
14668 if (Op0.getOpcode() == X86ISD::SETCC) {
14669 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14670 bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
14674 CCode = X86::GetOppositeBranchCondition(CCode);
14675 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14676 DAG.getConstant(CCode, dl, MVT::i8),
14677 Op0.getOperand(1));
14679 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14683 if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
14684 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14686 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14687 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14690 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14691 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14692 if (X86CC == X86::COND_INVALID)
14695 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14696 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14697 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14698 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14700 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14704 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14705 SDValue LHS = Op.getOperand(0);
14706 SDValue RHS = Op.getOperand(1);
14707 SDValue Carry = Op.getOperand(2);
14708 SDValue Cond = Op.getOperand(3);
14711 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14712 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14714 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14715 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14716 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14717 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14718 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14721 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14722 static bool isX86LogicalCmp(SDValue Op) {
14723 unsigned Opc = Op.getNode()->getOpcode();
14724 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14725 Opc == X86ISD::SAHF)
14727 if (Op.getResNo() == 1 &&
14728 (Opc == X86ISD::ADD ||
14729 Opc == X86ISD::SUB ||
14730 Opc == X86ISD::ADC ||
14731 Opc == X86ISD::SBB ||
14732 Opc == X86ISD::SMUL ||
14733 Opc == X86ISD::UMUL ||
14734 Opc == X86ISD::INC ||
14735 Opc == X86ISD::DEC ||
14736 Opc == X86ISD::OR ||
14737 Opc == X86ISD::XOR ||
14738 Opc == X86ISD::AND))
14741 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14747 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14748 if (V.getOpcode() != ISD::TRUNCATE)
14751 SDValue VOp0 = V.getOperand(0);
14752 unsigned InBits = VOp0.getValueSizeInBits();
14753 unsigned Bits = V.getValueSizeInBits();
14754 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14757 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14758 bool addTest = true;
14759 SDValue Cond = Op.getOperand(0);
14760 SDValue Op1 = Op.getOperand(1);
14761 SDValue Op2 = Op.getOperand(2);
14763 MVT VT = Op1.getSimpleValueType();
14766 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14767 // are available or VBLENDV if AVX is available.
14768 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14769 if (Cond.getOpcode() == ISD::SETCC &&
14770 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14771 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14772 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14773 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14774 int SSECC = translateX86FSETCC(
14775 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14778 if (Subtarget->hasAVX512()) {
14779 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14780 DAG.getConstant(SSECC, DL, MVT::i8));
14781 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14784 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14785 DAG.getConstant(SSECC, DL, MVT::i8));
14787 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14788 // of 3 logic instructions for size savings and potentially speed.
14789 // Unfortunately, there is no scalar form of VBLENDV.
14791 // If either operand is a constant, don't try this. We can expect to
14792 // optimize away at least one of the logic instructions later in that
14793 // case, so that sequence would be faster than a variable blend.
14795 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14796 // uses XMM0 as the selection register. That may need just as many
14797 // instructions as the AND/ANDN/OR sequence due to register moves, so
14800 if (Subtarget->hasAVX() &&
14801 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14803 // Convert to vectors, do a VSELECT, and convert back to scalar.
14804 // All of the conversions should be optimized away.
14806 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14807 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14808 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14809 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14811 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14812 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14814 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14816 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14817 VSel, DAG.getIntPtrConstant(0, DL));
14819 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14820 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14821 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14825 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
14827 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14828 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14829 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14830 Op1Scalar = Op1.getOperand(0);
14832 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14833 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14834 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14835 Op2Scalar = Op2.getOperand(0);
14836 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14837 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14838 Op1Scalar.getValueType(),
14839 Cond, Op1Scalar, Op2Scalar);
14840 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14841 return DAG.getBitcast(VT, newSelect);
14842 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14843 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14844 DAG.getIntPtrConstant(0, DL));
14848 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14849 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14850 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14851 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14852 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14853 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14854 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14856 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14859 if (Cond.getOpcode() == ISD::SETCC) {
14860 SDValue NewCond = LowerSETCC(Cond, DAG);
14861 if (NewCond.getNode())
14865 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14866 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14867 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14868 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14869 if (Cond.getOpcode() == X86ISD::SETCC &&
14870 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14871 isNullConstant(Cond.getOperand(1).getOperand(1))) {
14872 SDValue Cmp = Cond.getOperand(1);
14874 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14876 if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
14877 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14878 SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
14880 SDValue CmpOp0 = Cmp.getOperand(0);
14881 // Apply further optimizations for special cases
14882 // (select (x != 0), -1, 0) -> neg & sbb
14883 // (select (x == 0), 0, -1) -> neg & sbb
14884 if (isNullConstant(Y) &&
14885 (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
14886 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14887 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14888 DAG.getConstant(0, DL,
14889 CmpOp0.getValueType()),
14891 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14892 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14893 SDValue(Neg.getNode(), 1));
14897 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14898 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14899 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14901 SDValue Res = // Res = 0 or -1.
14902 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14903 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14905 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
14906 Res = DAG.getNOT(DL, Res, Res.getValueType());
14908 if (!isNullConstant(Op2))
14909 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14914 // Look past (and (setcc_carry (cmp ...)), 1).
14915 if (Cond.getOpcode() == ISD::AND &&
14916 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
14917 isOneConstant(Cond.getOperand(1)))
14918 Cond = Cond.getOperand(0);
14920 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14921 // setting operand in place of the X86ISD::SETCC.
14922 unsigned CondOpcode = Cond.getOpcode();
14923 if (CondOpcode == X86ISD::SETCC ||
14924 CondOpcode == X86ISD::SETCC_CARRY) {
14925 CC = Cond.getOperand(0);
14927 SDValue Cmp = Cond.getOperand(1);
14928 unsigned Opc = Cmp.getOpcode();
14929 MVT VT = Op.getSimpleValueType();
14931 bool IllegalFPCMov = false;
14932 if (VT.isFloatingPoint() && !VT.isVector() &&
14933 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14934 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14936 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14937 Opc == X86ISD::BT) { // FIXME
14941 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14942 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14943 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14944 Cond.getOperand(0).getValueType() != MVT::i8)) {
14945 SDValue LHS = Cond.getOperand(0);
14946 SDValue RHS = Cond.getOperand(1);
14947 unsigned X86Opcode;
14950 switch (CondOpcode) {
14951 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14952 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14953 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14954 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14955 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14956 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14957 default: llvm_unreachable("unexpected overflowing operator");
14959 if (CondOpcode == ISD::UMULO)
14960 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14963 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14965 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14967 if (CondOpcode == ISD::UMULO)
14968 Cond = X86Op.getValue(2);
14970 Cond = X86Op.getValue(1);
14972 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14977 // Look past the truncate if the high bits are known zero.
14978 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14979 Cond = Cond.getOperand(0);
14981 // We know the result of AND is compared against zero. Try to match
14983 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14984 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
14985 CC = NewSetCC.getOperand(0);
14986 Cond = NewSetCC.getOperand(1);
14993 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14994 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14997 // a < b ? -1 : 0 -> RES = ~setcc_carry
14998 // a < b ? 0 : -1 -> RES = setcc_carry
14999 // a >= b ? -1 : 0 -> RES = setcc_carry
15000 // a >= b ? 0 : -1 -> RES = ~setcc_carry
15001 if (Cond.getOpcode() == X86ISD::SUB) {
15002 Cond = ConvertCmpIfNecessary(Cond, DAG);
15003 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
15005 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
15006 (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15007 (isNullConstant(Op1) || isNullConstant(Op2))) {
15008 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15009 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15011 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
15012 return DAG.getNOT(DL, Res, Res.getValueType());
15017 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15018 // widen the cmov and push the truncate through. This avoids introducing a new
15019 // branch during isel and doesn't add any extensions.
15020 if (Op.getValueType() == MVT::i8 &&
15021 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15022 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15023 if (T1.getValueType() == T2.getValueType() &&
15024 // Blacklist CopyFromReg to avoid partial register stalls.
15025 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15026 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15027 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15028 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15032 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15033 // condition is true.
15034 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15035 SDValue Ops[] = { Op2, Op1, CC, Cond };
15036 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15039 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
15040 const X86Subtarget *Subtarget,
15041 SelectionDAG &DAG) {
15042 MVT VT = Op->getSimpleValueType(0);
15043 SDValue In = Op->getOperand(0);
15044 MVT InVT = In.getSimpleValueType();
15045 MVT VTElt = VT.getVectorElementType();
15046 MVT InVTElt = InVT.getVectorElementType();
15050 if ((InVTElt == MVT::i1) &&
15051 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15052 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15054 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15055 VTElt.getSizeInBits() <= 16)) ||
15057 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15058 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15060 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15061 VTElt.getSizeInBits() >= 32))))
15062 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15064 unsigned int NumElts = VT.getVectorNumElements();
15066 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
15069 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15070 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15071 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15072 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15075 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15076 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
15078 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
15081 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
15083 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
15084 if (VT.is512BitVector())
15086 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
15089 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
15090 const X86Subtarget *Subtarget,
15091 SelectionDAG &DAG) {
15092 SDValue In = Op->getOperand(0);
15093 MVT VT = Op->getSimpleValueType(0);
15094 MVT InVT = In.getSimpleValueType();
15095 assert(VT.getSizeInBits() == InVT.getSizeInBits());
15097 MVT InSVT = InVT.getVectorElementType();
15098 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
15100 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
15102 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15107 // SSE41 targets can use the pmovsx* instructions directly.
15108 if (Subtarget->hasSSE41())
15109 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15111 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15115 // As SRAI is only available on i16/i32 types, we expand only up to i32
15116 // and handle i64 separately.
15117 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15118 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15119 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15120 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15121 Curr = DAG.getBitcast(CurrVT, Curr);
15124 SDValue SignExt = Curr;
15125 if (CurrVT != InVT) {
15126 unsigned SignExtShift =
15127 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15128 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15129 DAG.getConstant(SignExtShift, dl, MVT::i8));
15135 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15136 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15137 DAG.getConstant(31, dl, MVT::i8));
15138 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15139 return DAG.getBitcast(VT, Ext);
15145 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15146 SelectionDAG &DAG) {
15147 MVT VT = Op->getSimpleValueType(0);
15148 SDValue In = Op->getOperand(0);
15149 MVT InVT = In.getSimpleValueType();
15152 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15153 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15155 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15156 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15157 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15160 if (Subtarget->hasInt256())
15161 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15163 // Optimize vectors in AVX mode
15164 // Sign extend v8i16 to v8i32 and
15167 // Divide input vector into two parts
15168 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15169 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15170 // concat the vectors to original VT
15172 unsigned NumElems = InVT.getVectorNumElements();
15173 SDValue Undef = DAG.getUNDEF(InVT);
15175 SmallVector<int,8> ShufMask1(NumElems, -1);
15176 for (unsigned i = 0; i != NumElems/2; ++i)
15179 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15181 SmallVector<int,8> ShufMask2(NumElems, -1);
15182 for (unsigned i = 0; i != NumElems/2; ++i)
15183 ShufMask2[i] = i + NumElems/2;
15185 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15187 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15188 VT.getVectorNumElements()/2);
15190 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15191 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15193 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15196 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15197 // may emit an illegal shuffle but the expansion is still better than scalar
15198 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15199 // we'll emit a shuffle and a arithmetic shift.
15200 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15201 // TODO: It is possible to support ZExt by zeroing the undef values during
15202 // the shuffle phase or after the shuffle.
15203 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15204 SelectionDAG &DAG) {
15205 MVT RegVT = Op.getSimpleValueType();
15206 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15207 assert(RegVT.isInteger() &&
15208 "We only custom lower integer vector sext loads.");
15210 // Nothing useful we can do without SSE2 shuffles.
15211 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15213 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15215 EVT MemVT = Ld->getMemoryVT();
15216 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15217 unsigned RegSz = RegVT.getSizeInBits();
15219 ISD::LoadExtType Ext = Ld->getExtensionType();
15221 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15222 && "Only anyext and sext are currently implemented.");
15223 assert(MemVT != RegVT && "Cannot extend to the same type");
15224 assert(MemVT.isVector() && "Must load a vector from memory");
15226 unsigned NumElems = RegVT.getVectorNumElements();
15227 unsigned MemSz = MemVT.getSizeInBits();
15228 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15230 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15231 // The only way in which we have a legal 256-bit vector result but not the
15232 // integer 256-bit operations needed to directly lower a sextload is if we
15233 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15234 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15235 // correctly legalized. We do this late to allow the canonical form of
15236 // sextload to persist throughout the rest of the DAG combiner -- it wants
15237 // to fold together any extensions it can, and so will fuse a sign_extend
15238 // of an sextload into a sextload targeting a wider value.
15240 if (MemSz == 128) {
15241 // Just switch this to a normal load.
15242 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15243 "it must be a legal 128-bit vector "
15245 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15246 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15247 Ld->isInvariant(), Ld->getAlignment());
15249 assert(MemSz < 128 &&
15250 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15251 // Do an sext load to a 128-bit vector type. We want to use the same
15252 // number of elements, but elements half as wide. This will end up being
15253 // recursively lowered by this routine, but will succeed as we definitely
15254 // have all the necessary features if we're using AVX1.
15256 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15257 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15259 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15260 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15261 Ld->isNonTemporal(), Ld->isInvariant(),
15262 Ld->getAlignment());
15265 // Replace chain users with the new chain.
15266 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15267 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15269 // Finally, do a normal sign-extend to the desired register.
15270 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15273 // All sizes must be a power of two.
15274 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15275 "Non-power-of-two elements are not custom lowered!");
15277 // Attempt to load the original value using scalar loads.
15278 // Find the largest scalar type that divides the total loaded size.
15279 MVT SclrLoadTy = MVT::i8;
15280 for (MVT Tp : MVT::integer_valuetypes()) {
15281 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15286 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15287 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15289 SclrLoadTy = MVT::f64;
15291 // Calculate the number of scalar loads that we need to perform
15292 // in order to load our vector from memory.
15293 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15295 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15296 "Can only lower sext loads with a single scalar load!");
15298 unsigned loadRegZize = RegSz;
15299 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15302 // Represent our vector as a sequence of elements which are the
15303 // largest scalar that we can load.
15304 EVT LoadUnitVecVT = EVT::getVectorVT(
15305 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15307 // Represent the data using the same element type that is stored in
15308 // memory. In practice, we ''widen'' MemVT.
15310 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15311 loadRegZize / MemVT.getScalarSizeInBits());
15313 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15314 "Invalid vector type");
15316 // We can't shuffle using an illegal type.
15317 assert(TLI.isTypeLegal(WideVecVT) &&
15318 "We only lower types that form legal widened vector types");
15320 SmallVector<SDValue, 8> Chains;
15321 SDValue Ptr = Ld->getBasePtr();
15322 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15323 TLI.getPointerTy(DAG.getDataLayout()));
15324 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15326 for (unsigned i = 0; i < NumLoads; ++i) {
15327 // Perform a single load.
15328 SDValue ScalarLoad =
15329 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15330 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15331 Ld->getAlignment());
15332 Chains.push_back(ScalarLoad.getValue(1));
15333 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15334 // another round of DAGCombining.
15336 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15338 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15339 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15341 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15344 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15346 // Bitcast the loaded value to a vector of the original element type, in
15347 // the size of the target vector type.
15348 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15349 unsigned SizeRatio = RegSz / MemSz;
15351 if (Ext == ISD::SEXTLOAD) {
15352 // If we have SSE4.1, we can directly emit a VSEXT node.
15353 if (Subtarget->hasSSE41()) {
15354 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15355 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15359 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15361 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15362 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15364 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15365 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15369 // Redistribute the loaded elements into the different locations.
15370 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15371 for (unsigned i = 0; i != NumElems; ++i)
15372 ShuffleVec[i * SizeRatio] = i;
15374 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15375 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15377 // Bitcast to the requested type.
15378 Shuff = DAG.getBitcast(RegVT, Shuff);
15379 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15383 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15384 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15385 // from the AND / OR.
15386 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15387 Opc = Op.getOpcode();
15388 if (Opc != ISD::OR && Opc != ISD::AND)
15390 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15391 Op.getOperand(0).hasOneUse() &&
15392 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15393 Op.getOperand(1).hasOneUse());
15396 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15397 // 1 and that the SETCC node has a single use.
15398 static bool isXor1OfSetCC(SDValue Op) {
15399 if (Op.getOpcode() != ISD::XOR)
15401 if (isOneConstant(Op.getOperand(1)))
15402 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15403 Op.getOperand(0).hasOneUse();
15407 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15408 bool addTest = true;
15409 SDValue Chain = Op.getOperand(0);
15410 SDValue Cond = Op.getOperand(1);
15411 SDValue Dest = Op.getOperand(2);
15414 bool Inverted = false;
15416 if (Cond.getOpcode() == ISD::SETCC) {
15417 // Check for setcc([su]{add,sub,mul}o == 0).
15418 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15419 isNullConstant(Cond.getOperand(1)) &&
15420 Cond.getOperand(0).getResNo() == 1 &&
15421 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15422 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15423 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15424 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15425 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15426 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15428 Cond = Cond.getOperand(0);
15430 SDValue NewCond = LowerSETCC(Cond, DAG);
15431 if (NewCond.getNode())
15436 // FIXME: LowerXALUO doesn't handle these!!
15437 else if (Cond.getOpcode() == X86ISD::ADD ||
15438 Cond.getOpcode() == X86ISD::SUB ||
15439 Cond.getOpcode() == X86ISD::SMUL ||
15440 Cond.getOpcode() == X86ISD::UMUL)
15441 Cond = LowerXALUO(Cond, DAG);
15444 // Look pass (and (setcc_carry (cmp ...)), 1).
15445 if (Cond.getOpcode() == ISD::AND &&
15446 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15447 isOneConstant(Cond.getOperand(1)))
15448 Cond = Cond.getOperand(0);
15450 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15451 // setting operand in place of the X86ISD::SETCC.
15452 unsigned CondOpcode = Cond.getOpcode();
15453 if (CondOpcode == X86ISD::SETCC ||
15454 CondOpcode == X86ISD::SETCC_CARRY) {
15455 CC = Cond.getOperand(0);
15457 SDValue Cmp = Cond.getOperand(1);
15458 unsigned Opc = Cmp.getOpcode();
15459 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15460 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15464 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15468 // These can only come from an arithmetic instruction with overflow,
15469 // e.g. SADDO, UADDO.
15470 Cond = Cond.getNode()->getOperand(1);
15476 CondOpcode = Cond.getOpcode();
15477 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15478 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15479 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15480 Cond.getOperand(0).getValueType() != MVT::i8)) {
15481 SDValue LHS = Cond.getOperand(0);
15482 SDValue RHS = Cond.getOperand(1);
15483 unsigned X86Opcode;
15486 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15487 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15489 switch (CondOpcode) {
15490 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15492 if (isOneConstant(RHS)) {
15493 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15496 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15497 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15499 if (isOneConstant(RHS)) {
15500 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15503 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15504 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15505 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15506 default: llvm_unreachable("unexpected overflowing operator");
15509 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15510 if (CondOpcode == ISD::UMULO)
15511 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15514 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15516 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15518 if (CondOpcode == ISD::UMULO)
15519 Cond = X86Op.getValue(2);
15521 Cond = X86Op.getValue(1);
15523 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15527 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15528 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15529 if (CondOpc == ISD::OR) {
15530 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15531 // two branches instead of an explicit OR instruction with a
15533 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15534 isX86LogicalCmp(Cmp)) {
15535 CC = Cond.getOperand(0).getOperand(0);
15536 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15537 Chain, Dest, CC, Cmp);
15538 CC = Cond.getOperand(1).getOperand(0);
15542 } else { // ISD::AND
15543 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15544 // two branches instead of an explicit AND instruction with a
15545 // separate test. However, we only do this if this block doesn't
15546 // have a fall-through edge, because this requires an explicit
15547 // jmp when the condition is false.
15548 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15549 isX86LogicalCmp(Cmp) &&
15550 Op.getNode()->hasOneUse()) {
15551 X86::CondCode CCode =
15552 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15553 CCode = X86::GetOppositeBranchCondition(CCode);
15554 CC = DAG.getConstant(CCode, dl, MVT::i8);
15555 SDNode *User = *Op.getNode()->use_begin();
15556 // Look for an unconditional branch following this conditional branch.
15557 // We need this because we need to reverse the successors in order
15558 // to implement FCMP_OEQ.
15559 if (User->getOpcode() == ISD::BR) {
15560 SDValue FalseBB = User->getOperand(1);
15562 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15563 assert(NewBR == User);
15567 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15568 Chain, Dest, CC, Cmp);
15569 X86::CondCode CCode =
15570 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15571 CCode = X86::GetOppositeBranchCondition(CCode);
15572 CC = DAG.getConstant(CCode, dl, MVT::i8);
15578 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15579 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15580 // It should be transformed during dag combiner except when the condition
15581 // is set by a arithmetics with overflow node.
15582 X86::CondCode CCode =
15583 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15584 CCode = X86::GetOppositeBranchCondition(CCode);
15585 CC = DAG.getConstant(CCode, dl, MVT::i8);
15586 Cond = Cond.getOperand(0).getOperand(1);
15588 } else if (Cond.getOpcode() == ISD::SETCC &&
15589 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15590 // For FCMP_OEQ, we can emit
15591 // two branches instead of an explicit AND instruction with a
15592 // separate test. However, we only do this if this block doesn't
15593 // have a fall-through edge, because this requires an explicit
15594 // jmp when the condition is false.
15595 if (Op.getNode()->hasOneUse()) {
15596 SDNode *User = *Op.getNode()->use_begin();
15597 // Look for an unconditional branch following this conditional branch.
15598 // We need this because we need to reverse the successors in order
15599 // to implement FCMP_OEQ.
15600 if (User->getOpcode() == ISD::BR) {
15601 SDValue FalseBB = User->getOperand(1);
15603 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15604 assert(NewBR == User);
15608 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15609 Cond.getOperand(0), Cond.getOperand(1));
15610 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15611 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15612 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15613 Chain, Dest, CC, Cmp);
15614 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15619 } else if (Cond.getOpcode() == ISD::SETCC &&
15620 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15621 // For FCMP_UNE, we can emit
15622 // two branches instead of an explicit AND instruction with a
15623 // separate test. However, we only do this if this block doesn't
15624 // have a fall-through edge, because this requires an explicit
15625 // jmp when the condition is false.
15626 if (Op.getNode()->hasOneUse()) {
15627 SDNode *User = *Op.getNode()->use_begin();
15628 // Look for an unconditional branch following this conditional branch.
15629 // We need this because we need to reverse the successors in order
15630 // to implement FCMP_UNE.
15631 if (User->getOpcode() == ISD::BR) {
15632 SDValue FalseBB = User->getOperand(1);
15634 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15635 assert(NewBR == User);
15638 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15639 Cond.getOperand(0), Cond.getOperand(1));
15640 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15641 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15642 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15643 Chain, Dest, CC, Cmp);
15644 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15654 // Look pass the truncate if the high bits are known zero.
15655 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15656 Cond = Cond.getOperand(0);
15658 // We know the result of AND is compared against zero. Try to match
15660 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15661 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15662 CC = NewSetCC.getOperand(0);
15663 Cond = NewSetCC.getOperand(1);
15670 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15671 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15672 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15674 Cond = ConvertCmpIfNecessary(Cond, DAG);
15675 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15676 Chain, Dest, CC, Cond);
15679 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15680 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15681 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15682 // that the guard pages used by the OS virtual memory manager are allocated in
15683 // correct sequence.
15685 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15686 SelectionDAG &DAG) const {
15687 MachineFunction &MF = DAG.getMachineFunction();
15688 bool SplitStack = MF.shouldSplitStack();
15689 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15694 SDNode *Node = Op.getNode();
15695 SDValue Chain = Op.getOperand(0);
15696 SDValue Size = Op.getOperand(1);
15697 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15698 EVT VT = Node->getValueType(0);
15700 // Chain the dynamic stack allocation so that it doesn't modify the stack
15701 // pointer when other instructions are using the stack.
15702 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
15704 bool Is64Bit = Subtarget->is64Bit();
15705 MVT SPTy = getPointerTy(DAG.getDataLayout());
15709 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15710 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15711 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15712 " not tell us which reg is the stack pointer!");
15713 EVT VT = Node->getValueType(0);
15714 SDValue Tmp3 = Node->getOperand(2);
15716 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15717 Chain = SP.getValue(1);
15718 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15719 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15720 unsigned StackAlign = TFI.getStackAlignment();
15721 Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15722 if (Align > StackAlign)
15723 Result = DAG.getNode(ISD::AND, dl, VT, Result,
15724 DAG.getConstant(-(uint64_t)Align, dl, VT));
15725 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain
15726 } else if (SplitStack) {
15727 MachineRegisterInfo &MRI = MF.getRegInfo();
15730 // The 64 bit implementation of segmented stacks needs to clobber both r10
15731 // r11. This makes it impossible to use it along with nested parameters.
15732 const Function *F = MF.getFunction();
15734 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15736 if (I->hasNestAttr())
15737 report_fatal_error("Cannot use segmented stacks with functions that "
15738 "have nested arguments.");
15741 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15742 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15743 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15744 Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15745 DAG.getRegister(Vreg, SPTy));
15748 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15750 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15751 Flag = Chain.getValue(1);
15752 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15754 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15756 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15757 unsigned SPReg = RegInfo->getStackRegister();
15758 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15759 Chain = SP.getValue(1);
15762 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15763 DAG.getConstant(-(uint64_t)Align, dl, VT));
15764 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15770 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15771 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
15773 SDValue Ops[2] = {Result, Chain};
15774 return DAG.getMergeValues(Ops, dl);
15777 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15778 MachineFunction &MF = DAG.getMachineFunction();
15779 auto PtrVT = getPointerTy(MF.getDataLayout());
15780 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15782 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15785 if (!Subtarget->is64Bit() ||
15786 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15787 // vastart just stores the address of the VarArgsFrameIndex slot into the
15788 // memory location argument.
15789 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15790 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15791 MachinePointerInfo(SV), false, false, 0);
15795 // gp_offset (0 - 6 * 8)
15796 // fp_offset (48 - 48 + 8 * 16)
15797 // overflow_arg_area (point to parameters coming in memory).
15799 SmallVector<SDValue, 8> MemOps;
15800 SDValue FIN = Op.getOperand(1);
15802 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15803 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15805 FIN, MachinePointerInfo(SV), false, false, 0);
15806 MemOps.push_back(Store);
15809 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15810 Store = DAG.getStore(Op.getOperand(0), DL,
15811 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15813 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15814 MemOps.push_back(Store);
15816 // Store ptr to overflow_arg_area
15817 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15818 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15819 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15820 MachinePointerInfo(SV, 8),
15822 MemOps.push_back(Store);
15824 // Store ptr to reg_save_area.
15825 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15826 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15827 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15828 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15829 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15830 MemOps.push_back(Store);
15831 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15834 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15835 assert(Subtarget->is64Bit() &&
15836 "LowerVAARG only handles 64-bit va_arg!");
15837 assert(Op.getNode()->getNumOperands() == 4);
15839 MachineFunction &MF = DAG.getMachineFunction();
15840 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15841 // The Win64 ABI uses char* instead of a structure.
15842 return DAG.expandVAArg(Op.getNode());
15844 SDValue Chain = Op.getOperand(0);
15845 SDValue SrcPtr = Op.getOperand(1);
15846 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15847 unsigned Align = Op.getConstantOperandVal(3);
15850 EVT ArgVT = Op.getNode()->getValueType(0);
15851 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15852 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15855 // Decide which area this value should be read from.
15856 // TODO: Implement the AMD64 ABI in its entirety. This simple
15857 // selection mechanism works only for the basic types.
15858 if (ArgVT == MVT::f80) {
15859 llvm_unreachable("va_arg for f80 not yet implemented");
15860 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15861 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15862 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15863 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15865 llvm_unreachable("Unhandled argument type in LowerVAARG");
15868 if (ArgMode == 2) {
15869 // Sanity Check: Make sure using fp_offset makes sense.
15870 assert(!Subtarget->useSoftFloat() &&
15871 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15872 Subtarget->hasSSE1());
15875 // Insert VAARG_64 node into the DAG
15876 // VAARG_64 returns two values: Variable Argument Address, Chain
15877 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15878 DAG.getConstant(ArgMode, dl, MVT::i8),
15879 DAG.getConstant(Align, dl, MVT::i32)};
15880 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15881 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15882 VTs, InstOps, MVT::i64,
15883 MachinePointerInfo(SV),
15885 /*Volatile=*/false,
15887 /*WriteMem=*/true);
15888 Chain = VAARG.getValue(1);
15890 // Load the next argument and return it
15891 return DAG.getLoad(ArgVT, dl,
15894 MachinePointerInfo(),
15895 false, false, false, 0);
15898 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15899 SelectionDAG &DAG) {
15900 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15901 // where a va_list is still an i8*.
15902 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15903 if (Subtarget->isCallingConvWin64(
15904 DAG.getMachineFunction().getFunction()->getCallingConv()))
15905 // Probably a Win64 va_copy.
15906 return DAG.expandVACopy(Op.getNode());
15908 SDValue Chain = Op.getOperand(0);
15909 SDValue DstPtr = Op.getOperand(1);
15910 SDValue SrcPtr = Op.getOperand(2);
15911 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15912 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15915 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15916 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15918 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15921 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15922 // amount is a constant. Takes immediate version of shift as input.
15923 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15924 SDValue SrcOp, uint64_t ShiftAmt,
15925 SelectionDAG &DAG) {
15926 MVT ElementType = VT.getVectorElementType();
15928 // Fold this packed shift into its first operand if ShiftAmt is 0.
15932 // Check for ShiftAmt >= element width
15933 if (ShiftAmt >= ElementType.getSizeInBits()) {
15934 if (Opc == X86ISD::VSRAI)
15935 ShiftAmt = ElementType.getSizeInBits() - 1;
15937 return DAG.getConstant(0, dl, VT);
15940 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15941 && "Unknown target vector shift-by-constant node");
15943 // Fold this packed vector shift into a build vector if SrcOp is a
15944 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15945 if (VT == SrcOp.getSimpleValueType() &&
15946 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15947 SmallVector<SDValue, 8> Elts;
15948 unsigned NumElts = SrcOp->getNumOperands();
15949 ConstantSDNode *ND;
15952 default: llvm_unreachable(nullptr);
15953 case X86ISD::VSHLI:
15954 for (unsigned i=0; i!=NumElts; ++i) {
15955 SDValue CurrentOp = SrcOp->getOperand(i);
15956 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15957 Elts.push_back(CurrentOp);
15960 ND = cast<ConstantSDNode>(CurrentOp);
15961 const APInt &C = ND->getAPIntValue();
15962 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15965 case X86ISD::VSRLI:
15966 for (unsigned i=0; i!=NumElts; ++i) {
15967 SDValue CurrentOp = SrcOp->getOperand(i);
15968 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15969 Elts.push_back(CurrentOp);
15972 ND = cast<ConstantSDNode>(CurrentOp);
15973 const APInt &C = ND->getAPIntValue();
15974 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15977 case X86ISD::VSRAI:
15978 for (unsigned i=0; i!=NumElts; ++i) {
15979 SDValue CurrentOp = SrcOp->getOperand(i);
15980 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15981 Elts.push_back(CurrentOp);
15984 ND = cast<ConstantSDNode>(CurrentOp);
15985 const APInt &C = ND->getAPIntValue();
15986 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15991 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15994 return DAG.getNode(Opc, dl, VT, SrcOp,
15995 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15998 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15999 // may or may not be a constant. Takes immediate version of shift as input.
16000 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
16001 SDValue SrcOp, SDValue ShAmt,
16002 SelectionDAG &DAG) {
16003 MVT SVT = ShAmt.getSimpleValueType();
16004 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
16006 // Catch shift-by-constant.
16007 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16008 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16009 CShAmt->getZExtValue(), DAG);
16011 // Change opcode to non-immediate version
16013 default: llvm_unreachable("Unknown target vector shift node");
16014 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16015 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16016 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16019 const X86Subtarget &Subtarget =
16020 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16021 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16022 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16023 // Let the shuffle legalizer expand this shift amount node.
16024 SDValue Op0 = ShAmt.getOperand(0);
16025 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16026 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16028 // Need to build a vector containing shift amount.
16029 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16030 SmallVector<SDValue, 4> ShOps;
16031 ShOps.push_back(ShAmt);
16032 if (SVT == MVT::i32) {
16033 ShOps.push_back(DAG.getConstant(0, dl, SVT));
16034 ShOps.push_back(DAG.getUNDEF(SVT));
16036 ShOps.push_back(DAG.getUNDEF(SVT));
16038 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16039 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16042 // The return type has to be a 128-bit type with the same element
16043 // type as the input type.
16044 MVT EltVT = VT.getVectorElementType();
16045 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16047 ShAmt = DAG.getBitcast(ShVT, ShAmt);
16048 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16051 /// \brief Return Mask with the necessary casting or extending
16052 /// for \p Mask according to \p MaskVT when lowering masking intrinsics
16053 static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
16054 const X86Subtarget *Subtarget,
16055 SelectionDAG &DAG, SDLoc dl) {
16057 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
16058 // Mask should be extended
16059 Mask = DAG.getNode(ISD::ANY_EXTEND, dl,
16060 MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask);
16063 if (Mask.getSimpleValueType() == MVT::i64 && Subtarget->is32Bit()) {
16064 if (MaskVT == MVT::v64i1) {
16065 assert(Subtarget->hasBWI() && "Expected AVX512BW target!");
16066 // In case 32bit mode, bitcast i64 is illegal, extend/split it.
16068 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16069 DAG.getConstant(0, dl, MVT::i32));
16070 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16071 DAG.getConstant(1, dl, MVT::i32));
16073 Lo = DAG.getBitcast(MVT::v32i1, Lo);
16074 Hi = DAG.getBitcast(MVT::v32i1, Hi);
16076 return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
16078 // MaskVT require < 64bit. Truncate mask (should succeed in any case),
16080 MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
16081 return DAG.getBitcast(MaskVT,
16082 DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask));
16086 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16087 Mask.getSimpleValueType().getSizeInBits());
16088 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16089 // are extracted by EXTRACT_SUBVECTOR.
16090 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16091 DAG.getBitcast(BitcastVT, Mask),
16092 DAG.getIntPtrConstant(0, dl));
16096 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16097 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16098 /// necessary casting or extending for \p Mask when lowering masking intrinsics
16099 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16100 SDValue PreservedSrc,
16101 const X86Subtarget *Subtarget,
16102 SelectionDAG &DAG) {
16103 MVT VT = Op.getSimpleValueType();
16104 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16105 unsigned OpcodeSelect = ISD::VSELECT;
16108 if (isAllOnesConstant(Mask))
16111 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16113 switch (Op.getOpcode()) {
16115 case X86ISD::PCMPEQM:
16116 case X86ISD::PCMPGTM:
16118 case X86ISD::CMPMU:
16119 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16120 case X86ISD::VFPCLASS:
16121 case X86ISD::VFPCLASSS:
16122 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16123 case X86ISD::VTRUNC:
16124 case X86ISD::VTRUNCS:
16125 case X86ISD::VTRUNCUS:
16126 // We can't use ISD::VSELECT here because it is not always "Legal"
16127 // for the destination type. For example vpmovqb require only AVX512
16128 // and vselect that can operate on byte element type require BWI
16129 OpcodeSelect = X86ISD::SELECT;
16132 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16133 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16134 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16137 /// \brief Creates an SDNode for a predicated scalar operation.
16138 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16139 /// The mask is coming as MVT::i8 and it should be truncated
16140 /// to MVT::i1 while lowering masking intrinsics.
16141 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16142 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16143 /// for a scalar instruction.
16144 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16145 SDValue PreservedSrc,
16146 const X86Subtarget *Subtarget,
16147 SelectionDAG &DAG) {
16148 if (isAllOnesConstant(Mask))
16151 MVT VT = Op.getSimpleValueType();
16153 // The mask should be of type MVT::i1
16154 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16156 if (Op.getOpcode() == X86ISD::FSETCC)
16157 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16158 if (Op.getOpcode() == X86ISD::VFPCLASS ||
16159 Op.getOpcode() == X86ISD::VFPCLASSS)
16160 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16162 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16163 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16164 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16167 static int getSEHRegistrationNodeSize(const Function *Fn) {
16168 if (!Fn->hasPersonalityFn())
16169 report_fatal_error(
16170 "querying registration node size for function without personality");
16171 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16172 // WinEHStatePass for the full struct definition.
16173 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16174 case EHPersonality::MSVC_X86SEH: return 24;
16175 case EHPersonality::MSVC_CXX: return 16;
16178 report_fatal_error(
16179 "can only recover FP for 32-bit MSVC EH personality functions");
16182 /// When the MSVC runtime transfers control to us, either to an outlined
16183 /// function or when returning to a parent frame after catching an exception, we
16184 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16185 /// Here's the math:
16186 /// RegNodeBase = EntryEBP - RegNodeSize
16187 /// ParentFP = RegNodeBase - ParentFrameOffset
16188 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16189 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16190 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16191 SDValue EntryEBP) {
16192 MachineFunction &MF = DAG.getMachineFunction();
16195 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16196 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16198 // It's possible that the parent function no longer has a personality function
16199 // if the exceptional code was optimized away, in which case we just return
16200 // the incoming EBP.
16201 if (!Fn->hasPersonalityFn())
16204 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16205 // registration, or the .set_setframe offset.
16206 MCSymbol *OffsetSym =
16207 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16208 GlobalValue::getRealLinkageName(Fn->getName()));
16209 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16210 SDValue ParentFrameOffset =
16211 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16213 // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after
16214 // prologue to RBP in the parent function.
16215 const X86Subtarget &Subtarget =
16216 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16217 if (Subtarget.is64Bit())
16218 return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset);
16220 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16221 // RegNodeBase = EntryEBP - RegNodeSize
16222 // ParentFP = RegNodeBase - ParentFrameOffset
16223 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16224 DAG.getConstant(RegNodeSize, dl, PtrVT));
16225 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset);
16228 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16229 SelectionDAG &DAG) {
16231 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16232 MVT VT = Op.getSimpleValueType();
16233 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16235 switch(IntrData->Type) {
16236 case INTR_TYPE_1OP:
16237 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16238 case INTR_TYPE_2OP:
16239 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16241 case INTR_TYPE_2OP_IMM8:
16242 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16243 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16244 case INTR_TYPE_3OP:
16245 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16246 Op.getOperand(2), Op.getOperand(3));
16247 case INTR_TYPE_4OP:
16248 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16249 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16250 case INTR_TYPE_1OP_MASK_RM: {
16251 SDValue Src = Op.getOperand(1);
16252 SDValue PassThru = Op.getOperand(2);
16253 SDValue Mask = Op.getOperand(3);
16254 SDValue RoundingMode;
16255 // We allways add rounding mode to the Node.
16256 // If the rounding mode is not specified, we add the
16257 // "current direction" mode.
16258 if (Op.getNumOperands() == 4)
16260 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16262 RoundingMode = Op.getOperand(4);
16263 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16264 if (IntrWithRoundingModeOpcode != 0)
16265 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16266 X86::STATIC_ROUNDING::CUR_DIRECTION)
16267 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16268 dl, Op.getValueType(), Src, RoundingMode),
16269 Mask, PassThru, Subtarget, DAG);
16270 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16272 Mask, PassThru, Subtarget, DAG);
16274 case INTR_TYPE_1OP_MASK: {
16275 SDValue Src = Op.getOperand(1);
16276 SDValue PassThru = Op.getOperand(2);
16277 SDValue Mask = Op.getOperand(3);
16278 // We add rounding mode to the Node when
16279 // - RM Opcode is specified and
16280 // - RM is not "current direction".
16281 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16282 if (IntrWithRoundingModeOpcode != 0) {
16283 SDValue Rnd = Op.getOperand(4);
16284 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16285 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16286 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16287 dl, Op.getValueType(),
16289 Mask, PassThru, Subtarget, DAG);
16292 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16293 Mask, PassThru, Subtarget, DAG);
16295 case INTR_TYPE_SCALAR_MASK: {
16296 SDValue Src1 = Op.getOperand(1);
16297 SDValue Src2 = Op.getOperand(2);
16298 SDValue passThru = Op.getOperand(3);
16299 SDValue Mask = Op.getOperand(4);
16300 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16301 Mask, passThru, Subtarget, DAG);
16303 case INTR_TYPE_SCALAR_MASK_RM: {
16304 SDValue Src1 = Op.getOperand(1);
16305 SDValue Src2 = Op.getOperand(2);
16306 SDValue Src0 = Op.getOperand(3);
16307 SDValue Mask = Op.getOperand(4);
16308 // There are 2 kinds of intrinsics in this group:
16309 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16310 // (2) With rounding mode and sae - 7 operands.
16311 if (Op.getNumOperands() == 6) {
16312 SDValue Sae = Op.getOperand(5);
16313 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16314 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16316 Mask, Src0, Subtarget, DAG);
16318 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16319 SDValue RoundingMode = Op.getOperand(5);
16320 SDValue Sae = Op.getOperand(6);
16321 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16322 RoundingMode, Sae),
16323 Mask, Src0, Subtarget, DAG);
16325 case INTR_TYPE_2OP_MASK:
16326 case INTR_TYPE_2OP_IMM8_MASK: {
16327 SDValue Src1 = Op.getOperand(1);
16328 SDValue Src2 = Op.getOperand(2);
16329 SDValue PassThru = Op.getOperand(3);
16330 SDValue Mask = Op.getOperand(4);
16332 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16333 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16335 // We specify 2 possible opcodes for intrinsics with rounding modes.
16336 // First, we check if the intrinsic may have non-default rounding mode,
16337 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16338 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16339 if (IntrWithRoundingModeOpcode != 0) {
16340 SDValue Rnd = Op.getOperand(5);
16341 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16342 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16343 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16344 dl, Op.getValueType(),
16346 Mask, PassThru, Subtarget, DAG);
16349 // TODO: Intrinsics should have fast-math-flags to propagate.
16350 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16351 Mask, PassThru, Subtarget, DAG);
16353 case INTR_TYPE_2OP_MASK_RM: {
16354 SDValue Src1 = Op.getOperand(1);
16355 SDValue Src2 = Op.getOperand(2);
16356 SDValue PassThru = Op.getOperand(3);
16357 SDValue Mask = Op.getOperand(4);
16358 // We specify 2 possible modes for intrinsics, with/without rounding
16360 // First, we check if the intrinsic have rounding mode (6 operands),
16361 // if not, we set rounding mode to "current".
16363 if (Op.getNumOperands() == 6)
16364 Rnd = Op.getOperand(5);
16366 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16367 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16369 Mask, PassThru, Subtarget, DAG);
16371 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16372 SDValue Src1 = Op.getOperand(1);
16373 SDValue Src2 = Op.getOperand(2);
16374 SDValue Src3 = Op.getOperand(3);
16375 SDValue PassThru = Op.getOperand(4);
16376 SDValue Mask = Op.getOperand(5);
16377 SDValue Sae = Op.getOperand(6);
16379 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16381 Mask, PassThru, Subtarget, DAG);
16383 case INTR_TYPE_3OP_MASK_RM: {
16384 SDValue Src1 = Op.getOperand(1);
16385 SDValue Src2 = Op.getOperand(2);
16386 SDValue Imm = Op.getOperand(3);
16387 SDValue PassThru = Op.getOperand(4);
16388 SDValue Mask = Op.getOperand(5);
16389 // We specify 2 possible modes for intrinsics, with/without rounding
16391 // First, we check if the intrinsic have rounding mode (7 operands),
16392 // if not, we set rounding mode to "current".
16394 if (Op.getNumOperands() == 7)
16395 Rnd = Op.getOperand(6);
16397 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16398 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16399 Src1, Src2, Imm, Rnd),
16400 Mask, PassThru, Subtarget, DAG);
16402 case INTR_TYPE_3OP_IMM8_MASK:
16403 case INTR_TYPE_3OP_MASK:
16404 case INSERT_SUBVEC: {
16405 SDValue Src1 = Op.getOperand(1);
16406 SDValue Src2 = Op.getOperand(2);
16407 SDValue Src3 = Op.getOperand(3);
16408 SDValue PassThru = Op.getOperand(4);
16409 SDValue Mask = Op.getOperand(5);
16411 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16412 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16413 else if (IntrData->Type == INSERT_SUBVEC) {
16414 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16415 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16416 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16417 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16418 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16421 // We specify 2 possible opcodes for intrinsics with rounding modes.
16422 // First, we check if the intrinsic may have non-default rounding mode,
16423 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16424 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16425 if (IntrWithRoundingModeOpcode != 0) {
16426 SDValue Rnd = Op.getOperand(6);
16427 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16428 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16429 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16430 dl, Op.getValueType(),
16431 Src1, Src2, Src3, Rnd),
16432 Mask, PassThru, Subtarget, DAG);
16435 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16437 Mask, PassThru, Subtarget, DAG);
16439 case VPERM_3OP_MASKZ:
16440 case VPERM_3OP_MASK:{
16441 // Src2 is the PassThru
16442 SDValue Src1 = Op.getOperand(1);
16443 SDValue Src2 = Op.getOperand(2);
16444 SDValue Src3 = Op.getOperand(3);
16445 SDValue Mask = Op.getOperand(4);
16446 MVT VT = Op.getSimpleValueType();
16447 SDValue PassThru = SDValue();
16449 // set PassThru element
16450 if (IntrData->Type == VPERM_3OP_MASKZ)
16451 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16453 PassThru = DAG.getBitcast(VT, Src2);
16455 // Swap Src1 and Src2 in the node creation
16456 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16457 dl, Op.getValueType(),
16459 Mask, PassThru, Subtarget, DAG);
16463 case FMA_OP_MASK: {
16464 SDValue Src1 = Op.getOperand(1);
16465 SDValue Src2 = Op.getOperand(2);
16466 SDValue Src3 = Op.getOperand(3);
16467 SDValue Mask = Op.getOperand(4);
16468 MVT VT = Op.getSimpleValueType();
16469 SDValue PassThru = SDValue();
16471 // set PassThru element
16472 if (IntrData->Type == FMA_OP_MASKZ)
16473 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16474 else if (IntrData->Type == FMA_OP_MASK3)
16479 // We specify 2 possible opcodes for intrinsics with rounding modes.
16480 // First, we check if the intrinsic may have non-default rounding mode,
16481 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16482 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16483 if (IntrWithRoundingModeOpcode != 0) {
16484 SDValue Rnd = Op.getOperand(5);
16485 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16486 X86::STATIC_ROUNDING::CUR_DIRECTION)
16487 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16488 dl, Op.getValueType(),
16489 Src1, Src2, Src3, Rnd),
16490 Mask, PassThru, Subtarget, DAG);
16492 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16493 dl, Op.getValueType(),
16495 Mask, PassThru, Subtarget, DAG);
16497 case TERLOG_OP_MASK:
16498 case TERLOG_OP_MASKZ: {
16499 SDValue Src1 = Op.getOperand(1);
16500 SDValue Src2 = Op.getOperand(2);
16501 SDValue Src3 = Op.getOperand(3);
16502 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16503 SDValue Mask = Op.getOperand(5);
16504 MVT VT = Op.getSimpleValueType();
16505 SDValue PassThru = Src1;
16506 // Set PassThru element.
16507 if (IntrData->Type == TERLOG_OP_MASKZ)
16508 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16510 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16511 Src1, Src2, Src3, Src4),
16512 Mask, PassThru, Subtarget, DAG);
16515 // FPclass intrinsics with mask
16516 SDValue Src1 = Op.getOperand(1);
16517 MVT VT = Src1.getSimpleValueType();
16518 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16519 SDValue Imm = Op.getOperand(2);
16520 SDValue Mask = Op.getOperand(3);
16521 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16522 Mask.getSimpleValueType().getSizeInBits());
16523 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16524 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16525 DAG.getTargetConstant(0, dl, MaskVT),
16527 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16528 DAG.getUNDEF(BitcastVT), FPclassMask,
16529 DAG.getIntPtrConstant(0, dl));
16530 return DAG.getBitcast(Op.getValueType(), Res);
16533 SDValue Src1 = Op.getOperand(1);
16534 SDValue Imm = Op.getOperand(2);
16535 SDValue Mask = Op.getOperand(3);
16536 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16537 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16538 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16539 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16542 case CMP_MASK_CC: {
16543 // Comparison intrinsics with masks.
16544 // Example of transformation:
16545 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16546 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16548 // (v8i1 (insert_subvector undef,
16549 // (v2i1 (and (PCMPEQM %a, %b),
16550 // (extract_subvector
16551 // (v8i1 (bitcast %mask)), 0))), 0))))
16552 MVT VT = Op.getOperand(1).getSimpleValueType();
16553 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16554 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16555 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16556 Mask.getSimpleValueType().getSizeInBits());
16558 if (IntrData->Type == CMP_MASK_CC) {
16559 SDValue CC = Op.getOperand(3);
16560 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16561 // We specify 2 possible opcodes for intrinsics with rounding modes.
16562 // First, we check if the intrinsic may have non-default rounding mode,
16563 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16564 if (IntrData->Opc1 != 0) {
16565 SDValue Rnd = Op.getOperand(5);
16566 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16567 X86::STATIC_ROUNDING::CUR_DIRECTION)
16568 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16569 Op.getOperand(2), CC, Rnd);
16571 //default rounding mode
16573 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16574 Op.getOperand(2), CC);
16577 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16578 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16581 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16582 DAG.getTargetConstant(0, dl,
16585 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16586 DAG.getUNDEF(BitcastVT), CmpMask,
16587 DAG.getIntPtrConstant(0, dl));
16588 return DAG.getBitcast(Op.getValueType(), Res);
16590 case CMP_MASK_SCALAR_CC: {
16591 SDValue Src1 = Op.getOperand(1);
16592 SDValue Src2 = Op.getOperand(2);
16593 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16594 SDValue Mask = Op.getOperand(4);
16597 if (IntrData->Opc1 != 0) {
16598 SDValue Rnd = Op.getOperand(5);
16599 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16600 X86::STATIC_ROUNDING::CUR_DIRECTION)
16601 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16603 //default rounding mode
16605 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16607 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16608 DAG.getTargetConstant(0, dl,
16612 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16613 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16614 DAG.getValueType(MVT::i1));
16616 case COMI: { // Comparison intrinsics
16617 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16618 SDValue LHS = Op.getOperand(1);
16619 SDValue RHS = Op.getOperand(2);
16620 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16621 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16622 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16623 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16624 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16625 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16627 case COMI_RM: { // Comparison intrinsics with Sae
16628 SDValue LHS = Op.getOperand(1);
16629 SDValue RHS = Op.getOperand(2);
16630 SDValue CC = Op.getOperand(3);
16631 SDValue Sae = Op.getOperand(4);
16632 auto ComiType = TranslateX86ConstCondToX86CC(CC);
16633 // choose between ordered and unordered (comi/ucomi)
16634 unsigned comiOp = std::get<0>(ComiType) ? IntrData->Opc0 : IntrData->Opc1;
16636 if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
16637 X86::STATIC_ROUNDING::CUR_DIRECTION)
16638 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
16640 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
16641 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16642 DAG.getConstant(std::get<1>(ComiType), dl, MVT::i8), Cond);
16643 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16646 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16647 Op.getOperand(1), Op.getOperand(2), DAG);
16649 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16650 Op.getSimpleValueType(),
16652 Op.getOperand(2), DAG),
16653 Op.getOperand(4), Op.getOperand(3), Subtarget,
16655 case COMPRESS_EXPAND_IN_REG: {
16656 SDValue Mask = Op.getOperand(3);
16657 SDValue DataToCompress = Op.getOperand(1);
16658 SDValue PassThru = Op.getOperand(2);
16659 if (isAllOnesConstant(Mask)) // return data as is
16660 return Op.getOperand(1);
16662 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16664 Mask, PassThru, Subtarget, DAG);
16667 SDValue Mask = Op.getOperand(1);
16668 MVT MaskVT = MVT::getVectorVT(MVT::i1, Mask.getSimpleValueType().getSizeInBits());
16669 Mask = DAG.getBitcast(MaskVT, Mask);
16670 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16673 SDValue Mask = Op.getOperand(3);
16674 MVT VT = Op.getSimpleValueType();
16675 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16676 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16677 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16681 MVT VT = Op.getSimpleValueType();
16682 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2);
16684 SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl);
16685 SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl);
16686 // Arguments should be swapped.
16687 SDValue Res = DAG.getNode(IntrData->Opc0, dl,
16688 MVT::getVectorVT(MVT::i1, VT.getSizeInBits()),
16690 return DAG.getBitcast(VT, Res);
16698 default: return SDValue(); // Don't custom lower most intrinsics.
16700 case Intrinsic::x86_avx2_permd:
16701 case Intrinsic::x86_avx2_permps:
16702 // Operands intentionally swapped. Mask is last operand to intrinsic,
16703 // but second operand for node/instruction.
16704 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16705 Op.getOperand(2), Op.getOperand(1));
16707 // ptest and testp intrinsics. The intrinsic these come from are designed to
16708 // return an integer value, not just an instruction so lower it to the ptest
16709 // or testp pattern and a setcc for the result.
16710 case Intrinsic::x86_sse41_ptestz:
16711 case Intrinsic::x86_sse41_ptestc:
16712 case Intrinsic::x86_sse41_ptestnzc:
16713 case Intrinsic::x86_avx_ptestz_256:
16714 case Intrinsic::x86_avx_ptestc_256:
16715 case Intrinsic::x86_avx_ptestnzc_256:
16716 case Intrinsic::x86_avx_vtestz_ps:
16717 case Intrinsic::x86_avx_vtestc_ps:
16718 case Intrinsic::x86_avx_vtestnzc_ps:
16719 case Intrinsic::x86_avx_vtestz_pd:
16720 case Intrinsic::x86_avx_vtestc_pd:
16721 case Intrinsic::x86_avx_vtestnzc_pd:
16722 case Intrinsic::x86_avx_vtestz_ps_256:
16723 case Intrinsic::x86_avx_vtestc_ps_256:
16724 case Intrinsic::x86_avx_vtestnzc_ps_256:
16725 case Intrinsic::x86_avx_vtestz_pd_256:
16726 case Intrinsic::x86_avx_vtestc_pd_256:
16727 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16728 bool IsTestPacked = false;
16731 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16732 case Intrinsic::x86_avx_vtestz_ps:
16733 case Intrinsic::x86_avx_vtestz_pd:
16734 case Intrinsic::x86_avx_vtestz_ps_256:
16735 case Intrinsic::x86_avx_vtestz_pd_256:
16736 IsTestPacked = true; // Fallthrough
16737 case Intrinsic::x86_sse41_ptestz:
16738 case Intrinsic::x86_avx_ptestz_256:
16740 X86CC = X86::COND_E;
16742 case Intrinsic::x86_avx_vtestc_ps:
16743 case Intrinsic::x86_avx_vtestc_pd:
16744 case Intrinsic::x86_avx_vtestc_ps_256:
16745 case Intrinsic::x86_avx_vtestc_pd_256:
16746 IsTestPacked = true; // Fallthrough
16747 case Intrinsic::x86_sse41_ptestc:
16748 case Intrinsic::x86_avx_ptestc_256:
16750 X86CC = X86::COND_B;
16752 case Intrinsic::x86_avx_vtestnzc_ps:
16753 case Intrinsic::x86_avx_vtestnzc_pd:
16754 case Intrinsic::x86_avx_vtestnzc_ps_256:
16755 case Intrinsic::x86_avx_vtestnzc_pd_256:
16756 IsTestPacked = true; // Fallthrough
16757 case Intrinsic::x86_sse41_ptestnzc:
16758 case Intrinsic::x86_avx_ptestnzc_256:
16760 X86CC = X86::COND_A;
16764 SDValue LHS = Op.getOperand(1);
16765 SDValue RHS = Op.getOperand(2);
16766 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16767 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16768 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16769 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16770 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16772 case Intrinsic::x86_avx512_kortestz_w:
16773 case Intrinsic::x86_avx512_kortestc_w: {
16774 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16775 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16776 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16777 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16778 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16779 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16780 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16783 case Intrinsic::x86_sse42_pcmpistria128:
16784 case Intrinsic::x86_sse42_pcmpestria128:
16785 case Intrinsic::x86_sse42_pcmpistric128:
16786 case Intrinsic::x86_sse42_pcmpestric128:
16787 case Intrinsic::x86_sse42_pcmpistrio128:
16788 case Intrinsic::x86_sse42_pcmpestrio128:
16789 case Intrinsic::x86_sse42_pcmpistris128:
16790 case Intrinsic::x86_sse42_pcmpestris128:
16791 case Intrinsic::x86_sse42_pcmpistriz128:
16792 case Intrinsic::x86_sse42_pcmpestriz128: {
16796 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16797 case Intrinsic::x86_sse42_pcmpistria128:
16798 Opcode = X86ISD::PCMPISTRI;
16799 X86CC = X86::COND_A;
16801 case Intrinsic::x86_sse42_pcmpestria128:
16802 Opcode = X86ISD::PCMPESTRI;
16803 X86CC = X86::COND_A;
16805 case Intrinsic::x86_sse42_pcmpistric128:
16806 Opcode = X86ISD::PCMPISTRI;
16807 X86CC = X86::COND_B;
16809 case Intrinsic::x86_sse42_pcmpestric128:
16810 Opcode = X86ISD::PCMPESTRI;
16811 X86CC = X86::COND_B;
16813 case Intrinsic::x86_sse42_pcmpistrio128:
16814 Opcode = X86ISD::PCMPISTRI;
16815 X86CC = X86::COND_O;
16817 case Intrinsic::x86_sse42_pcmpestrio128:
16818 Opcode = X86ISD::PCMPESTRI;
16819 X86CC = X86::COND_O;
16821 case Intrinsic::x86_sse42_pcmpistris128:
16822 Opcode = X86ISD::PCMPISTRI;
16823 X86CC = X86::COND_S;
16825 case Intrinsic::x86_sse42_pcmpestris128:
16826 Opcode = X86ISD::PCMPESTRI;
16827 X86CC = X86::COND_S;
16829 case Intrinsic::x86_sse42_pcmpistriz128:
16830 Opcode = X86ISD::PCMPISTRI;
16831 X86CC = X86::COND_E;
16833 case Intrinsic::x86_sse42_pcmpestriz128:
16834 Opcode = X86ISD::PCMPESTRI;
16835 X86CC = X86::COND_E;
16838 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16839 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16840 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16841 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16842 DAG.getConstant(X86CC, dl, MVT::i8),
16843 SDValue(PCMP.getNode(), 1));
16844 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16847 case Intrinsic::x86_sse42_pcmpistri128:
16848 case Intrinsic::x86_sse42_pcmpestri128: {
16850 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16851 Opcode = X86ISD::PCMPISTRI;
16853 Opcode = X86ISD::PCMPESTRI;
16855 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16856 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16857 return DAG.getNode(Opcode, dl, VTs, NewOps);
16860 case Intrinsic::x86_seh_lsda: {
16861 // Compute the symbol for the LSDA. We know it'll get emitted later.
16862 MachineFunction &MF = DAG.getMachineFunction();
16863 SDValue Op1 = Op.getOperand(1);
16864 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16865 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16866 GlobalValue::getRealLinkageName(Fn->getName()));
16868 // Generate a simple absolute symbol reference. This intrinsic is only
16869 // supported on 32-bit Windows, which isn't PIC.
16870 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16871 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16874 case Intrinsic::x86_seh_recoverfp: {
16875 SDValue FnOp = Op.getOperand(1);
16876 SDValue IncomingFPOp = Op.getOperand(2);
16877 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16878 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16880 report_fatal_error(
16881 "llvm.x86.seh.recoverfp must take a function as the first argument");
16882 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16885 case Intrinsic::localaddress: {
16886 // Returns one of the stack, base, or frame pointer registers, depending on
16887 // which is used to reference local variables.
16888 MachineFunction &MF = DAG.getMachineFunction();
16889 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16891 if (RegInfo->hasBasePointer(MF))
16892 Reg = RegInfo->getBaseRegister();
16893 else // This function handles the SP or FP case.
16894 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16895 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16900 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16901 SDValue Src, SDValue Mask, SDValue Base,
16902 SDValue Index, SDValue ScaleOp, SDValue Chain,
16903 const X86Subtarget * Subtarget) {
16905 auto *C = cast<ConstantSDNode>(ScaleOp);
16906 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16907 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16908 Index.getSimpleValueType().getVectorNumElements());
16910 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16912 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16914 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16915 Mask.getSimpleValueType().getSizeInBits());
16917 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16918 // are extracted by EXTRACT_SUBVECTOR.
16919 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16920 DAG.getBitcast(BitcastVT, Mask),
16921 DAG.getIntPtrConstant(0, dl));
16923 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16924 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16925 SDValue Segment = DAG.getRegister(0, MVT::i32);
16926 if (Src.getOpcode() == ISD::UNDEF)
16927 Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl);
16928 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16929 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16930 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16931 return DAG.getMergeValues(RetOps, dl);
16934 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16935 SDValue Src, SDValue Mask, SDValue Base,
16936 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16938 auto *C = cast<ConstantSDNode>(ScaleOp);
16939 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16940 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16941 SDValue Segment = DAG.getRegister(0, MVT::i32);
16942 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16943 Index.getSimpleValueType().getVectorNumElements());
16945 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16947 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16949 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16950 Mask.getSimpleValueType().getSizeInBits());
16952 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16953 // are extracted by EXTRACT_SUBVECTOR.
16954 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16955 DAG.getBitcast(BitcastVT, Mask),
16956 DAG.getIntPtrConstant(0, dl));
16958 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16959 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16960 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16961 return SDValue(Res, 1);
16964 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16965 SDValue Mask, SDValue Base, SDValue Index,
16966 SDValue ScaleOp, SDValue Chain) {
16968 auto *C = cast<ConstantSDNode>(ScaleOp);
16969 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16970 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16971 SDValue Segment = DAG.getRegister(0, MVT::i32);
16973 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16975 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16977 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16979 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16980 //SDVTList VTs = DAG.getVTList(MVT::Other);
16981 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16982 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16983 return SDValue(Res, 0);
16986 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16987 // read performance monitor counters (x86_rdpmc).
16988 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16989 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16990 SmallVectorImpl<SDValue> &Results) {
16991 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16992 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16995 // The ECX register is used to select the index of the performance counter
16997 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16999 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17001 // Reads the content of a 64-bit performance counter and returns it in the
17002 // registers EDX:EAX.
17003 if (Subtarget->is64Bit()) {
17004 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17005 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17008 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17009 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17012 Chain = HI.getValue(1);
17014 if (Subtarget->is64Bit()) {
17015 // The EAX register is loaded with the low-order 32 bits. The EDX register
17016 // is loaded with the supported high-order bits of the counter.
17017 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17018 DAG.getConstant(32, DL, MVT::i8));
17019 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17020 Results.push_back(Chain);
17024 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17025 SDValue Ops[] = { LO, HI };
17026 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17027 Results.push_back(Pair);
17028 Results.push_back(Chain);
17031 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17032 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17033 // also used to custom lower READCYCLECOUNTER nodes.
17034 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17035 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17036 SmallVectorImpl<SDValue> &Results) {
17037 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17038 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17041 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17042 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17043 // and the EAX register is loaded with the low-order 32 bits.
17044 if (Subtarget->is64Bit()) {
17045 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17046 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17049 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17050 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17053 SDValue Chain = HI.getValue(1);
17055 if (Opcode == X86ISD::RDTSCP_DAG) {
17056 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17058 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17059 // the ECX register. Add 'ecx' explicitly to the chain.
17060 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17062 // Explicitly store the content of ECX at the location passed in input
17063 // to the 'rdtscp' intrinsic.
17064 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17065 MachinePointerInfo(), false, false, 0);
17068 if (Subtarget->is64Bit()) {
17069 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17070 // the EAX register is loaded with the low-order 32 bits.
17071 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17072 DAG.getConstant(32, DL, MVT::i8));
17073 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17074 Results.push_back(Chain);
17078 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17079 SDValue Ops[] = { LO, HI };
17080 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17081 Results.push_back(Pair);
17082 Results.push_back(Chain);
17085 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17086 SelectionDAG &DAG) {
17087 SmallVector<SDValue, 2> Results;
17089 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17091 return DAG.getMergeValues(Results, DL);
17094 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
17095 MachineFunction &MF = DAG.getMachineFunction();
17096 SDValue Chain = Op.getOperand(0);
17097 SDValue RegNode = Op.getOperand(2);
17098 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
17100 report_fatal_error("EH registrations only live in functions using WinEH");
17102 // Cast the operand to an alloca, and remember the frame index.
17103 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
17105 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
17106 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
17108 // Return the chain operand without making any DAG nodes.
17112 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
17113 /// return truncate Store/MaskedStore Node
17114 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17118 SDValue Mask = Op.getOperand(4);
17119 SDValue DataToTruncate = Op.getOperand(3);
17120 SDValue Addr = Op.getOperand(2);
17121 SDValue Chain = Op.getOperand(0);
17123 MVT VT = DataToTruncate.getSimpleValueType();
17124 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17126 if (isAllOnesConstant(Mask)) // return just a truncate store
17127 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17128 MachinePointerInfo(), SVT, false, false,
17129 SVT.getScalarSizeInBits()/8);
17131 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17132 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17133 Mask.getSimpleValueType().getSizeInBits());
17134 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17135 // are extracted by EXTRACT_SUBVECTOR.
17136 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17137 DAG.getBitcast(BitcastVT, Mask),
17138 DAG.getIntPtrConstant(0, dl));
17140 MachineMemOperand *MMO = DAG.getMachineFunction().
17141 getMachineMemOperand(MachinePointerInfo(),
17142 MachineMemOperand::MOStore, SVT.getStoreSize(),
17143 SVT.getScalarSizeInBits()/8);
17145 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17146 VMask, SVT, MMO, true);
17149 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17150 SelectionDAG &DAG) {
17151 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17153 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17155 if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17156 return MarkEHRegistrationNode(Op, DAG);
17161 switch(IntrData->Type) {
17162 default: llvm_unreachable("Unknown Intrinsic Type");
17165 // Emit the node with the right value type.
17166 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17167 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17169 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17170 // Otherwise return the value from Rand, which is always 0, casted to i32.
17171 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17172 DAG.getConstant(1, dl, Op->getValueType(1)),
17173 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17174 SDValue(Result.getNode(), 1) };
17175 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17176 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17179 // Return { result, isValid, chain }.
17180 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17181 SDValue(Result.getNode(), 2));
17184 //gather(v1, mask, index, base, scale);
17185 SDValue Chain = Op.getOperand(0);
17186 SDValue Src = Op.getOperand(2);
17187 SDValue Base = Op.getOperand(3);
17188 SDValue Index = Op.getOperand(4);
17189 SDValue Mask = Op.getOperand(5);
17190 SDValue Scale = Op.getOperand(6);
17191 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17195 //scatter(base, mask, index, v1, scale);
17196 SDValue Chain = Op.getOperand(0);
17197 SDValue Base = Op.getOperand(2);
17198 SDValue Mask = Op.getOperand(3);
17199 SDValue Index = Op.getOperand(4);
17200 SDValue Src = Op.getOperand(5);
17201 SDValue Scale = Op.getOperand(6);
17202 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17206 SDValue Hint = Op.getOperand(6);
17207 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17208 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17209 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17210 SDValue Chain = Op.getOperand(0);
17211 SDValue Mask = Op.getOperand(2);
17212 SDValue Index = Op.getOperand(3);
17213 SDValue Base = Op.getOperand(4);
17214 SDValue Scale = Op.getOperand(5);
17215 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17217 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17219 SmallVector<SDValue, 2> Results;
17220 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17222 return DAG.getMergeValues(Results, dl);
17224 // Read Performance Monitoring Counters.
17226 SmallVector<SDValue, 2> Results;
17227 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17228 return DAG.getMergeValues(Results, dl);
17230 // XTEST intrinsics.
17232 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17233 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17234 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17235 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17237 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17238 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17239 Ret, SDValue(InTrans.getNode(), 1));
17243 SmallVector<SDValue, 2> Results;
17244 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17245 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17246 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17247 DAG.getConstant(-1, dl, MVT::i8));
17248 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17249 Op.getOperand(4), GenCF.getValue(1));
17250 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17251 Op.getOperand(5), MachinePointerInfo(),
17253 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17254 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17256 Results.push_back(SetCC);
17257 Results.push_back(Store);
17258 return DAG.getMergeValues(Results, dl);
17260 case COMPRESS_TO_MEM: {
17262 SDValue Mask = Op.getOperand(4);
17263 SDValue DataToCompress = Op.getOperand(3);
17264 SDValue Addr = Op.getOperand(2);
17265 SDValue Chain = Op.getOperand(0);
17267 MVT VT = DataToCompress.getSimpleValueType();
17268 if (isAllOnesConstant(Mask)) // return just a store
17269 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17270 MachinePointerInfo(), false, false,
17271 VT.getScalarSizeInBits()/8);
17273 SDValue Compressed =
17274 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17275 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17276 return DAG.getStore(Chain, dl, Compressed, Addr,
17277 MachinePointerInfo(), false, false,
17278 VT.getScalarSizeInBits()/8);
17280 case TRUNCATE_TO_MEM_VI8:
17281 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17282 case TRUNCATE_TO_MEM_VI16:
17283 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17284 case TRUNCATE_TO_MEM_VI32:
17285 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17286 case EXPAND_FROM_MEM: {
17288 SDValue Mask = Op.getOperand(4);
17289 SDValue PassThru = Op.getOperand(3);
17290 SDValue Addr = Op.getOperand(2);
17291 SDValue Chain = Op.getOperand(0);
17292 MVT VT = Op.getSimpleValueType();
17294 if (isAllOnesConstant(Mask)) // return just a load
17295 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17296 false, VT.getScalarSizeInBits()/8);
17298 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17299 false, false, false,
17300 VT.getScalarSizeInBits()/8);
17302 SDValue Results[] = {
17303 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17304 Mask, PassThru, Subtarget, DAG), Chain};
17305 return DAG.getMergeValues(Results, dl);
17310 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17311 SelectionDAG &DAG) const {
17312 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17313 MFI->setReturnAddressIsTaken(true);
17315 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17318 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17320 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17323 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17324 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17325 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17326 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17327 DAG.getNode(ISD::ADD, dl, PtrVT,
17328 FrameAddr, Offset),
17329 MachinePointerInfo(), false, false, false, 0);
17332 // Just load the return address.
17333 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17334 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17335 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17338 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17339 MachineFunction &MF = DAG.getMachineFunction();
17340 MachineFrameInfo *MFI = MF.getFrameInfo();
17341 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17342 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17343 EVT VT = Op.getValueType();
17345 MFI->setFrameAddressIsTaken(true);
17347 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17348 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17349 // is not possible to crawl up the stack without looking at the unwind codes
17351 int FrameAddrIndex = FuncInfo->getFAIndex();
17352 if (!FrameAddrIndex) {
17353 // Set up a frame object for the return address.
17354 unsigned SlotSize = RegInfo->getSlotSize();
17355 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17356 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17357 FuncInfo->setFAIndex(FrameAddrIndex);
17359 return DAG.getFrameIndex(FrameAddrIndex, VT);
17362 unsigned FrameReg =
17363 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17364 SDLoc dl(Op); // FIXME probably not meaningful
17365 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17366 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17367 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17368 "Invalid Frame Register!");
17369 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17371 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17372 MachinePointerInfo(),
17373 false, false, false, 0);
17377 // FIXME? Maybe this could be a TableGen attribute on some registers and
17378 // this table could be generated automatically from RegInfo.
17379 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17380 SelectionDAG &DAG) const {
17381 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17382 const MachineFunction &MF = DAG.getMachineFunction();
17384 unsigned Reg = StringSwitch<unsigned>(RegName)
17385 .Case("esp", X86::ESP)
17386 .Case("rsp", X86::RSP)
17387 .Case("ebp", X86::EBP)
17388 .Case("rbp", X86::RBP)
17391 if (Reg == X86::EBP || Reg == X86::RBP) {
17392 if (!TFI.hasFP(MF))
17393 report_fatal_error("register " + StringRef(RegName) +
17394 " is allocatable: function has no frame pointer");
17397 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17398 unsigned FrameReg =
17399 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17400 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17401 "Invalid Frame Register!");
17409 report_fatal_error("Invalid register name global variable");
17412 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17413 SelectionDAG &DAG) const {
17414 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17415 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17418 unsigned X86TargetLowering::getExceptionPointerRegister(
17419 const Constant *PersonalityFn) const {
17420 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17421 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17423 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17426 unsigned X86TargetLowering::getExceptionSelectorRegister(
17427 const Constant *PersonalityFn) const {
17428 // Funclet personalities don't use selectors (the runtime does the selection).
17429 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17430 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17433 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17434 SDValue Chain = Op.getOperand(0);
17435 SDValue Offset = Op.getOperand(1);
17436 SDValue Handler = Op.getOperand(2);
17439 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17440 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17441 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17442 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17443 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17444 "Invalid Frame Register!");
17445 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17446 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17448 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17449 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17451 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17452 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17454 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17456 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17457 DAG.getRegister(StoreAddrReg, PtrVT));
17460 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17461 SelectionDAG &DAG) const {
17463 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17464 DAG.getVTList(MVT::i32, MVT::Other),
17465 Op.getOperand(0), Op.getOperand(1));
17468 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17469 SelectionDAG &DAG) const {
17471 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17472 Op.getOperand(0), Op.getOperand(1));
17475 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17476 return Op.getOperand(0);
17479 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17480 SelectionDAG &DAG) const {
17481 SDValue Root = Op.getOperand(0);
17482 SDValue Trmp = Op.getOperand(1); // trampoline
17483 SDValue FPtr = Op.getOperand(2); // nested function
17484 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17487 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17488 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17490 if (Subtarget->is64Bit()) {
17491 SDValue OutChains[6];
17493 // Large code-model.
17494 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17495 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17497 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17498 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17500 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17502 // Load the pointer to the nested function into R11.
17503 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17504 SDValue Addr = Trmp;
17505 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17506 Addr, MachinePointerInfo(TrmpAddr),
17509 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17510 DAG.getConstant(2, dl, MVT::i64));
17511 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17512 MachinePointerInfo(TrmpAddr, 2),
17515 // Load the 'nest' parameter value into R10.
17516 // R10 is specified in X86CallingConv.td
17517 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17518 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17519 DAG.getConstant(10, dl, MVT::i64));
17520 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17521 Addr, MachinePointerInfo(TrmpAddr, 10),
17524 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17525 DAG.getConstant(12, dl, MVT::i64));
17526 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17527 MachinePointerInfo(TrmpAddr, 12),
17530 // Jump to the nested function.
17531 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17532 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17533 DAG.getConstant(20, dl, MVT::i64));
17534 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17535 Addr, MachinePointerInfo(TrmpAddr, 20),
17538 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17539 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17540 DAG.getConstant(22, dl, MVT::i64));
17541 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17542 Addr, MachinePointerInfo(TrmpAddr, 22),
17545 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17547 const Function *Func =
17548 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17549 CallingConv::ID CC = Func->getCallingConv();
17554 llvm_unreachable("Unsupported calling convention");
17555 case CallingConv::C:
17556 case CallingConv::X86_StdCall: {
17557 // Pass 'nest' parameter in ECX.
17558 // Must be kept in sync with X86CallingConv.td
17559 NestReg = X86::ECX;
17561 // Check that ECX wasn't needed by an 'inreg' parameter.
17562 FunctionType *FTy = Func->getFunctionType();
17563 const AttributeSet &Attrs = Func->getAttributes();
17565 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17566 unsigned InRegCount = 0;
17569 for (FunctionType::param_iterator I = FTy->param_begin(),
17570 E = FTy->param_end(); I != E; ++I, ++Idx)
17571 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17572 auto &DL = DAG.getDataLayout();
17573 // FIXME: should only count parameters that are lowered to integers.
17574 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17577 if (InRegCount > 2) {
17578 report_fatal_error("Nest register in use - reduce number of inreg"
17584 case CallingConv::X86_FastCall:
17585 case CallingConv::X86_ThisCall:
17586 case CallingConv::Fast:
17587 // Pass 'nest' parameter in EAX.
17588 // Must be kept in sync with X86CallingConv.td
17589 NestReg = X86::EAX;
17593 SDValue OutChains[4];
17594 SDValue Addr, Disp;
17596 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17597 DAG.getConstant(10, dl, MVT::i32));
17598 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17600 // This is storing the opcode for MOV32ri.
17601 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17602 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17603 OutChains[0] = DAG.getStore(Root, dl,
17604 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17605 Trmp, MachinePointerInfo(TrmpAddr),
17608 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17609 DAG.getConstant(1, dl, MVT::i32));
17610 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17611 MachinePointerInfo(TrmpAddr, 1),
17614 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17615 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17616 DAG.getConstant(5, dl, MVT::i32));
17617 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17618 Addr, MachinePointerInfo(TrmpAddr, 5),
17621 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17622 DAG.getConstant(6, dl, MVT::i32));
17623 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17624 MachinePointerInfo(TrmpAddr, 6),
17627 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17631 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17632 SelectionDAG &DAG) const {
17634 The rounding mode is in bits 11:10 of FPSR, and has the following
17636 00 Round to nearest
17641 FLT_ROUNDS, on the other hand, expects the following:
17648 To perform the conversion, we do:
17649 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17652 MachineFunction &MF = DAG.getMachineFunction();
17653 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17654 unsigned StackAlignment = TFI.getStackAlignment();
17655 MVT VT = Op.getSimpleValueType();
17658 // Save FP Control Word to stack slot
17659 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17660 SDValue StackSlot =
17661 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17663 MachineMemOperand *MMO =
17664 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17665 MachineMemOperand::MOStore, 2, 2);
17667 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17668 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17669 DAG.getVTList(MVT::Other),
17670 Ops, MVT::i16, MMO);
17672 // Load FP Control Word from stack slot
17673 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17674 MachinePointerInfo(), false, false, false, 0);
17676 // Transform as necessary
17678 DAG.getNode(ISD::SRL, DL, MVT::i16,
17679 DAG.getNode(ISD::AND, DL, MVT::i16,
17680 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17681 DAG.getConstant(11, DL, MVT::i8));
17683 DAG.getNode(ISD::SRL, DL, MVT::i16,
17684 DAG.getNode(ISD::AND, DL, MVT::i16,
17685 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17686 DAG.getConstant(9, DL, MVT::i8));
17689 DAG.getNode(ISD::AND, DL, MVT::i16,
17690 DAG.getNode(ISD::ADD, DL, MVT::i16,
17691 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17692 DAG.getConstant(1, DL, MVT::i16)),
17693 DAG.getConstant(3, DL, MVT::i16));
17695 return DAG.getNode((VT.getSizeInBits() < 16 ?
17696 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17699 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17701 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17702 // to 512-bit vector.
17703 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17704 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17705 // split the vector, perform operation on it's Lo a Hi part and
17706 // concatenate the results.
17707 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17709 MVT VT = Op.getSimpleValueType();
17710 MVT EltVT = VT.getVectorElementType();
17711 unsigned NumElems = VT.getVectorNumElements();
17713 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17714 // Extend to 512 bit vector.
17715 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17716 "Unsupported value type for operation");
17718 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17719 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17720 DAG.getUNDEF(NewVT),
17722 DAG.getIntPtrConstant(0, dl));
17723 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17725 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17726 DAG.getIntPtrConstant(0, dl));
17729 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17730 "Unsupported element type");
17732 if (16 < NumElems) {
17733 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17735 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17736 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17738 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17739 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17741 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17744 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17746 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17747 "Unsupported value type for operation");
17749 // Use native supported vector instruction vplzcntd.
17750 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17751 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17752 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17753 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17755 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17758 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17759 SelectionDAG &DAG) {
17760 MVT VT = Op.getSimpleValueType();
17762 unsigned NumBits = VT.getSizeInBits();
17765 if (VT.isVector() && Subtarget->hasAVX512())
17766 return LowerVectorCTLZ_AVX512(Op, DAG);
17768 Op = Op.getOperand(0);
17769 if (VT == MVT::i8) {
17770 // Zero extend to i32 since there is not an i8 bsr.
17772 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17775 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17776 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17777 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17779 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17782 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17783 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17786 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17788 // Finally xor with NumBits-1.
17789 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17790 DAG.getConstant(NumBits - 1, dl, OpVT));
17793 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17797 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17798 SelectionDAG &DAG) {
17799 MVT VT = Op.getSimpleValueType();
17801 unsigned NumBits = VT.getSizeInBits();
17804 if (VT.isVector() && Subtarget->hasAVX512())
17805 return LowerVectorCTLZ_AVX512(Op, DAG);
17807 Op = Op.getOperand(0);
17808 if (VT == MVT::i8) {
17809 // Zero extend to i32 since there is not an i8 bsr.
17811 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17814 // Issue a bsr (scan bits in reverse).
17815 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17816 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17818 // And xor with NumBits-1.
17819 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17820 DAG.getConstant(NumBits - 1, dl, OpVT));
17823 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17827 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17828 MVT VT = Op.getSimpleValueType();
17829 unsigned NumBits = VT.getScalarSizeInBits();
17832 if (VT.isVector()) {
17833 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17835 SDValue N0 = Op.getOperand(0);
17836 SDValue Zero = DAG.getConstant(0, dl, VT);
17838 // lsb(x) = (x & -x)
17839 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17840 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17842 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17843 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17844 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17845 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17846 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17847 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17850 // cttz(x) = ctpop(lsb - 1)
17851 SDValue One = DAG.getConstant(1, dl, VT);
17852 return DAG.getNode(ISD::CTPOP, dl, VT,
17853 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17856 assert(Op.getOpcode() == ISD::CTTZ &&
17857 "Only scalar CTTZ requires custom lowering");
17859 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17860 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17861 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17863 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17866 DAG.getConstant(NumBits, dl, VT),
17867 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17870 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17873 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17874 // ones, and then concatenate the result back.
17875 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17876 MVT VT = Op.getSimpleValueType();
17878 assert(VT.is256BitVector() && VT.isInteger() &&
17879 "Unsupported value type for operation");
17881 unsigned NumElems = VT.getVectorNumElements();
17884 // Extract the LHS vectors
17885 SDValue LHS = Op.getOperand(0);
17886 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17887 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17889 // Extract the RHS vectors
17890 SDValue RHS = Op.getOperand(1);
17891 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17892 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17894 MVT EltVT = VT.getVectorElementType();
17895 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17897 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17898 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17899 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17902 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17903 if (Op.getValueType() == MVT::i1)
17904 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17905 Op.getOperand(0), Op.getOperand(1));
17906 assert(Op.getSimpleValueType().is256BitVector() &&
17907 Op.getSimpleValueType().isInteger() &&
17908 "Only handle AVX 256-bit vector integer operation");
17909 return Lower256IntArith(Op, DAG);
17912 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17913 if (Op.getValueType() == MVT::i1)
17914 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17915 Op.getOperand(0), Op.getOperand(1));
17916 assert(Op.getSimpleValueType().is256BitVector() &&
17917 Op.getSimpleValueType().isInteger() &&
17918 "Only handle AVX 256-bit vector integer operation");
17919 return Lower256IntArith(Op, DAG);
17922 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17923 assert(Op.getSimpleValueType().is256BitVector() &&
17924 Op.getSimpleValueType().isInteger() &&
17925 "Only handle AVX 256-bit vector integer operation");
17926 return Lower256IntArith(Op, DAG);
17929 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17930 SelectionDAG &DAG) {
17932 MVT VT = Op.getSimpleValueType();
17935 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17937 // Decompose 256-bit ops into smaller 128-bit ops.
17938 if (VT.is256BitVector() && !Subtarget->hasInt256())
17939 return Lower256IntArith(Op, DAG);
17941 SDValue A = Op.getOperand(0);
17942 SDValue B = Op.getOperand(1);
17944 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17945 // pairs, multiply and truncate.
17946 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17947 if (Subtarget->hasInt256()) {
17948 if (VT == MVT::v32i8) {
17949 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17950 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17951 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17952 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17953 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17954 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17955 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17956 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17957 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17958 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17961 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17962 return DAG.getNode(
17963 ISD::TRUNCATE, dl, VT,
17964 DAG.getNode(ISD::MUL, dl, ExVT,
17965 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17966 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17969 assert(VT == MVT::v16i8 &&
17970 "Pre-AVX2 support only supports v16i8 multiplication");
17971 MVT ExVT = MVT::v8i16;
17973 // Extract the lo parts and sign extend to i16
17975 if (Subtarget->hasSSE41()) {
17976 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17977 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17979 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17980 -1, 4, -1, 5, -1, 6, -1, 7};
17981 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17982 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17983 ALo = DAG.getBitcast(ExVT, ALo);
17984 BLo = DAG.getBitcast(ExVT, BLo);
17985 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17986 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17989 // Extract the hi parts and sign extend to i16
17991 if (Subtarget->hasSSE41()) {
17992 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17993 -1, -1, -1, -1, -1, -1, -1, -1};
17994 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17995 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17996 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17997 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17999 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
18000 -1, 12, -1, 13, -1, 14, -1, 15};
18001 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18002 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18003 AHi = DAG.getBitcast(ExVT, AHi);
18004 BHi = DAG.getBitcast(ExVT, BHi);
18005 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
18006 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
18009 // Multiply, mask the lower 8bits of the lo/hi results and pack
18010 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
18011 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
18012 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
18013 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
18014 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18017 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18018 if (VT == MVT::v4i32) {
18019 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18020 "Should not custom lower when pmuldq is available!");
18022 // Extract the odd parts.
18023 static const int UnpackMask[] = { 1, -1, 3, -1 };
18024 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18025 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18027 // Multiply the even parts.
18028 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18029 // Now multiply odd parts.
18030 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18032 Evens = DAG.getBitcast(VT, Evens);
18033 Odds = DAG.getBitcast(VT, Odds);
18035 // Merge the two vectors back together with a shuffle. This expands into 2
18037 static const int ShufMask[] = { 0, 4, 2, 6 };
18038 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18041 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18042 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18044 // Ahi = psrlqi(a, 32);
18045 // Bhi = psrlqi(b, 32);
18047 // AloBlo = pmuludq(a, b);
18048 // AloBhi = pmuludq(a, Bhi);
18049 // AhiBlo = pmuludq(Ahi, b);
18051 // AloBhi = psllqi(AloBhi, 32);
18052 // AhiBlo = psllqi(AhiBlo, 32);
18053 // return AloBlo + AloBhi + AhiBlo;
18055 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18056 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18058 SDValue AhiBlo = Ahi;
18059 SDValue AloBhi = Bhi;
18060 // Bit cast to 32-bit vectors for MULUDQ
18061 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18062 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18063 A = DAG.getBitcast(MulVT, A);
18064 B = DAG.getBitcast(MulVT, B);
18065 Ahi = DAG.getBitcast(MulVT, Ahi);
18066 Bhi = DAG.getBitcast(MulVT, Bhi);
18068 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18069 // After shifting right const values the result may be all-zero.
18070 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
18071 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18072 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18074 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
18075 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18076 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18079 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18080 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18083 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18084 assert(Subtarget->isTargetWin64() && "Unexpected target");
18085 EVT VT = Op.getValueType();
18086 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18087 "Unexpected return type for lowering");
18091 switch (Op->getOpcode()) {
18092 default: llvm_unreachable("Unexpected request for libcall!");
18093 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18094 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18095 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18096 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18097 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18098 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18102 SDValue InChain = DAG.getEntryNode();
18104 TargetLowering::ArgListTy Args;
18105 TargetLowering::ArgListEntry Entry;
18106 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18107 EVT ArgVT = Op->getOperand(i).getValueType();
18108 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18109 "Unexpected argument type for lowering");
18110 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18111 Entry.Node = StackPtr;
18112 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18114 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18115 Entry.Ty = PointerType::get(ArgTy,0);
18116 Entry.isSExt = false;
18117 Entry.isZExt = false;
18118 Args.push_back(Entry);
18121 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18122 getPointerTy(DAG.getDataLayout()));
18124 TargetLowering::CallLoweringInfo CLI(DAG);
18125 CLI.setDebugLoc(dl).setChain(InChain)
18126 .setCallee(getLibcallCallingConv(LC),
18127 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18128 Callee, std::move(Args), 0)
18129 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18131 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18132 return DAG.getBitcast(VT, CallInfo.first);
18135 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18136 SelectionDAG &DAG) {
18137 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18138 MVT VT = Op0.getSimpleValueType();
18141 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18142 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18144 // PMULxD operations multiply each even value (starting at 0) of LHS with
18145 // the related value of RHS and produce a widen result.
18146 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18147 // => <2 x i64> <ae|cg>
18149 // In other word, to have all the results, we need to perform two PMULxD:
18150 // 1. one with the even values.
18151 // 2. one with the odd values.
18152 // To achieve #2, with need to place the odd values at an even position.
18154 // Place the odd value at an even position (basically, shift all values 1
18155 // step to the left):
18156 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18157 // <a|b|c|d> => <b|undef|d|undef>
18158 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18159 // <e|f|g|h> => <f|undef|h|undef>
18160 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18162 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18164 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18165 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18167 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18168 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18169 // => <2 x i64> <ae|cg>
18170 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18171 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18172 // => <2 x i64> <bf|dh>
18173 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18175 // Shuffle it back into the right order.
18176 SDValue Highs, Lows;
18177 if (VT == MVT::v8i32) {
18178 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18179 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18180 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18181 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18183 const int HighMask[] = {1, 5, 3, 7};
18184 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18185 const int LowMask[] = {0, 4, 2, 6};
18186 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18189 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18190 // unsigned multiply.
18191 if (IsSigned && !Subtarget->hasSSE41()) {
18192 SDValue ShAmt = DAG.getConstant(
18194 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18195 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18196 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18197 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18198 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18200 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18201 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18204 // The first result of MUL_LOHI is actually the low value, followed by the
18206 SDValue Ops[] = {Lows, Highs};
18207 return DAG.getMergeValues(Ops, dl);
18210 // Return true if the required (according to Opcode) shift-imm form is natively
18211 // supported by the Subtarget
18212 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18214 if (VT.getScalarSizeInBits() < 16)
18217 if (VT.is512BitVector() &&
18218 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18221 bool LShift = VT.is128BitVector() ||
18222 (VT.is256BitVector() && Subtarget->hasInt256());
18224 bool AShift = LShift && (Subtarget->hasVLX() ||
18225 (VT != MVT::v2i64 && VT != MVT::v4i64));
18226 return (Opcode == ISD::SRA) ? AShift : LShift;
18229 // The shift amount is a variable, but it is the same for all vector lanes.
18230 // These instructions are defined together with shift-immediate.
18232 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18234 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18237 // Return true if the required (according to Opcode) variable-shift form is
18238 // natively supported by the Subtarget
18239 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18242 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18245 // vXi16 supported only on AVX-512, BWI
18246 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18249 if (VT.is512BitVector() || Subtarget->hasVLX())
18252 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18253 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18254 return (Opcode == ISD::SRA) ? AShift : LShift;
18257 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18258 const X86Subtarget *Subtarget) {
18259 MVT VT = Op.getSimpleValueType();
18261 SDValue R = Op.getOperand(0);
18262 SDValue Amt = Op.getOperand(1);
18264 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18265 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18267 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18268 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18269 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18270 SDValue Ex = DAG.getBitcast(ExVT, R);
18272 if (ShiftAmt >= 32) {
18273 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18275 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18276 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18277 ShiftAmt - 32, DAG);
18278 if (VT == MVT::v2i64)
18279 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18280 if (VT == MVT::v4i64)
18281 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18282 {9, 1, 11, 3, 13, 5, 15, 7});
18284 // SRA upper i32, SHL whole i64 and select lower i32.
18285 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18288 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18289 Lower = DAG.getBitcast(ExVT, Lower);
18290 if (VT == MVT::v2i64)
18291 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18292 if (VT == MVT::v4i64)
18293 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18294 {8, 1, 10, 3, 12, 5, 14, 7});
18296 return DAG.getBitcast(VT, Ex);
18299 // Optimize shl/srl/sra with constant shift amount.
18300 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18301 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18302 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18304 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18305 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18307 // i64 SRA needs to be performed as partial shifts.
18308 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18309 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18310 return ArithmeticShiftRight64(ShiftAmt);
18312 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18313 unsigned NumElts = VT.getVectorNumElements();
18314 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18316 // Simple i8 add case
18317 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18318 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18320 // ashr(R, 7) === cmp_slt(R, 0)
18321 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18322 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18323 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18326 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18327 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18330 if (Op.getOpcode() == ISD::SHL) {
18331 // Make a large shift.
18332 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18334 SHL = DAG.getBitcast(VT, SHL);
18335 // Zero out the rightmost bits.
18336 SmallVector<SDValue, 32> V(
18337 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18338 return DAG.getNode(ISD::AND, dl, VT, SHL,
18339 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18341 if (Op.getOpcode() == ISD::SRL) {
18342 // Make a large shift.
18343 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18345 SRL = DAG.getBitcast(VT, SRL);
18346 // Zero out the leftmost bits.
18347 SmallVector<SDValue, 32> V(
18348 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18349 return DAG.getNode(ISD::AND, dl, VT, SRL,
18350 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18352 if (Op.getOpcode() == ISD::SRA) {
18353 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18354 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18355 SmallVector<SDValue, 32> V(NumElts,
18356 DAG.getConstant(128 >> ShiftAmt, dl,
18358 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18359 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18360 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18363 llvm_unreachable("Unknown shift opcode.");
18368 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18369 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18370 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18372 // Peek through any splat that was introduced for i64 shift vectorization.
18373 int SplatIndex = -1;
18374 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18375 if (SVN->isSplat()) {
18376 SplatIndex = SVN->getSplatIndex();
18377 Amt = Amt.getOperand(0);
18378 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18379 "Splat shuffle referencing second operand");
18382 if (Amt.getOpcode() != ISD::BITCAST ||
18383 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18386 Amt = Amt.getOperand(0);
18387 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18388 VT.getVectorNumElements();
18389 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18390 uint64_t ShiftAmt = 0;
18391 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18392 for (unsigned i = 0; i != Ratio; ++i) {
18393 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18397 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18400 // Check remaining shift amounts (if not a splat).
18401 if (SplatIndex < 0) {
18402 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18403 uint64_t ShAmt = 0;
18404 for (unsigned j = 0; j != Ratio; ++j) {
18405 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18409 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18411 if (ShAmt != ShiftAmt)
18416 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18417 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18419 if (Op.getOpcode() == ISD::SRA)
18420 return ArithmeticShiftRight64(ShiftAmt);
18426 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18427 const X86Subtarget* Subtarget) {
18428 MVT VT = Op.getSimpleValueType();
18430 SDValue R = Op.getOperand(0);
18431 SDValue Amt = Op.getOperand(1);
18433 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18434 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18436 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18437 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18439 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18441 MVT EltVT = VT.getVectorElementType();
18443 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18444 // Check if this build_vector node is doing a splat.
18445 // If so, then set BaseShAmt equal to the splat value.
18446 BaseShAmt = BV->getSplatValue();
18447 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18448 BaseShAmt = SDValue();
18450 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18451 Amt = Amt.getOperand(0);
18453 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18454 if (SVN && SVN->isSplat()) {
18455 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18456 SDValue InVec = Amt.getOperand(0);
18457 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18458 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18459 "Unexpected shuffle index found!");
18460 BaseShAmt = InVec.getOperand(SplatIdx);
18461 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18462 if (ConstantSDNode *C =
18463 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18464 if (C->getZExtValue() == SplatIdx)
18465 BaseShAmt = InVec.getOperand(1);
18470 // Avoid introducing an extract element from a shuffle.
18471 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18472 DAG.getIntPtrConstant(SplatIdx, dl));
18476 if (BaseShAmt.getNode()) {
18477 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18478 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18479 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18480 else if (EltVT.bitsLT(MVT::i32))
18481 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18483 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18487 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18488 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18489 Amt.getOpcode() == ISD::BITCAST &&
18490 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18491 Amt = Amt.getOperand(0);
18492 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18493 VT.getVectorNumElements();
18494 std::vector<SDValue> Vals(Ratio);
18495 for (unsigned i = 0; i != Ratio; ++i)
18496 Vals[i] = Amt.getOperand(i);
18497 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18498 for (unsigned j = 0; j != Ratio; ++j)
18499 if (Vals[j] != Amt.getOperand(i + j))
18503 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18504 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18509 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18510 SelectionDAG &DAG) {
18511 MVT VT = Op.getSimpleValueType();
18513 SDValue R = Op.getOperand(0);
18514 SDValue Amt = Op.getOperand(1);
18516 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18517 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18519 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18522 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18525 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18528 // XOP has 128-bit variable logical/arithmetic shifts.
18529 // +ve/-ve Amt = shift left/right.
18530 if (Subtarget->hasXOP() &&
18531 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18532 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18533 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18534 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18535 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18537 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18538 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18539 if (Op.getOpcode() == ISD::SRA)
18540 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18543 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18544 // shifts per-lane and then shuffle the partial results back together.
18545 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18546 // Splat the shift amounts so the scalar shifts above will catch it.
18547 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18548 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18549 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18550 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18551 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18554 // i64 vector arithmetic shift can be emulated with the transform:
18555 // M = lshr(SIGN_BIT, Amt)
18556 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18557 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18558 Op.getOpcode() == ISD::SRA) {
18559 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18560 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18561 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18562 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18563 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18567 // If possible, lower this packed shift into a vector multiply instead of
18568 // expanding it into a sequence of scalar shifts.
18569 // Do this only if the vector shift count is a constant build_vector.
18570 if (Op.getOpcode() == ISD::SHL &&
18571 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18572 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18573 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18574 SmallVector<SDValue, 8> Elts;
18575 MVT SVT = VT.getVectorElementType();
18576 unsigned SVTBits = SVT.getSizeInBits();
18577 APInt One(SVTBits, 1);
18578 unsigned NumElems = VT.getVectorNumElements();
18580 for (unsigned i=0; i !=NumElems; ++i) {
18581 SDValue Op = Amt->getOperand(i);
18582 if (Op->getOpcode() == ISD::UNDEF) {
18583 Elts.push_back(Op);
18587 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18588 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18589 uint64_t ShAmt = C.getZExtValue();
18590 if (ShAmt >= SVTBits) {
18591 Elts.push_back(DAG.getUNDEF(SVT));
18594 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18596 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18597 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18600 // Lower SHL with variable shift amount.
18601 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18602 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18604 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18605 DAG.getConstant(0x3f800000U, dl, VT));
18606 Op = DAG.getBitcast(MVT::v4f32, Op);
18607 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18608 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18611 // If possible, lower this shift as a sequence of two shifts by
18612 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18614 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18616 // Could be rewritten as:
18617 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18619 // The advantage is that the two shifts from the example would be
18620 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18621 // the vector shift into four scalar shifts plus four pairs of vector
18623 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18624 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18625 unsigned TargetOpcode = X86ISD::MOVSS;
18626 bool CanBeSimplified;
18627 // The splat value for the first packed shift (the 'X' from the example).
18628 SDValue Amt1 = Amt->getOperand(0);
18629 // The splat value for the second packed shift (the 'Y' from the example).
18630 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18631 Amt->getOperand(2);
18633 // See if it is possible to replace this node with a sequence of
18634 // two shifts followed by a MOVSS/MOVSD
18635 if (VT == MVT::v4i32) {
18636 // Check if it is legal to use a MOVSS.
18637 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18638 Amt2 == Amt->getOperand(3);
18639 if (!CanBeSimplified) {
18640 // Otherwise, check if we can still simplify this node using a MOVSD.
18641 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18642 Amt->getOperand(2) == Amt->getOperand(3);
18643 TargetOpcode = X86ISD::MOVSD;
18644 Amt2 = Amt->getOperand(2);
18647 // Do similar checks for the case where the machine value type
18649 CanBeSimplified = Amt1 == Amt->getOperand(1);
18650 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18651 CanBeSimplified = Amt2 == Amt->getOperand(i);
18653 if (!CanBeSimplified) {
18654 TargetOpcode = X86ISD::MOVSD;
18655 CanBeSimplified = true;
18656 Amt2 = Amt->getOperand(4);
18657 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18658 CanBeSimplified = Amt1 == Amt->getOperand(i);
18659 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18660 CanBeSimplified = Amt2 == Amt->getOperand(j);
18664 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18665 isa<ConstantSDNode>(Amt2)) {
18666 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18667 MVT CastVT = MVT::v4i32;
18669 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18670 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18672 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18673 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18674 if (TargetOpcode == X86ISD::MOVSD)
18675 CastVT = MVT::v2i64;
18676 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18677 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18678 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18680 return DAG.getBitcast(VT, Result);
18684 // v4i32 Non Uniform Shifts.
18685 // If the shift amount is constant we can shift each lane using the SSE2
18686 // immediate shifts, else we need to zero-extend each lane to the lower i64
18687 // and shift using the SSE2 variable shifts.
18688 // The separate results can then be blended together.
18689 if (VT == MVT::v4i32) {
18690 unsigned Opc = Op.getOpcode();
18691 SDValue Amt0, Amt1, Amt2, Amt3;
18692 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18693 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18694 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18695 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18696 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18698 // ISD::SHL is handled above but we include it here for completeness.
18701 llvm_unreachable("Unknown target vector shift node");
18703 Opc = X86ISD::VSHL;
18706 Opc = X86ISD::VSRL;
18709 Opc = X86ISD::VSRA;
18712 // The SSE2 shifts use the lower i64 as the same shift amount for
18713 // all lanes and the upper i64 is ignored. These shuffle masks
18714 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18715 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18716 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18717 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18718 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18719 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18722 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18723 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18724 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18725 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18726 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18727 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18728 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18731 if (VT == MVT::v16i8 ||
18732 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18733 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18734 unsigned ShiftOpcode = Op->getOpcode();
18736 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18737 // On SSE41 targets we make use of the fact that VSELECT lowers
18738 // to PBLENDVB which selects bytes based just on the sign bit.
18739 if (Subtarget->hasSSE41()) {
18740 V0 = DAG.getBitcast(VT, V0);
18741 V1 = DAG.getBitcast(VT, V1);
18742 Sel = DAG.getBitcast(VT, Sel);
18743 return DAG.getBitcast(SelVT,
18744 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18746 // On pre-SSE41 targets we test for the sign bit by comparing to
18747 // zero - a negative value will set all bits of the lanes to true
18748 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18749 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18750 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18751 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18754 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18755 // We can safely do this using i16 shifts as we're only interested in
18756 // the 3 lower bits of each byte.
18757 Amt = DAG.getBitcast(ExtVT, Amt);
18758 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18759 Amt = DAG.getBitcast(VT, Amt);
18761 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18762 // r = VSELECT(r, shift(r, 4), a);
18764 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18765 R = SignBitSelect(VT, Amt, M, R);
18768 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18770 // r = VSELECT(r, shift(r, 2), a);
18771 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18772 R = SignBitSelect(VT, Amt, M, R);
18775 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18777 // return VSELECT(r, shift(r, 1), a);
18778 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18779 R = SignBitSelect(VT, Amt, M, R);
18783 if (Op->getOpcode() == ISD::SRA) {
18784 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18785 // so we can correctly sign extend. We don't care what happens to the
18787 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18788 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18789 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18790 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18791 ALo = DAG.getBitcast(ExtVT, ALo);
18792 AHi = DAG.getBitcast(ExtVT, AHi);
18793 RLo = DAG.getBitcast(ExtVT, RLo);
18794 RHi = DAG.getBitcast(ExtVT, RHi);
18796 // r = VSELECT(r, shift(r, 4), a);
18797 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18798 DAG.getConstant(4, dl, ExtVT));
18799 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18800 DAG.getConstant(4, dl, ExtVT));
18801 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18802 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18805 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18806 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18808 // r = VSELECT(r, shift(r, 2), a);
18809 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18810 DAG.getConstant(2, dl, ExtVT));
18811 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18812 DAG.getConstant(2, dl, ExtVT));
18813 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18814 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18817 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18818 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18820 // r = VSELECT(r, shift(r, 1), a);
18821 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18822 DAG.getConstant(1, dl, ExtVT));
18823 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18824 DAG.getConstant(1, dl, ExtVT));
18825 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18826 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18828 // Logical shift the result back to the lower byte, leaving a zero upper
18830 // meaning that we can safely pack with PACKUSWB.
18832 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18834 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18835 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18839 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18840 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18841 // solution better.
18842 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18843 MVT ExtVT = MVT::v8i32;
18845 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18846 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18847 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18848 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18849 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18852 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18853 MVT ExtVT = MVT::v8i32;
18854 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18855 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18856 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18857 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18858 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18859 ALo = DAG.getBitcast(ExtVT, ALo);
18860 AHi = DAG.getBitcast(ExtVT, AHi);
18861 RLo = DAG.getBitcast(ExtVT, RLo);
18862 RHi = DAG.getBitcast(ExtVT, RHi);
18863 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18864 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18865 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18866 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18867 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18870 if (VT == MVT::v8i16) {
18871 unsigned ShiftOpcode = Op->getOpcode();
18873 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18874 // On SSE41 targets we make use of the fact that VSELECT lowers
18875 // to PBLENDVB which selects bytes based just on the sign bit.
18876 if (Subtarget->hasSSE41()) {
18877 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18878 V0 = DAG.getBitcast(ExtVT, V0);
18879 V1 = DAG.getBitcast(ExtVT, V1);
18880 Sel = DAG.getBitcast(ExtVT, Sel);
18881 return DAG.getBitcast(
18882 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18884 // On pre-SSE41 targets we splat the sign bit - a negative value will
18885 // set all bits of the lanes to true and VSELECT uses that in
18886 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18888 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18889 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18892 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18893 if (Subtarget->hasSSE41()) {
18894 // On SSE41 targets we need to replicate the shift mask in both
18895 // bytes for PBLENDVB.
18898 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18899 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18901 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18904 // r = VSELECT(r, shift(r, 8), a);
18905 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18906 R = SignBitSelect(Amt, M, R);
18909 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18911 // r = VSELECT(r, shift(r, 4), a);
18912 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18913 R = SignBitSelect(Amt, M, R);
18916 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18918 // r = VSELECT(r, shift(r, 2), a);
18919 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18920 R = SignBitSelect(Amt, M, R);
18923 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18925 // return VSELECT(r, shift(r, 1), a);
18926 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18927 R = SignBitSelect(Amt, M, R);
18931 // Decompose 256-bit shifts into smaller 128-bit shifts.
18932 if (VT.is256BitVector()) {
18933 unsigned NumElems = VT.getVectorNumElements();
18934 MVT EltVT = VT.getVectorElementType();
18935 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18937 // Extract the two vectors
18938 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18939 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18941 // Recreate the shift amount vectors
18942 SDValue Amt1, Amt2;
18943 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18944 // Constant shift amount
18945 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18946 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18947 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18949 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18950 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18952 // Variable shift amount
18953 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18954 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18957 // Issue new vector shifts for the smaller types
18958 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18959 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18961 // Concatenate the result back
18962 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18968 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18969 SelectionDAG &DAG) {
18970 MVT VT = Op.getSimpleValueType();
18972 SDValue R = Op.getOperand(0);
18973 SDValue Amt = Op.getOperand(1);
18975 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18976 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18977 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18979 // XOP has 128-bit vector variable + immediate rotates.
18980 // +ve/-ve Amt = rotate left/right.
18982 // Split 256-bit integers.
18983 if (VT.is256BitVector())
18984 return Lower256IntArith(Op, DAG);
18986 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
18988 // Attempt to rotate by immediate.
18989 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18990 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18991 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18992 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18993 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18994 DAG.getConstant(RotateAmt, DL, MVT::i8));
18998 // Use general rotate by variable (per-element).
18999 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
19002 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19003 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19004 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19005 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19006 // has only one use.
19007 SDNode *N = Op.getNode();
19008 SDValue LHS = N->getOperand(0);
19009 SDValue RHS = N->getOperand(1);
19010 unsigned BaseOp = 0;
19013 switch (Op.getOpcode()) {
19014 default: llvm_unreachable("Unknown ovf instruction!");
19016 // A subtract of one will be selected as a INC. Note that INC doesn't
19017 // set CF, so we can't do this for UADDO.
19018 if (isOneConstant(RHS)) {
19019 BaseOp = X86ISD::INC;
19020 Cond = X86::COND_O;
19023 BaseOp = X86ISD::ADD;
19024 Cond = X86::COND_O;
19027 BaseOp = X86ISD::ADD;
19028 Cond = X86::COND_B;
19031 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19032 // set CF, so we can't do this for USUBO.
19033 if (isOneConstant(RHS)) {
19034 BaseOp = X86ISD::DEC;
19035 Cond = X86::COND_O;
19038 BaseOp = X86ISD::SUB;
19039 Cond = X86::COND_O;
19042 BaseOp = X86ISD::SUB;
19043 Cond = X86::COND_B;
19046 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19047 Cond = X86::COND_O;
19049 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19050 if (N->getValueType(0) == MVT::i8) {
19051 BaseOp = X86ISD::UMUL8;
19052 Cond = X86::COND_O;
19055 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19057 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19060 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19061 DAG.getConstant(X86::COND_O, DL, MVT::i32),
19062 SDValue(Sum.getNode(), 2));
19064 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19068 // Also sets EFLAGS.
19069 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19070 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19073 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19074 DAG.getConstant(Cond, DL, MVT::i32),
19075 SDValue(Sum.getNode(), 1));
19077 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19080 /// Returns true if the operand type is exactly twice the native width, and
19081 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19082 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19083 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19084 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
19085 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19088 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19089 else if (OpWidth == 128)
19090 return Subtarget->hasCmpxchg16b();
19095 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19096 return needsCmpXchgNb(SI->getValueOperand()->getType());
19099 // Note: this turns large loads into lock cmpxchg8b/16b.
19100 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19101 TargetLowering::AtomicExpansionKind
19102 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19103 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19104 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
19105 : AtomicExpansionKind::None;
19108 TargetLowering::AtomicExpansionKind
19109 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19110 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19111 Type *MemType = AI->getType();
19113 // If the operand is too big, we must see if cmpxchg8/16b is available
19114 // and default to library calls otherwise.
19115 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
19116 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19117 : AtomicExpansionKind::None;
19120 AtomicRMWInst::BinOp Op = AI->getOperation();
19123 llvm_unreachable("Unknown atomic operation");
19124 case AtomicRMWInst::Xchg:
19125 case AtomicRMWInst::Add:
19126 case AtomicRMWInst::Sub:
19127 // It's better to use xadd, xsub or xchg for these in all cases.
19128 return AtomicExpansionKind::None;
19129 case AtomicRMWInst::Or:
19130 case AtomicRMWInst::And:
19131 case AtomicRMWInst::Xor:
19132 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19133 // prefix to a normal instruction for these operations.
19134 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19135 : AtomicExpansionKind::None;
19136 case AtomicRMWInst::Nand:
19137 case AtomicRMWInst::Max:
19138 case AtomicRMWInst::Min:
19139 case AtomicRMWInst::UMax:
19140 case AtomicRMWInst::UMin:
19141 // These always require a non-trivial set of data operations on x86. We must
19142 // use a cmpxchg loop.
19143 return AtomicExpansionKind::CmpXChg;
19147 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19148 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19149 // no-sse2). There isn't any reason to disable it if the target processor
19151 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19155 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19156 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19157 Type *MemType = AI->getType();
19158 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19159 // there is no benefit in turning such RMWs into loads, and it is actually
19160 // harmful as it introduces a mfence.
19161 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19164 auto Builder = IRBuilder<>(AI);
19165 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19166 auto SynchScope = AI->getSynchScope();
19167 // We must restrict the ordering to avoid generating loads with Release or
19168 // ReleaseAcquire orderings.
19169 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19170 auto Ptr = AI->getPointerOperand();
19172 // Before the load we need a fence. Here is an example lifted from
19173 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19176 // x.store(1, relaxed);
19177 // r1 = y.fetch_add(0, release);
19179 // y.fetch_add(42, acquire);
19180 // r2 = x.load(relaxed);
19181 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19182 // lowered to just a load without a fence. A mfence flushes the store buffer,
19183 // making the optimization clearly correct.
19184 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19185 // otherwise, we might be able to be more aggressive on relaxed idempotent
19186 // rmw. In practice, they do not look useful, so we don't try to be
19187 // especially clever.
19188 if (SynchScope == SingleThread)
19189 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19190 // the IR level, so we must wrap it in an intrinsic.
19193 if (!hasMFENCE(*Subtarget))
19194 // FIXME: it might make sense to use a locked operation here but on a
19195 // different cache-line to prevent cache-line bouncing. In practice it
19196 // is probably a small win, and x86 processors without mfence are rare
19197 // enough that we do not bother.
19201 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19202 Builder.CreateCall(MFence, {});
19204 // Finally we can emit the atomic load.
19205 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19206 AI->getType()->getPrimitiveSizeInBits());
19207 Loaded->setAtomic(Order, SynchScope);
19208 AI->replaceAllUsesWith(Loaded);
19209 AI->eraseFromParent();
19213 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19214 SelectionDAG &DAG) {
19216 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19217 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19218 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19219 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19221 // The only fence that needs an instruction is a sequentially-consistent
19222 // cross-thread fence.
19223 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19224 if (hasMFENCE(*Subtarget))
19225 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19227 SDValue Chain = Op.getOperand(0);
19228 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19230 DAG.getRegister(X86::ESP, MVT::i32), // Base
19231 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19232 DAG.getRegister(0, MVT::i32), // Index
19233 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19234 DAG.getRegister(0, MVT::i32), // Segment.
19238 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19239 return SDValue(Res, 0);
19242 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19243 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19246 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19247 SelectionDAG &DAG) {
19248 MVT T = Op.getSimpleValueType();
19252 switch(T.SimpleTy) {
19253 default: llvm_unreachable("Invalid value type!");
19254 case MVT::i8: Reg = X86::AL; size = 1; break;
19255 case MVT::i16: Reg = X86::AX; size = 2; break;
19256 case MVT::i32: Reg = X86::EAX; size = 4; break;
19258 assert(Subtarget->is64Bit() && "Node not type legal!");
19259 Reg = X86::RAX; size = 8;
19262 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19263 Op.getOperand(2), SDValue());
19264 SDValue Ops[] = { cpIn.getValue(0),
19267 DAG.getTargetConstant(size, DL, MVT::i8),
19268 cpIn.getValue(1) };
19269 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19270 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19271 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19275 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19276 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19277 MVT::i32, cpOut.getValue(2));
19278 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19279 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19282 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19283 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19284 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19288 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19289 SelectionDAG &DAG) {
19290 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19291 MVT DstVT = Op.getSimpleValueType();
19293 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19294 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19295 if (DstVT != MVT::f64)
19296 // This conversion needs to be expanded.
19299 SDValue InVec = Op->getOperand(0);
19301 unsigned NumElts = SrcVT.getVectorNumElements();
19302 MVT SVT = SrcVT.getVectorElementType();
19304 // Widen the vector in input in the case of MVT::v2i32.
19305 // Example: from MVT::v2i32 to MVT::v4i32.
19306 SmallVector<SDValue, 16> Elts;
19307 for (unsigned i = 0, e = NumElts; i != e; ++i)
19308 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19309 DAG.getIntPtrConstant(i, dl)));
19311 // Explicitly mark the extra elements as Undef.
19312 Elts.append(NumElts, DAG.getUNDEF(SVT));
19314 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19315 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19316 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19317 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19318 DAG.getIntPtrConstant(0, dl));
19321 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19322 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19323 assert((DstVT == MVT::i64 ||
19324 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19325 "Unexpected custom BITCAST");
19326 // i64 <=> MMX conversions are Legal.
19327 if (SrcVT==MVT::i64 && DstVT.isVector())
19329 if (DstVT==MVT::i64 && SrcVT.isVector())
19331 // MMX <=> MMX conversions are Legal.
19332 if (SrcVT.isVector() && DstVT.isVector())
19334 // All other conversions need to be expanded.
19338 /// Compute the horizontal sum of bytes in V for the elements of VT.
19340 /// Requires V to be a byte vector and VT to be an integer vector type with
19341 /// wider elements than V's type. The width of the elements of VT determines
19342 /// how many bytes of V are summed horizontally to produce each element of the
19344 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19345 const X86Subtarget *Subtarget,
19346 SelectionDAG &DAG) {
19348 MVT ByteVecVT = V.getSimpleValueType();
19349 MVT EltVT = VT.getVectorElementType();
19350 int NumElts = VT.getVectorNumElements();
19351 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19352 "Expected value to have byte element type.");
19353 assert(EltVT != MVT::i8 &&
19354 "Horizontal byte sum only makes sense for wider elements!");
19355 unsigned VecSize = VT.getSizeInBits();
19356 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19358 // PSADBW instruction horizontally add all bytes and leave the result in i64
19359 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19360 if (EltVT == MVT::i64) {
19361 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19362 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19363 V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
19364 return DAG.getBitcast(VT, V);
19367 if (EltVT == MVT::i32) {
19368 // We unpack the low half and high half into i32s interleaved with zeros so
19369 // that we can use PSADBW to horizontally sum them. The most useful part of
19370 // this is that it lines up the results of two PSADBW instructions to be
19371 // two v2i64 vectors which concatenated are the 4 population counts. We can
19372 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19373 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19374 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19375 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19377 // Do the horizontal sums into two v2i64s.
19378 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19379 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19380 Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19381 DAG.getBitcast(ByteVecVT, Low), Zeros);
19382 High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19383 DAG.getBitcast(ByteVecVT, High), Zeros);
19385 // Merge them together.
19386 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19387 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19388 DAG.getBitcast(ShortVecVT, Low),
19389 DAG.getBitcast(ShortVecVT, High));
19391 return DAG.getBitcast(VT, V);
19394 // The only element type left is i16.
19395 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19397 // To obtain pop count for each i16 element starting from the pop count for
19398 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19399 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19400 // directly supported.
19401 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19402 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19403 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19404 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19405 DAG.getBitcast(ByteVecVT, V));
19406 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19409 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19410 const X86Subtarget *Subtarget,
19411 SelectionDAG &DAG) {
19412 MVT VT = Op.getSimpleValueType();
19413 MVT EltVT = VT.getVectorElementType();
19414 unsigned VecSize = VT.getSizeInBits();
19416 // Implement a lookup table in register by using an algorithm based on:
19417 // http://wm.ite.pl/articles/sse-popcount.html
19419 // The general idea is that every lower byte nibble in the input vector is an
19420 // index into a in-register pre-computed pop count table. We then split up the
19421 // input vector in two new ones: (1) a vector with only the shifted-right
19422 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19423 // masked out higher ones) for each byte. PSHUB is used separately with both
19424 // to index the in-register table. Next, both are added and the result is a
19425 // i8 vector where each element contains the pop count for input byte.
19427 // To obtain the pop count for elements != i8, we follow up with the same
19428 // approach and use additional tricks as described below.
19430 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19431 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19432 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19433 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19435 int NumByteElts = VecSize / 8;
19436 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19437 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19438 SmallVector<SDValue, 16> LUTVec;
19439 for (int i = 0; i < NumByteElts; ++i)
19440 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19441 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19442 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19443 DAG.getConstant(0x0F, DL, MVT::i8));
19444 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19447 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19448 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19449 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19452 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19454 // The input vector is used as the shuffle mask that index elements into the
19455 // LUT. After counting low and high nibbles, add the vector to obtain the
19456 // final pop count per i8 element.
19457 SDValue HighPopCnt =
19458 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19459 SDValue LowPopCnt =
19460 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19461 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19463 if (EltVT == MVT::i8)
19466 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19469 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19470 const X86Subtarget *Subtarget,
19471 SelectionDAG &DAG) {
19472 MVT VT = Op.getSimpleValueType();
19473 assert(VT.is128BitVector() &&
19474 "Only 128-bit vector bitmath lowering supported.");
19476 int VecSize = VT.getSizeInBits();
19477 MVT EltVT = VT.getVectorElementType();
19478 int Len = EltVT.getSizeInBits();
19480 // This is the vectorized version of the "best" algorithm from
19481 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19482 // with a minor tweak to use a series of adds + shifts instead of vector
19483 // multiplications. Implemented for all integer vector types. We only use
19484 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19485 // much faster, even faster than using native popcnt instructions.
19487 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19488 MVT VT = V.getSimpleValueType();
19489 SmallVector<SDValue, 32> Shifters(
19490 VT.getVectorNumElements(),
19491 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19492 return DAG.getNode(OpCode, DL, VT, V,
19493 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19495 auto GetMask = [&](SDValue V, APInt Mask) {
19496 MVT VT = V.getSimpleValueType();
19497 SmallVector<SDValue, 32> Masks(
19498 VT.getVectorNumElements(),
19499 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19500 return DAG.getNode(ISD::AND, DL, VT, V,
19501 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19504 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19505 // x86, so set the SRL type to have elements at least i16 wide. This is
19506 // correct because all of our SRLs are followed immediately by a mask anyways
19507 // that handles any bits that sneak into the high bits of the byte elements.
19508 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19512 // v = v - ((v >> 1) & 0x55555555...)
19514 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19515 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19516 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19518 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19519 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19520 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19521 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19522 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19524 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19525 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19526 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19527 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19529 // At this point, V contains the byte-wise population count, and we are
19530 // merely doing a horizontal sum if necessary to get the wider element
19532 if (EltVT == MVT::i8)
19535 return LowerHorizontalByteSum(
19536 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19540 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19541 SelectionDAG &DAG) {
19542 MVT VT = Op.getSimpleValueType();
19543 // FIXME: Need to add AVX-512 support here!
19544 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19545 "Unknown CTPOP type to handle");
19546 SDLoc DL(Op.getNode());
19547 SDValue Op0 = Op.getOperand(0);
19549 if (!Subtarget->hasSSSE3()) {
19550 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19551 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19552 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19555 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19556 unsigned NumElems = VT.getVectorNumElements();
19558 // Extract each 128-bit vector, compute pop count and concat the result.
19559 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19560 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19562 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19563 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19564 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19567 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19570 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19571 SelectionDAG &DAG) {
19572 assert(Op.getSimpleValueType().isVector() &&
19573 "We only do custom lowering for vector population count.");
19574 return LowerVectorCTPOP(Op, Subtarget, DAG);
19577 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19578 SDNode *Node = Op.getNode();
19580 EVT T = Node->getValueType(0);
19581 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19582 DAG.getConstant(0, dl, T), Node->getOperand(2));
19583 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19584 cast<AtomicSDNode>(Node)->getMemoryVT(),
19585 Node->getOperand(0),
19586 Node->getOperand(1), negOp,
19587 cast<AtomicSDNode>(Node)->getMemOperand(),
19588 cast<AtomicSDNode>(Node)->getOrdering(),
19589 cast<AtomicSDNode>(Node)->getSynchScope());
19592 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19593 SDNode *Node = Op.getNode();
19595 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19597 // Convert seq_cst store -> xchg
19598 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19599 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19600 // (The only way to get a 16-byte store is cmpxchg16b)
19601 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19602 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19603 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19604 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19605 cast<AtomicSDNode>(Node)->getMemoryVT(),
19606 Node->getOperand(0),
19607 Node->getOperand(1), Node->getOperand(2),
19608 cast<AtomicSDNode>(Node)->getMemOperand(),
19609 cast<AtomicSDNode>(Node)->getOrdering(),
19610 cast<AtomicSDNode>(Node)->getSynchScope());
19611 return Swap.getValue(1);
19613 // Other atomic stores have a simple pattern.
19617 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19618 MVT VT = Op.getNode()->getSimpleValueType(0);
19620 // Let legalize expand this if it isn't a legal type yet.
19621 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19624 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19627 bool ExtraOp = false;
19628 switch (Op.getOpcode()) {
19629 default: llvm_unreachable("Invalid code");
19630 case ISD::ADDC: Opc = X86ISD::ADD; break;
19631 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19632 case ISD::SUBC: Opc = X86ISD::SUB; break;
19633 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19637 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19639 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19640 Op.getOperand(1), Op.getOperand(2));
19643 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19644 SelectionDAG &DAG) {
19645 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19647 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19648 // which returns the values as { float, float } (in XMM0) or
19649 // { double, double } (which is returned in XMM0, XMM1).
19651 SDValue Arg = Op.getOperand(0);
19652 EVT ArgVT = Arg.getValueType();
19653 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19655 TargetLowering::ArgListTy Args;
19656 TargetLowering::ArgListEntry Entry;
19660 Entry.isSExt = false;
19661 Entry.isZExt = false;
19662 Args.push_back(Entry);
19664 bool isF64 = ArgVT == MVT::f64;
19665 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19666 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19667 // the results are returned via SRet in memory.
19668 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19669 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19671 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19673 Type *RetTy = isF64
19674 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19675 : (Type*)VectorType::get(ArgTy, 4);
19677 TargetLowering::CallLoweringInfo CLI(DAG);
19678 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19679 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19681 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19684 // Returned in xmm0 and xmm1.
19685 return CallResult.first;
19687 // Returned in bits 0:31 and 32:64 xmm0.
19688 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19689 CallResult.first, DAG.getIntPtrConstant(0, dl));
19690 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19691 CallResult.first, DAG.getIntPtrConstant(1, dl));
19692 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19693 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19696 /// Widen a vector input to a vector of NVT. The
19697 /// input vector must have the same element type as NVT.
19698 static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG,
19699 bool FillWithZeroes = false) {
19700 // Check if InOp already has the right width.
19701 MVT InVT = InOp.getSimpleValueType();
19705 if (InOp.isUndef())
19706 return DAG.getUNDEF(NVT);
19708 assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
19709 "input and widen element type must match");
19711 unsigned InNumElts = InVT.getVectorNumElements();
19712 unsigned WidenNumElts = NVT.getVectorNumElements();
19713 assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&
19714 "Unexpected request for vector widening");
19716 EVT EltVT = NVT.getVectorElementType();
19719 if (InOp.getOpcode() == ISD::CONCAT_VECTORS &&
19720 InOp.getNumOperands() == 2) {
19721 SDValue N1 = InOp.getOperand(1);
19722 if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) ||
19724 InOp = InOp.getOperand(0);
19725 InVT = InOp.getSimpleValueType();
19726 InNumElts = InVT.getVectorNumElements();
19729 if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) ||
19730 ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) {
19731 SmallVector<SDValue, 16> Ops;
19732 for (unsigned i = 0; i < InNumElts; ++i)
19733 Ops.push_back(InOp.getOperand(i));
19735 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) :
19736 DAG.getUNDEF(EltVT);
19737 for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i)
19738 Ops.push_back(FillVal);
19739 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
19741 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) :
19743 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal,
19744 InOp, DAG.getIntPtrConstant(0, dl));
19747 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19748 SelectionDAG &DAG) {
19749 assert(Subtarget->hasAVX512() &&
19750 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19752 // X86 scatter kills mask register, so its type should be added to
19753 // the list of return values.
19754 // If the "scatter" has 2 return values, it is already handled.
19755 if (Op.getNode()->getNumValues() == 2)
19758 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19759 SDValue Src = N->getValue();
19760 MVT VT = Src.getSimpleValueType();
19761 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19764 SDValue NewScatter;
19765 SDValue Index = N->getIndex();
19766 SDValue Mask = N->getMask();
19767 SDValue Chain = N->getChain();
19768 SDValue BasePtr = N->getBasePtr();
19769 MVT MemVT = N->getMemoryVT().getSimpleVT();
19770 MVT IndexVT = Index.getSimpleValueType();
19771 MVT MaskVT = Mask.getSimpleValueType();
19773 if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) {
19774 // The v2i32 value was promoted to v2i64.
19775 // Now we "redo" the type legalizer's work and widen the original
19776 // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64
19778 assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&
19779 "Unexpected memory type");
19780 int ShuffleMask[] = {0, 2, -1, -1};
19781 Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src),
19782 DAG.getUNDEF(MVT::v4i32), ShuffleMask);
19783 // Now we have 4 elements instead of 2.
19784 // Expand the index.
19785 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4);
19786 Index = ExtendToType(Index, NewIndexVT, DAG);
19788 // Expand the mask with zeroes
19789 // Mask may be <2 x i64> or <2 x i1> at this moment
19790 assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&
19791 "Unexpected mask type");
19792 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4);
19793 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
19797 unsigned NumElts = VT.getVectorNumElements();
19798 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19799 !Index.getSimpleValueType().is512BitVector()) {
19800 // AVX512F supports only 512-bit vectors. Or data or index should
19801 // be 512 bit wide. If now the both index and data are 256-bit, but
19802 // the vector contains 8 elements, we just sign-extend the index
19803 if (IndexVT == MVT::v8i32)
19804 // Just extend index
19805 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19807 // The minimal number of elts in scatter is 8
19810 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
19811 // Use original index here, do not modify the index twice
19812 Index = ExtendToType(N->getIndex(), NewIndexVT, DAG);
19813 if (IndexVT.getScalarType() == MVT::i32)
19814 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19817 // At this point we have promoted mask operand
19818 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
19819 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
19820 // Use the original mask here, do not modify the mask twice
19821 Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true);
19823 // The value that should be stored
19824 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
19825 Src = ExtendToType(Src, NewVT, DAG);
19828 // If the mask is "wide" at this point - truncate it to i1 vector
19829 MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts);
19830 Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask);
19832 // The mask is killed by scatter, add it to the values
19833 SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other);
19834 SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
19835 NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops,
19836 N->getMemOperand());
19837 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19838 return SDValue(NewScatter.getNode(), 0);
19841 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget *Subtarget,
19842 SelectionDAG &DAG) {
19844 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
19845 MVT VT = Op.getSimpleValueType();
19846 SDValue Mask = N->getMask();
19849 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
19850 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
19851 // This operation is legal for targets with VLX, but without
19852 // VLX the vector should be widened to 512 bit
19853 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
19854 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
19855 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
19856 SDValue Src0 = N->getSrc0();
19857 Src0 = ExtendToType(Src0, WideDataVT, DAG);
19858 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
19859 SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(),
19860 N->getBasePtr(), Mask, Src0,
19861 N->getMemoryVT(), N->getMemOperand(),
19862 N->getExtensionType());
19864 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
19865 NewLoad.getValue(0),
19866 DAG.getIntPtrConstant(0, dl));
19867 SDValue RetOps[] = {Exract, NewLoad.getValue(1)};
19868 return DAG.getMergeValues(RetOps, dl);
19873 static SDValue LowerMSTORE(SDValue Op, const X86Subtarget *Subtarget,
19874 SelectionDAG &DAG) {
19875 MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode());
19876 SDValue DataToStore = N->getValue();
19877 MVT VT = DataToStore.getSimpleValueType();
19878 SDValue Mask = N->getMask();
19881 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
19882 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
19883 // This operation is legal for targets with VLX, but without
19884 // VLX the vector should be widened to 512 bit
19885 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
19886 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
19887 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
19888 DataToStore = ExtendToType(DataToStore, WideDataVT, DAG);
19889 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
19890 return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(),
19891 Mask, N->getMemoryVT(), N->getMemOperand(),
19892 N->isTruncatingStore());
19897 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19898 SelectionDAG &DAG) {
19899 assert(Subtarget->hasAVX512() &&
19900 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19902 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19904 MVT VT = Op.getSimpleValueType();
19905 SDValue Index = N->getIndex();
19906 SDValue Mask = N->getMask();
19907 SDValue Src0 = N->getValue();
19908 MVT IndexVT = Index.getSimpleValueType();
19909 MVT MaskVT = Mask.getSimpleValueType();
19911 unsigned NumElts = VT.getVectorNumElements();
19912 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19914 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19915 !Index.getSimpleValueType().is512BitVector()) {
19916 // AVX512F supports only 512-bit vectors. Or data or index should
19917 // be 512 bit wide. If now the both index and data are 256-bit, but
19918 // the vector contains 8 elements, we just sign-extend the index
19919 if (NumElts == 8) {
19920 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19921 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19922 N->getOperand(3), Index };
19923 DAG.UpdateNodeOperands(N, Ops);
19927 // Minimal number of elements in Gather
19930 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
19931 Index = ExtendToType(Index, NewIndexVT, DAG);
19932 if (IndexVT.getScalarType() == MVT::i32)
19933 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19936 MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts);
19937 // At this point we have promoted mask operand
19938 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
19939 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
19940 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
19941 Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask);
19943 // The pass-thru value
19944 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
19945 Src0 = ExtendToType(Src0, NewVT, DAG);
19947 SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
19948 SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other),
19949 N->getMemoryVT(), dl, Ops,
19950 N->getMemOperand());
19951 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
19952 NewGather.getValue(0),
19953 DAG.getIntPtrConstant(0, dl));
19954 SDValue RetOps[] = {Exract, NewGather.getValue(1)};
19955 return DAG.getMergeValues(RetOps, dl);
19960 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19961 SelectionDAG &DAG) const {
19962 // TODO: Eventually, the lowering of these nodes should be informed by or
19963 // deferred to the GC strategy for the function in which they appear. For
19964 // now, however, they must be lowered to something. Since they are logically
19965 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19966 // require special handling for these nodes), lower them as literal NOOPs for
19968 SmallVector<SDValue, 2> Ops;
19970 Ops.push_back(Op.getOperand(0));
19971 if (Op->getGluedNode())
19972 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19975 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19976 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19981 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19982 SelectionDAG &DAG) const {
19983 // TODO: Eventually, the lowering of these nodes should be informed by or
19984 // deferred to the GC strategy for the function in which they appear. For
19985 // now, however, they must be lowered to something. Since they are logically
19986 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19987 // require special handling for these nodes), lower them as literal NOOPs for
19989 SmallVector<SDValue, 2> Ops;
19991 Ops.push_back(Op.getOperand(0));
19992 if (Op->getGluedNode())
19993 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19996 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19997 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20002 /// LowerOperation - Provide custom lowering hooks for some operations.
20004 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20005 switch (Op.getOpcode()) {
20006 default: llvm_unreachable("Should not custom lower this!");
20007 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20008 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20009 return LowerCMP_SWAP(Op, Subtarget, DAG);
20010 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20011 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20012 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20013 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20014 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
20015 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
20016 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20017 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20018 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20019 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20020 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20021 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20022 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20023 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20024 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20025 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20026 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20027 case ISD::SHL_PARTS:
20028 case ISD::SRA_PARTS:
20029 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20030 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20031 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20032 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20033 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20034 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20035 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20036 case ISD::SIGN_EXTEND_VECTOR_INREG:
20037 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
20038 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20039 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20040 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20041 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20043 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20044 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20045 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20046 case ISD::SETCC: return LowerSETCC(Op, DAG);
20047 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
20048 case ISD::SELECT: return LowerSELECT(Op, DAG);
20049 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20050 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20051 case ISD::VASTART: return LowerVASTART(Op, DAG);
20052 case ISD::VAARG: return LowerVAARG(Op, DAG);
20053 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20054 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20055 case ISD::INTRINSIC_VOID:
20056 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20057 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20058 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20059 case ISD::FRAME_TO_ARGS_OFFSET:
20060 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20061 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20062 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20063 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20064 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20065 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20066 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20067 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20068 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
20069 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
20071 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
20072 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20073 case ISD::UMUL_LOHI:
20074 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20075 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
20078 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20084 case ISD::UMULO: return LowerXALUO(Op, DAG);
20085 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20086 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20090 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20091 case ISD::ADD: return LowerADD(Op, DAG);
20092 case ISD::SUB: return LowerSUB(Op, DAG);
20096 case ISD::UMIN: return LowerMINMAX(Op, DAG);
20097 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20098 case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
20099 case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);
20100 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
20101 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
20102 case ISD::GC_TRANSITION_START:
20103 return LowerGC_TRANSITION_START(Op, DAG);
20104 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
20108 /// ReplaceNodeResults - Replace a node with an illegal result type
20109 /// with a new node built out of custom code.
20110 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20111 SmallVectorImpl<SDValue>&Results,
20112 SelectionDAG &DAG) const {
20114 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20115 switch (N->getOpcode()) {
20117 llvm_unreachable("Do not know how to custom type legalize this operation!");
20118 case X86ISD::AVG: {
20119 // Legalize types for X86ISD::AVG by expanding vectors.
20120 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20122 auto InVT = N->getValueType(0);
20123 auto InVTSize = InVT.getSizeInBits();
20124 const unsigned RegSize =
20125 (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
20126 assert((!Subtarget->hasAVX512() || RegSize < 512) &&
20127 "512-bit vector requires AVX512");
20128 assert((!Subtarget->hasAVX2() || RegSize < 256) &&
20129 "256-bit vector requires AVX2");
20131 auto ElemVT = InVT.getVectorElementType();
20132 auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
20133 RegSize / ElemVT.getSizeInBits());
20134 assert(RegSize % InVT.getSizeInBits() == 0);
20135 unsigned NumConcat = RegSize / InVT.getSizeInBits();
20137 SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
20138 Ops[0] = N->getOperand(0);
20139 SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20140 Ops[0] = N->getOperand(1);
20141 SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20143 SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
20144 Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
20145 DAG.getIntPtrConstant(0, dl)));
20148 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20149 case X86ISD::FMINC:
20151 case X86ISD::FMAXC:
20152 case X86ISD::FMAX: {
20153 EVT VT = N->getValueType(0);
20154 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
20155 SDValue UNDEF = DAG.getUNDEF(VT);
20156 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20157 N->getOperand(0), UNDEF);
20158 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20159 N->getOperand(1), UNDEF);
20160 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20163 case ISD::SIGN_EXTEND_INREG:
20168 // We don't want to expand or promote these.
20175 case ISD::UDIVREM: {
20176 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20177 Results.push_back(V);
20180 case ISD::FP_TO_SINT:
20181 case ISD::FP_TO_UINT: {
20182 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20184 std::pair<SDValue,SDValue> Vals =
20185 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20186 SDValue FIST = Vals.first, StackSlot = Vals.second;
20187 if (FIST.getNode()) {
20188 EVT VT = N->getValueType(0);
20189 // Return a load from the stack slot.
20190 if (StackSlot.getNode())
20191 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20192 MachinePointerInfo(),
20193 false, false, false, 0));
20195 Results.push_back(FIST);
20199 case ISD::UINT_TO_FP: {
20200 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20201 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20202 N->getValueType(0) != MVT::v2f32)
20204 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20206 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
20208 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20209 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20210 DAG.getBitcast(MVT::v2i64, VBias));
20211 Or = DAG.getBitcast(MVT::v2f64, Or);
20212 // TODO: Are there any fast-math-flags to propagate here?
20213 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20214 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20217 case ISD::FP_ROUND: {
20218 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20220 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20221 Results.push_back(V);
20224 case ISD::FP_EXTEND: {
20225 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
20226 // No other ValueType for FP_EXTEND should reach this point.
20227 assert(N->getValueType(0) == MVT::v2f32 &&
20228 "Do not know how to legalize this Node");
20231 case ISD::INTRINSIC_W_CHAIN: {
20232 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20234 default : llvm_unreachable("Do not know how to custom type "
20235 "legalize this intrinsic operation!");
20236 case Intrinsic::x86_rdtsc:
20237 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20239 case Intrinsic::x86_rdtscp:
20240 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20242 case Intrinsic::x86_rdpmc:
20243 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20246 case ISD::INTRINSIC_WO_CHAIN: {
20247 if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG))
20248 Results.push_back(V);
20251 case ISD::READCYCLECOUNTER: {
20252 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20255 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20256 EVT T = N->getValueType(0);
20257 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20258 bool Regs64bit = T == MVT::i128;
20259 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20260 SDValue cpInL, cpInH;
20261 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20262 DAG.getConstant(0, dl, HalfT));
20263 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20264 DAG.getConstant(1, dl, HalfT));
20265 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20266 Regs64bit ? X86::RAX : X86::EAX,
20268 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20269 Regs64bit ? X86::RDX : X86::EDX,
20270 cpInH, cpInL.getValue(1));
20271 SDValue swapInL, swapInH;
20272 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20273 DAG.getConstant(0, dl, HalfT));
20274 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20275 DAG.getConstant(1, dl, HalfT));
20276 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20277 Regs64bit ? X86::RBX : X86::EBX,
20278 swapInL, cpInH.getValue(1));
20279 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20280 Regs64bit ? X86::RCX : X86::ECX,
20281 swapInH, swapInL.getValue(1));
20282 SDValue Ops[] = { swapInH.getValue(0),
20284 swapInH.getValue(1) };
20285 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20286 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20287 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20288 X86ISD::LCMPXCHG8_DAG;
20289 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20290 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20291 Regs64bit ? X86::RAX : X86::EAX,
20292 HalfT, Result.getValue(1));
20293 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20294 Regs64bit ? X86::RDX : X86::EDX,
20295 HalfT, cpOutL.getValue(2));
20296 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20298 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20299 MVT::i32, cpOutH.getValue(2));
20301 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20302 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
20303 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20305 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20306 Results.push_back(Success);
20307 Results.push_back(EFLAGS.getValue(1));
20310 case ISD::ATOMIC_SWAP:
20311 case ISD::ATOMIC_LOAD_ADD:
20312 case ISD::ATOMIC_LOAD_SUB:
20313 case ISD::ATOMIC_LOAD_AND:
20314 case ISD::ATOMIC_LOAD_OR:
20315 case ISD::ATOMIC_LOAD_XOR:
20316 case ISD::ATOMIC_LOAD_NAND:
20317 case ISD::ATOMIC_LOAD_MIN:
20318 case ISD::ATOMIC_LOAD_MAX:
20319 case ISD::ATOMIC_LOAD_UMIN:
20320 case ISD::ATOMIC_LOAD_UMAX:
20321 case ISD::ATOMIC_LOAD: {
20322 // Delegate to generic TypeLegalization. Situations we can really handle
20323 // should have already been dealt with by AtomicExpandPass.cpp.
20326 case ISD::BITCAST: {
20327 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20328 EVT DstVT = N->getValueType(0);
20329 EVT SrcVT = N->getOperand(0)->getValueType(0);
20331 if (SrcVT != MVT::f64 ||
20332 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20335 unsigned NumElts = DstVT.getVectorNumElements();
20336 EVT SVT = DstVT.getVectorElementType();
20337 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20338 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20339 MVT::v2f64, N->getOperand(0));
20340 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
20342 if (ExperimentalVectorWideningLegalization) {
20343 // If we are legalizing vectors by widening, we already have the desired
20344 // legal vector type, just return it.
20345 Results.push_back(ToVecInt);
20349 SmallVector<SDValue, 8> Elts;
20350 for (unsigned i = 0, e = NumElts; i != e; ++i)
20351 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20352 ToVecInt, DAG.getIntPtrConstant(i, dl)));
20354 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20359 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20360 switch ((X86ISD::NodeType)Opcode) {
20361 case X86ISD::FIRST_NUMBER: break;
20362 case X86ISD::BSF: return "X86ISD::BSF";
20363 case X86ISD::BSR: return "X86ISD::BSR";
20364 case X86ISD::SHLD: return "X86ISD::SHLD";
20365 case X86ISD::SHRD: return "X86ISD::SHRD";
20366 case X86ISD::FAND: return "X86ISD::FAND";
20367 case X86ISD::FANDN: return "X86ISD::FANDN";
20368 case X86ISD::FOR: return "X86ISD::FOR";
20369 case X86ISD::FXOR: return "X86ISD::FXOR";
20370 case X86ISD::FILD: return "X86ISD::FILD";
20371 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20372 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20373 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20374 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20375 case X86ISD::FLD: return "X86ISD::FLD";
20376 case X86ISD::FST: return "X86ISD::FST";
20377 case X86ISD::CALL: return "X86ISD::CALL";
20378 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20379 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20380 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20381 case X86ISD::BT: return "X86ISD::BT";
20382 case X86ISD::CMP: return "X86ISD::CMP";
20383 case X86ISD::COMI: return "X86ISD::COMI";
20384 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20385 case X86ISD::CMPM: return "X86ISD::CMPM";
20386 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20387 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20388 case X86ISD::SETCC: return "X86ISD::SETCC";
20389 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20390 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20391 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20392 case X86ISD::CMOV: return "X86ISD::CMOV";
20393 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20394 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20395 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20396 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20397 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20398 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20399 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20400 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20401 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20402 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20403 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20404 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20405 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20406 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20407 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20408 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20409 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20410 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20411 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20412 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20413 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20414 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20415 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20416 case X86ISD::HADD: return "X86ISD::HADD";
20417 case X86ISD::HSUB: return "X86ISD::HSUB";
20418 case X86ISD::FHADD: return "X86ISD::FHADD";
20419 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20420 case X86ISD::ABS: return "X86ISD::ABS";
20421 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20422 case X86ISD::FMAX: return "X86ISD::FMAX";
20423 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20424 case X86ISD::FMIN: return "X86ISD::FMIN";
20425 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20426 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20427 case X86ISD::FMINC: return "X86ISD::FMINC";
20428 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20429 case X86ISD::FRCP: return "X86ISD::FRCP";
20430 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20431 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20432 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20433 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20434 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20435 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20436 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20437 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20438 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20439 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20440 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20441 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20442 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20443 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20444 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20445 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20446 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20447 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20448 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20449 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20450 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20451 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20452 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20453 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20454 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20455 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20456 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20457 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20458 case X86ISD::VSHL: return "X86ISD::VSHL";
20459 case X86ISD::VSRL: return "X86ISD::VSRL";
20460 case X86ISD::VSRA: return "X86ISD::VSRA";
20461 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20462 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20463 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20464 case X86ISD::CMPP: return "X86ISD::CMPP";
20465 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20466 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20467 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20468 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20469 case X86ISD::ADD: return "X86ISD::ADD";
20470 case X86ISD::SUB: return "X86ISD::SUB";
20471 case X86ISD::ADC: return "X86ISD::ADC";
20472 case X86ISD::SBB: return "X86ISD::SBB";
20473 case X86ISD::SMUL: return "X86ISD::SMUL";
20474 case X86ISD::UMUL: return "X86ISD::UMUL";
20475 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20476 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20477 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20478 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20479 case X86ISD::INC: return "X86ISD::INC";
20480 case X86ISD::DEC: return "X86ISD::DEC";
20481 case X86ISD::OR: return "X86ISD::OR";
20482 case X86ISD::XOR: return "X86ISD::XOR";
20483 case X86ISD::AND: return "X86ISD::AND";
20484 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20485 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20486 case X86ISD::PTEST: return "X86ISD::PTEST";
20487 case X86ISD::TESTP: return "X86ISD::TESTP";
20488 case X86ISD::TESTM: return "X86ISD::TESTM";
20489 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20490 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20491 case X86ISD::KTEST: return "X86ISD::KTEST";
20492 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20493 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20494 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20495 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20496 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20497 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20498 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20499 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20500 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20501 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20502 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20503 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20504 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20505 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20506 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20507 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20508 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20509 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20510 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20511 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20512 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20513 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20514 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20515 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20516 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20517 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20518 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20519 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20520 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20521 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20522 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20523 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20524 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20525 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20526 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20527 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20528 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20529 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20530 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20531 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20532 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20533 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20534 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20535 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20536 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20537 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20538 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20539 case X86ISD::SAHF: return "X86ISD::SAHF";
20540 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20541 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20542 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20543 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20544 case X86ISD::VPROT: return "X86ISD::VPROT";
20545 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20546 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20547 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20548 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20549 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20550 case X86ISD::FMADD: return "X86ISD::FMADD";
20551 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20552 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20553 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20554 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20555 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20556 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20557 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20558 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20559 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20560 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20561 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20562 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20563 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20564 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20565 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20566 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20567 case X86ISD::XTEST: return "X86ISD::XTEST";
20568 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20569 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20570 case X86ISD::SELECT: return "X86ISD::SELECT";
20571 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20572 case X86ISD::RCP28: return "X86ISD::RCP28";
20573 case X86ISD::EXP2: return "X86ISD::EXP2";
20574 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20575 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20576 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20577 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20578 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20579 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20580 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20581 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20582 case X86ISD::ADDS: return "X86ISD::ADDS";
20583 case X86ISD::SUBS: return "X86ISD::SUBS";
20584 case X86ISD::AVG: return "X86ISD::AVG";
20585 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20586 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20587 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20588 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20589 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20590 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20591 case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
20596 // isLegalAddressingMode - Return true if the addressing mode represented
20597 // by AM is legal for this target, for a load/store of the specified type.
20598 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20599 const AddrMode &AM, Type *Ty,
20600 unsigned AS) const {
20601 // X86 supports extremely general addressing modes.
20602 CodeModel::Model M = getTargetMachine().getCodeModel();
20603 Reloc::Model R = getTargetMachine().getRelocationModel();
20605 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20606 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20611 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20613 // If a reference to this global requires an extra load, we can't fold it.
20614 if (isGlobalStubReference(GVFlags))
20617 // If BaseGV requires a register for the PIC base, we cannot also have a
20618 // BaseReg specified.
20619 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20622 // If lower 4G is not available, then we must use rip-relative addressing.
20623 if ((M != CodeModel::Small || R != Reloc::Static) &&
20624 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20628 switch (AM.Scale) {
20634 // These scales always work.
20639 // These scales are formed with basereg+scalereg. Only accept if there is
20644 default: // Other stuff never works.
20651 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20652 unsigned Bits = Ty->getScalarSizeInBits();
20654 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20655 // particularly cheaper than those without.
20659 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20660 // variable shifts just as cheap as scalar ones.
20661 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20664 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20665 // fully general vector.
20669 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20670 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20672 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20673 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20674 return NumBits1 > NumBits2;
20677 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20678 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20681 if (!isTypeLegal(EVT::getEVT(Ty1)))
20684 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20686 // Assuming the caller doesn't have a zeroext or signext return parameter,
20687 // truncation all the way down to i1 is valid.
20691 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20692 return isInt<32>(Imm);
20695 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20696 // Can also use sub to handle negated immediates.
20697 return isInt<32>(Imm);
20700 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20701 if (!VT1.isInteger() || !VT2.isInteger())
20703 unsigned NumBits1 = VT1.getSizeInBits();
20704 unsigned NumBits2 = VT2.getSizeInBits();
20705 return NumBits1 > NumBits2;
20708 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20709 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20710 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20713 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20714 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20715 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20718 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20719 EVT VT1 = Val.getValueType();
20720 if (isZExtFree(VT1, VT2))
20723 if (Val.getOpcode() != ISD::LOAD)
20726 if (!VT1.isSimple() || !VT1.isInteger() ||
20727 !VT2.isSimple() || !VT2.isInteger())
20730 switch (VT1.getSimpleVT().SimpleTy) {
20735 // X86 has 8, 16, and 32-bit zero-extending loads.
20742 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20745 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20746 if (!Subtarget->hasAnyFMA())
20749 VT = VT.getScalarType();
20751 if (!VT.isSimple())
20754 switch (VT.getSimpleVT().SimpleTy) {
20765 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20766 // i16 instructions are longer (0x66 prefix) and potentially slower.
20767 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20770 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20771 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20772 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20773 /// are assumed to be legal.
20775 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20777 if (!VT.isSimple())
20780 // Not for i1 vectors
20781 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20784 // Very little shuffling can be done for 64-bit vectors right now.
20785 if (VT.getSimpleVT().getSizeInBits() == 64)
20788 // We only care that the types being shuffled are legal. The lowering can
20789 // handle any possible shuffle mask that results.
20790 return isTypeLegal(VT.getSimpleVT());
20794 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20796 // Just delegate to the generic legality, clear masks aren't special.
20797 return isShuffleMaskLegal(Mask, VT);
20800 //===----------------------------------------------------------------------===//
20801 // X86 Scheduler Hooks
20802 //===----------------------------------------------------------------------===//
20804 /// Utility function to emit xbegin specifying the start of an RTM region.
20805 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20806 const TargetInstrInfo *TII) {
20807 DebugLoc DL = MI->getDebugLoc();
20809 const BasicBlock *BB = MBB->getBasicBlock();
20810 MachineFunction::iterator I = ++MBB->getIterator();
20812 // For the v = xbegin(), we generate
20823 MachineBasicBlock *thisMBB = MBB;
20824 MachineFunction *MF = MBB->getParent();
20825 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20826 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20827 MF->insert(I, mainMBB);
20828 MF->insert(I, sinkMBB);
20830 // Transfer the remainder of BB and its successor edges to sinkMBB.
20831 sinkMBB->splice(sinkMBB->begin(), MBB,
20832 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20833 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20837 // # fallthrough to mainMBB
20838 // # abortion to sinkMBB
20839 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20840 thisMBB->addSuccessor(mainMBB);
20841 thisMBB->addSuccessor(sinkMBB);
20845 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20846 mainMBB->addSuccessor(sinkMBB);
20849 // EAX is live into the sinkMBB
20850 sinkMBB->addLiveIn(X86::EAX);
20851 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20852 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20855 MI->eraseFromParent();
20859 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20860 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20861 // in the .td file.
20862 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20863 const TargetInstrInfo *TII) {
20865 switch (MI->getOpcode()) {
20866 default: llvm_unreachable("illegal opcode!");
20867 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20868 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20869 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20870 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20871 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20872 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20873 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20874 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20877 DebugLoc dl = MI->getDebugLoc();
20878 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20880 unsigned NumArgs = MI->getNumOperands();
20881 for (unsigned i = 1; i < NumArgs; ++i) {
20882 MachineOperand &Op = MI->getOperand(i);
20883 if (!(Op.isReg() && Op.isImplicit()))
20884 MIB.addOperand(Op);
20886 if (MI->hasOneMemOperand())
20887 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20889 BuildMI(*BB, MI, dl,
20890 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20891 .addReg(X86::XMM0);
20893 MI->eraseFromParent();
20897 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20898 // defs in an instruction pattern
20899 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20900 const TargetInstrInfo *TII) {
20902 switch (MI->getOpcode()) {
20903 default: llvm_unreachable("illegal opcode!");
20904 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20905 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20906 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20907 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20908 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20909 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20910 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20911 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20914 DebugLoc dl = MI->getDebugLoc();
20915 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20917 unsigned NumArgs = MI->getNumOperands(); // remove the results
20918 for (unsigned i = 1; i < NumArgs; ++i) {
20919 MachineOperand &Op = MI->getOperand(i);
20920 if (!(Op.isReg() && Op.isImplicit()))
20921 MIB.addOperand(Op);
20923 if (MI->hasOneMemOperand())
20924 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20926 BuildMI(*BB, MI, dl,
20927 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20930 MI->eraseFromParent();
20934 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20935 const X86Subtarget *Subtarget) {
20936 DebugLoc dl = MI->getDebugLoc();
20937 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20938 // Address into RAX/EAX, other two args into ECX, EDX.
20939 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20940 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20941 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20942 for (int i = 0; i < X86::AddrNumOperands; ++i)
20943 MIB.addOperand(MI->getOperand(i));
20945 unsigned ValOps = X86::AddrNumOperands;
20946 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20947 .addReg(MI->getOperand(ValOps).getReg());
20948 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20949 .addReg(MI->getOperand(ValOps+1).getReg());
20951 // The instruction doesn't actually take any operands though.
20952 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20954 MI->eraseFromParent(); // The pseudo is gone now.
20958 MachineBasicBlock *
20959 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20960 MachineBasicBlock *MBB) const {
20961 // Emit va_arg instruction on X86-64.
20963 // Operands to this pseudo-instruction:
20964 // 0 ) Output : destination address (reg)
20965 // 1-5) Input : va_list address (addr, i64mem)
20966 // 6 ) ArgSize : Size (in bytes) of vararg type
20967 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20968 // 8 ) Align : Alignment of type
20969 // 9 ) EFLAGS (implicit-def)
20971 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20972 static_assert(X86::AddrNumOperands == 5,
20973 "VAARG_64 assumes 5 address operands");
20975 unsigned DestReg = MI->getOperand(0).getReg();
20976 MachineOperand &Base = MI->getOperand(1);
20977 MachineOperand &Scale = MI->getOperand(2);
20978 MachineOperand &Index = MI->getOperand(3);
20979 MachineOperand &Disp = MI->getOperand(4);
20980 MachineOperand &Segment = MI->getOperand(5);
20981 unsigned ArgSize = MI->getOperand(6).getImm();
20982 unsigned ArgMode = MI->getOperand(7).getImm();
20983 unsigned Align = MI->getOperand(8).getImm();
20985 // Memory Reference
20986 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20987 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20988 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20990 // Machine Information
20991 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20992 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20993 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20994 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20995 DebugLoc DL = MI->getDebugLoc();
20997 // struct va_list {
21000 // i64 overflow_area (address)
21001 // i64 reg_save_area (address)
21003 // sizeof(va_list) = 24
21004 // alignment(va_list) = 8
21006 unsigned TotalNumIntRegs = 6;
21007 unsigned TotalNumXMMRegs = 8;
21008 bool UseGPOffset = (ArgMode == 1);
21009 bool UseFPOffset = (ArgMode == 2);
21010 unsigned MaxOffset = TotalNumIntRegs * 8 +
21011 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21013 /* Align ArgSize to a multiple of 8 */
21014 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21015 bool NeedsAlign = (Align > 8);
21017 MachineBasicBlock *thisMBB = MBB;
21018 MachineBasicBlock *overflowMBB;
21019 MachineBasicBlock *offsetMBB;
21020 MachineBasicBlock *endMBB;
21022 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21023 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21024 unsigned OffsetReg = 0;
21026 if (!UseGPOffset && !UseFPOffset) {
21027 // If we only pull from the overflow region, we don't create a branch.
21028 // We don't need to alter control flow.
21029 OffsetDestReg = 0; // unused
21030 OverflowDestReg = DestReg;
21032 offsetMBB = nullptr;
21033 overflowMBB = thisMBB;
21036 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21037 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21038 // If not, pull from overflow_area. (branch to overflowMBB)
21043 // offsetMBB overflowMBB
21048 // Registers for the PHI in endMBB
21049 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21050 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21052 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21053 MachineFunction *MF = MBB->getParent();
21054 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21055 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21056 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21058 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21060 // Insert the new basic blocks
21061 MF->insert(MBBIter, offsetMBB);
21062 MF->insert(MBBIter, overflowMBB);
21063 MF->insert(MBBIter, endMBB);
21065 // Transfer the remainder of MBB and its successor edges to endMBB.
21066 endMBB->splice(endMBB->begin(), thisMBB,
21067 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21068 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21070 // Make offsetMBB and overflowMBB successors of thisMBB
21071 thisMBB->addSuccessor(offsetMBB);
21072 thisMBB->addSuccessor(overflowMBB);
21074 // endMBB is a successor of both offsetMBB and overflowMBB
21075 offsetMBB->addSuccessor(endMBB);
21076 overflowMBB->addSuccessor(endMBB);
21078 // Load the offset value into a register
21079 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21080 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21084 .addDisp(Disp, UseFPOffset ? 4 : 0)
21085 .addOperand(Segment)
21086 .setMemRefs(MMOBegin, MMOEnd);
21088 // Check if there is enough room left to pull this argument.
21089 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21091 .addImm(MaxOffset + 8 - ArgSizeA8);
21093 // Branch to "overflowMBB" if offset >= max
21094 // Fall through to "offsetMBB" otherwise
21095 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21096 .addMBB(overflowMBB);
21099 // In offsetMBB, emit code to use the reg_save_area.
21101 assert(OffsetReg != 0);
21103 // Read the reg_save_area address.
21104 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21105 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21110 .addOperand(Segment)
21111 .setMemRefs(MMOBegin, MMOEnd);
21113 // Zero-extend the offset
21114 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21115 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21118 .addImm(X86::sub_32bit);
21120 // Add the offset to the reg_save_area to get the final address.
21121 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21122 .addReg(OffsetReg64)
21123 .addReg(RegSaveReg);
21125 // Compute the offset for the next argument
21126 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21127 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21129 .addImm(UseFPOffset ? 16 : 8);
21131 // Store it back into the va_list.
21132 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21136 .addDisp(Disp, UseFPOffset ? 4 : 0)
21137 .addOperand(Segment)
21138 .addReg(NextOffsetReg)
21139 .setMemRefs(MMOBegin, MMOEnd);
21142 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21147 // Emit code to use overflow area
21150 // Load the overflow_area address into a register.
21151 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21152 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21157 .addOperand(Segment)
21158 .setMemRefs(MMOBegin, MMOEnd);
21160 // If we need to align it, do so. Otherwise, just copy the address
21161 // to OverflowDestReg.
21163 // Align the overflow address
21164 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21165 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21167 // aligned_addr = (addr + (align-1)) & ~(align-1)
21168 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21169 .addReg(OverflowAddrReg)
21172 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21174 .addImm(~(uint64_t)(Align-1));
21176 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21177 .addReg(OverflowAddrReg);
21180 // Compute the next overflow address after this argument.
21181 // (the overflow address should be kept 8-byte aligned)
21182 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21183 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21184 .addReg(OverflowDestReg)
21185 .addImm(ArgSizeA8);
21187 // Store the new overflow address.
21188 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21193 .addOperand(Segment)
21194 .addReg(NextAddrReg)
21195 .setMemRefs(MMOBegin, MMOEnd);
21197 // If we branched, emit the PHI to the front of endMBB.
21199 BuildMI(*endMBB, endMBB->begin(), DL,
21200 TII->get(X86::PHI), DestReg)
21201 .addReg(OffsetDestReg).addMBB(offsetMBB)
21202 .addReg(OverflowDestReg).addMBB(overflowMBB);
21205 // Erase the pseudo instruction
21206 MI->eraseFromParent();
21211 MachineBasicBlock *
21212 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21214 MachineBasicBlock *MBB) const {
21215 // Emit code to save XMM registers to the stack. The ABI says that the
21216 // number of registers to save is given in %al, so it's theoretically
21217 // possible to do an indirect jump trick to avoid saving all of them,
21218 // however this code takes a simpler approach and just executes all
21219 // of the stores if %al is non-zero. It's less code, and it's probably
21220 // easier on the hardware branch predictor, and stores aren't all that
21221 // expensive anyway.
21223 // Create the new basic blocks. One block contains all the XMM stores,
21224 // and one block is the final destination regardless of whether any
21225 // stores were performed.
21226 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21227 MachineFunction *F = MBB->getParent();
21228 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21229 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21230 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21231 F->insert(MBBIter, XMMSaveMBB);
21232 F->insert(MBBIter, EndMBB);
21234 // Transfer the remainder of MBB and its successor edges to EndMBB.
21235 EndMBB->splice(EndMBB->begin(), MBB,
21236 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21237 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21239 // The original block will now fall through to the XMM save block.
21240 MBB->addSuccessor(XMMSaveMBB);
21241 // The XMMSaveMBB will fall through to the end block.
21242 XMMSaveMBB->addSuccessor(EndMBB);
21244 // Now add the instructions.
21245 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21246 DebugLoc DL = MI->getDebugLoc();
21248 unsigned CountReg = MI->getOperand(0).getReg();
21249 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21250 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21252 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
21253 // If %al is 0, branch around the XMM save block.
21254 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21255 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21256 MBB->addSuccessor(EndMBB);
21259 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21260 // that was just emitted, but clearly shouldn't be "saved".
21261 assert((MI->getNumOperands() <= 3 ||
21262 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21263 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21264 && "Expected last argument to be EFLAGS");
21265 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21266 // In the XMM save block, save all the XMM argument registers.
21267 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21268 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21269 MachineMemOperand *MMO = F->getMachineMemOperand(
21270 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
21271 MachineMemOperand::MOStore,
21272 /*Size=*/16, /*Align=*/16);
21273 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21274 .addFrameIndex(RegSaveFrameIndex)
21275 .addImm(/*Scale=*/1)
21276 .addReg(/*IndexReg=*/0)
21277 .addImm(/*Disp=*/Offset)
21278 .addReg(/*Segment=*/0)
21279 .addReg(MI->getOperand(i).getReg())
21280 .addMemOperand(MMO);
21283 MI->eraseFromParent(); // The pseudo instruction is gone now.
21288 // The EFLAGS operand of SelectItr might be missing a kill marker
21289 // because there were multiple uses of EFLAGS, and ISel didn't know
21290 // which to mark. Figure out whether SelectItr should have had a
21291 // kill marker, and set it if it should. Returns the correct kill
21293 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21294 MachineBasicBlock* BB,
21295 const TargetRegisterInfo* TRI) {
21296 // Scan forward through BB for a use/def of EFLAGS.
21297 MachineBasicBlock::iterator miI(std::next(SelectItr));
21298 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21299 const MachineInstr& mi = *miI;
21300 if (mi.readsRegister(X86::EFLAGS))
21302 if (mi.definesRegister(X86::EFLAGS))
21303 break; // Should have kill-flag - update below.
21306 // If we hit the end of the block, check whether EFLAGS is live into a
21308 if (miI == BB->end()) {
21309 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21310 sEnd = BB->succ_end();
21311 sItr != sEnd; ++sItr) {
21312 MachineBasicBlock* succ = *sItr;
21313 if (succ->isLiveIn(X86::EFLAGS))
21318 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21319 // out. SelectMI should have a kill flag on EFLAGS.
21320 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21324 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
21325 // together with other CMOV pseudo-opcodes into a single basic-block with
21326 // conditional jump around it.
21327 static bool isCMOVPseudo(MachineInstr *MI) {
21328 switch (MI->getOpcode()) {
21329 case X86::CMOV_FR32:
21330 case X86::CMOV_FR64:
21331 case X86::CMOV_GR8:
21332 case X86::CMOV_GR16:
21333 case X86::CMOV_GR32:
21334 case X86::CMOV_RFP32:
21335 case X86::CMOV_RFP64:
21336 case X86::CMOV_RFP80:
21337 case X86::CMOV_V2F64:
21338 case X86::CMOV_V2I64:
21339 case X86::CMOV_V4F32:
21340 case X86::CMOV_V4F64:
21341 case X86::CMOV_V4I64:
21342 case X86::CMOV_V16F32:
21343 case X86::CMOV_V8F32:
21344 case X86::CMOV_V8F64:
21345 case X86::CMOV_V8I64:
21346 case X86::CMOV_V8I1:
21347 case X86::CMOV_V16I1:
21348 case X86::CMOV_V32I1:
21349 case X86::CMOV_V64I1:
21357 MachineBasicBlock *
21358 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21359 MachineBasicBlock *BB) const {
21360 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21361 DebugLoc DL = MI->getDebugLoc();
21363 // To "insert" a SELECT_CC instruction, we actually have to insert the
21364 // diamond control-flow pattern. The incoming instruction knows the
21365 // destination vreg to set, the condition code register to branch on, the
21366 // true/false values to select between, and a branch opcode to use.
21367 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21368 MachineFunction::iterator It = ++BB->getIterator();
21373 // cmpTY ccX, r1, r2
21375 // fallthrough --> copy0MBB
21376 MachineBasicBlock *thisMBB = BB;
21377 MachineFunction *F = BB->getParent();
21379 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21380 // as described above, by inserting a BB, and then making a PHI at the join
21381 // point to select the true and false operands of the CMOV in the PHI.
21383 // The code also handles two different cases of multiple CMOV opcodes
21387 // In this case, there are multiple CMOVs in a row, all which are based on
21388 // the same condition setting (or the exact opposite condition setting).
21389 // In this case we can lower all the CMOVs using a single inserted BB, and
21390 // then make a number of PHIs at the join point to model the CMOVs. The only
21391 // trickiness here, is that in a case like:
21393 // t2 = CMOV cond1 t1, f1
21394 // t3 = CMOV cond1 t2, f2
21396 // when rewriting this into PHIs, we have to perform some renaming on the
21397 // temps since you cannot have a PHI operand refer to a PHI result earlier
21398 // in the same block. The "simple" but wrong lowering would be:
21400 // t2 = PHI t1(BB1), f1(BB2)
21401 // t3 = PHI t2(BB1), f2(BB2)
21403 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21404 // renaming is to note that on the path through BB1, t2 is really just a
21405 // copy of t1, and do that renaming, properly generating:
21407 // t2 = PHI t1(BB1), f1(BB2)
21408 // t3 = PHI t1(BB1), f2(BB2)
21410 // Case 2, we lower cascaded CMOVs such as
21412 // (CMOV (CMOV F, T, cc1), T, cc2)
21414 // to two successives branches. For that, we look for another CMOV as the
21415 // following instruction.
21417 // Without this, we would add a PHI between the two jumps, which ends up
21418 // creating a few copies all around. For instance, for
21420 // (sitofp (zext (fcmp une)))
21422 // we would generate:
21424 // ucomiss %xmm1, %xmm0
21425 // movss <1.0f>, %xmm0
21426 // movaps %xmm0, %xmm1
21428 // xorps %xmm1, %xmm1
21431 // movaps %xmm1, %xmm0
21435 // because this custom-inserter would have generated:
21447 // A: X = ...; Y = ...
21449 // C: Z = PHI [X, A], [Y, B]
21451 // E: PHI [X, C], [Z, D]
21453 // If we lower both CMOVs in a single step, we can instead generate:
21465 // A: X = ...; Y = ...
21467 // E: PHI [X, A], [X, C], [Y, D]
21469 // Which, in our sitofp/fcmp example, gives us something like:
21471 // ucomiss %xmm1, %xmm0
21472 // movss <1.0f>, %xmm0
21475 // xorps %xmm0, %xmm0
21479 MachineInstr *CascadedCMOV = nullptr;
21480 MachineInstr *LastCMOV = MI;
21481 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21482 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21483 MachineBasicBlock::iterator NextMIIt =
21484 std::next(MachineBasicBlock::iterator(MI));
21486 // Check for case 1, where there are multiple CMOVs with the same condition
21487 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21488 // number of jumps the most.
21490 if (isCMOVPseudo(MI)) {
21491 // See if we have a string of CMOVS with the same condition.
21492 while (NextMIIt != BB->end() &&
21493 isCMOVPseudo(NextMIIt) &&
21494 (NextMIIt->getOperand(3).getImm() == CC ||
21495 NextMIIt->getOperand(3).getImm() == OppCC)) {
21496 LastCMOV = &*NextMIIt;
21501 // This checks for case 2, but only do this if we didn't already find
21502 // case 1, as indicated by LastCMOV == MI.
21503 if (LastCMOV == MI &&
21504 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21505 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21506 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21507 CascadedCMOV = &*NextMIIt;
21510 MachineBasicBlock *jcc1MBB = nullptr;
21512 // If we have a cascaded CMOV, we lower it to two successive branches to
21513 // the same block. EFLAGS is used by both, so mark it as live in the second.
21514 if (CascadedCMOV) {
21515 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21516 F->insert(It, jcc1MBB);
21517 jcc1MBB->addLiveIn(X86::EFLAGS);
21520 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21521 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21522 F->insert(It, copy0MBB);
21523 F->insert(It, sinkMBB);
21525 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21526 // live into the sink and copy blocks.
21527 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21529 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21530 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21531 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21532 copy0MBB->addLiveIn(X86::EFLAGS);
21533 sinkMBB->addLiveIn(X86::EFLAGS);
21536 // Transfer the remainder of BB and its successor edges to sinkMBB.
21537 sinkMBB->splice(sinkMBB->begin(), BB,
21538 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21539 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21541 // Add the true and fallthrough blocks as its successors.
21542 if (CascadedCMOV) {
21543 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21544 BB->addSuccessor(jcc1MBB);
21546 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21547 // jump to the sinkMBB.
21548 jcc1MBB->addSuccessor(copy0MBB);
21549 jcc1MBB->addSuccessor(sinkMBB);
21551 BB->addSuccessor(copy0MBB);
21554 // The true block target of the first (or only) branch is always sinkMBB.
21555 BB->addSuccessor(sinkMBB);
21557 // Create the conditional branch instruction.
21558 unsigned Opc = X86::GetCondBranchFromCond(CC);
21559 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21561 if (CascadedCMOV) {
21562 unsigned Opc2 = X86::GetCondBranchFromCond(
21563 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21564 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21568 // %FalseValue = ...
21569 // # fallthrough to sinkMBB
21570 copy0MBB->addSuccessor(sinkMBB);
21573 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21575 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21576 MachineBasicBlock::iterator MIItEnd =
21577 std::next(MachineBasicBlock::iterator(LastCMOV));
21578 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21579 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21580 MachineInstrBuilder MIB;
21582 // As we are creating the PHIs, we have to be careful if there is more than
21583 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21584 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21585 // That also means that PHI construction must work forward from earlier to
21586 // later, and that the code must maintain a mapping from earlier PHI's
21587 // destination registers, and the registers that went into the PHI.
21589 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21590 unsigned DestReg = MIIt->getOperand(0).getReg();
21591 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21592 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21594 // If this CMOV we are generating is the opposite condition from
21595 // the jump we generated, then we have to swap the operands for the
21596 // PHI that is going to be generated.
21597 if (MIIt->getOperand(3).getImm() == OppCC)
21598 std::swap(Op1Reg, Op2Reg);
21600 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21601 Op1Reg = RegRewriteTable[Op1Reg].first;
21603 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21604 Op2Reg = RegRewriteTable[Op2Reg].second;
21606 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21607 TII->get(X86::PHI), DestReg)
21608 .addReg(Op1Reg).addMBB(copy0MBB)
21609 .addReg(Op2Reg).addMBB(thisMBB);
21611 // Add this PHI to the rewrite table.
21612 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21615 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21616 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21617 if (CascadedCMOV) {
21618 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21619 // Copy the PHI result to the register defined by the second CMOV.
21620 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21621 DL, TII->get(TargetOpcode::COPY),
21622 CascadedCMOV->getOperand(0).getReg())
21623 .addReg(MI->getOperand(0).getReg());
21624 CascadedCMOV->eraseFromParent();
21627 // Now remove the CMOV(s).
21628 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21629 (MIIt++)->eraseFromParent();
21634 MachineBasicBlock *
21635 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21636 MachineBasicBlock *BB) const {
21637 // Combine the following atomic floating-point modification pattern:
21638 // a.store(reg OP a.load(acquire), release)
21639 // Transform them into:
21640 // OPss (%gpr), %xmm
21641 // movss %xmm, (%gpr)
21642 // Or sd equivalent for 64-bit operations.
21644 switch (MI->getOpcode()) {
21645 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21646 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21647 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21649 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21650 DebugLoc DL = MI->getDebugLoc();
21651 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21652 MachineOperand MSrc = MI->getOperand(0);
21653 unsigned VSrc = MI->getOperand(5).getReg();
21654 const MachineOperand &Disp = MI->getOperand(3);
21655 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21656 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21657 if (hasDisp && MSrc.isReg())
21658 MSrc.setIsKill(false);
21659 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21660 .addOperand(/*Base=*/MSrc)
21661 .addImm(/*Scale=*/1)
21662 .addReg(/*Index=*/0)
21663 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21665 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21666 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21668 .addOperand(/*Base=*/MSrc)
21669 .addImm(/*Scale=*/1)
21670 .addReg(/*Index=*/0)
21671 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21672 .addReg(/*Segment=*/0);
21673 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21674 MI->eraseFromParent(); // The pseudo instruction is gone now.
21678 MachineBasicBlock *
21679 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21680 MachineBasicBlock *BB) const {
21681 MachineFunction *MF = BB->getParent();
21682 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21683 DebugLoc DL = MI->getDebugLoc();
21684 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21686 assert(MF->shouldSplitStack());
21688 const bool Is64Bit = Subtarget->is64Bit();
21689 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21691 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21692 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21695 // ... [Till the alloca]
21696 // If stacklet is not large enough, jump to mallocMBB
21699 // Allocate by subtracting from RSP
21700 // Jump to continueMBB
21703 // Allocate by call to runtime
21707 // [rest of original BB]
21710 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21711 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21712 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21714 MachineRegisterInfo &MRI = MF->getRegInfo();
21715 const TargetRegisterClass *AddrRegClass =
21716 getRegClassFor(getPointerTy(MF->getDataLayout()));
21718 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21719 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21720 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21721 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21722 sizeVReg = MI->getOperand(1).getReg(),
21723 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21725 MachineFunction::iterator MBBIter = ++BB->getIterator();
21727 MF->insert(MBBIter, bumpMBB);
21728 MF->insert(MBBIter, mallocMBB);
21729 MF->insert(MBBIter, continueMBB);
21731 continueMBB->splice(continueMBB->begin(), BB,
21732 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21733 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21735 // Add code to the main basic block to check if the stack limit has been hit,
21736 // and if so, jump to mallocMBB otherwise to bumpMBB.
21737 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21738 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21739 .addReg(tmpSPVReg).addReg(sizeVReg);
21740 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21741 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21742 .addReg(SPLimitVReg);
21743 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21745 // bumpMBB simply decreases the stack pointer, since we know the current
21746 // stacklet has enough space.
21747 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21748 .addReg(SPLimitVReg);
21749 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21750 .addReg(SPLimitVReg);
21751 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21753 // Calls into a routine in libgcc to allocate more space from the heap.
21754 const uint32_t *RegMask =
21755 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21757 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21759 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21760 .addExternalSymbol("__morestack_allocate_stack_space")
21761 .addRegMask(RegMask)
21762 .addReg(X86::RDI, RegState::Implicit)
21763 .addReg(X86::RAX, RegState::ImplicitDefine);
21764 } else if (Is64Bit) {
21765 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21767 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21768 .addExternalSymbol("__morestack_allocate_stack_space")
21769 .addRegMask(RegMask)
21770 .addReg(X86::EDI, RegState::Implicit)
21771 .addReg(X86::EAX, RegState::ImplicitDefine);
21773 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21775 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21776 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21777 .addExternalSymbol("__morestack_allocate_stack_space")
21778 .addRegMask(RegMask)
21779 .addReg(X86::EAX, RegState::ImplicitDefine);
21783 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21786 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21787 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21788 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21790 // Set up the CFG correctly.
21791 BB->addSuccessor(bumpMBB);
21792 BB->addSuccessor(mallocMBB);
21793 mallocMBB->addSuccessor(continueMBB);
21794 bumpMBB->addSuccessor(continueMBB);
21796 // Take care of the PHI nodes.
21797 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21798 MI->getOperand(0).getReg())
21799 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21800 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21802 // Delete the original pseudo instruction.
21803 MI->eraseFromParent();
21806 return continueMBB;
21809 MachineBasicBlock *
21810 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21811 MachineBasicBlock *BB) const {
21812 assert(!Subtarget->isTargetMachO());
21813 DebugLoc DL = MI->getDebugLoc();
21814 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
21815 *BB->getParent(), *BB, MI, DL, false);
21816 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
21817 MI->eraseFromParent(); // The pseudo instruction is gone now.
21821 MachineBasicBlock *
21822 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
21823 MachineBasicBlock *BB) const {
21824 MachineFunction *MF = BB->getParent();
21825 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21826 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
21827 DebugLoc DL = MI->getDebugLoc();
21829 assert(!isAsynchronousEHPersonality(
21830 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
21831 "SEH does not use catchret!");
21833 // Only 32-bit EH needs to worry about manually restoring stack pointers.
21834 if (!Subtarget->is32Bit())
21837 // C++ EH creates a new target block to hold the restore code, and wires up
21838 // the new block to the return destination with a normal JMP_4.
21839 MachineBasicBlock *RestoreMBB =
21840 MF->CreateMachineBasicBlock(BB->getBasicBlock());
21841 assert(BB->succ_size() == 1);
21842 MF->insert(std::next(BB->getIterator()), RestoreMBB);
21843 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
21844 BB->addSuccessor(RestoreMBB);
21845 MI->getOperand(0).setMBB(RestoreMBB);
21847 auto RestoreMBBI = RestoreMBB->begin();
21848 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
21849 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
21853 MachineBasicBlock *
21854 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
21855 MachineBasicBlock *BB) const {
21856 MachineFunction *MF = BB->getParent();
21857 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
21858 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
21859 // Only 32-bit SEH requires special handling for catchpad.
21860 if (IsSEH && Subtarget->is32Bit()) {
21861 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21862 DebugLoc DL = MI->getDebugLoc();
21863 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
21865 MI->eraseFromParent();
21869 MachineBasicBlock *
21870 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21871 MachineBasicBlock *BB) const {
21872 // This is pretty easy. We're taking the value that we received from
21873 // our load from the relocation, sticking it in either RDI (x86-64)
21874 // or EAX and doing an indirect call. The return value will then
21875 // be in the normal return register.
21876 MachineFunction *F = BB->getParent();
21877 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21878 DebugLoc DL = MI->getDebugLoc();
21880 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21881 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21883 // Get a register mask for the lowered call.
21884 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21885 // proper register mask.
21886 const uint32_t *RegMask =
21887 Subtarget->is64Bit() ?
21888 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
21889 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21890 if (Subtarget->is64Bit()) {
21891 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21892 TII->get(X86::MOV64rm), X86::RDI)
21894 .addImm(0).addReg(0)
21895 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21896 MI->getOperand(3).getTargetFlags())
21898 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21899 addDirectMem(MIB, X86::RDI);
21900 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21901 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21902 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21903 TII->get(X86::MOV32rm), X86::EAX)
21905 .addImm(0).addReg(0)
21906 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21907 MI->getOperand(3).getTargetFlags())
21909 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21910 addDirectMem(MIB, X86::EAX);
21911 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21913 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21914 TII->get(X86::MOV32rm), X86::EAX)
21915 .addReg(TII->getGlobalBaseReg(F))
21916 .addImm(0).addReg(0)
21917 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21918 MI->getOperand(3).getTargetFlags())
21920 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21921 addDirectMem(MIB, X86::EAX);
21922 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21925 MI->eraseFromParent(); // The pseudo instruction is gone now.
21929 MachineBasicBlock *
21930 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21931 MachineBasicBlock *MBB) const {
21932 DebugLoc DL = MI->getDebugLoc();
21933 MachineFunction *MF = MBB->getParent();
21934 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21935 MachineRegisterInfo &MRI = MF->getRegInfo();
21937 const BasicBlock *BB = MBB->getBasicBlock();
21938 MachineFunction::iterator I = ++MBB->getIterator();
21940 // Memory Reference
21941 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21942 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21945 unsigned MemOpndSlot = 0;
21947 unsigned CurOp = 0;
21949 DstReg = MI->getOperand(CurOp++).getReg();
21950 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21951 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21952 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21953 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21955 MemOpndSlot = CurOp;
21957 MVT PVT = getPointerTy(MF->getDataLayout());
21958 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21959 "Invalid Pointer Size!");
21961 // For v = setjmp(buf), we generate
21964 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21965 // SjLjSetup restoreMBB
21971 // v = phi(main, restore)
21974 // if base pointer being used, load it from frame
21977 MachineBasicBlock *thisMBB = MBB;
21978 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21979 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21980 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21981 MF->insert(I, mainMBB);
21982 MF->insert(I, sinkMBB);
21983 MF->push_back(restoreMBB);
21984 restoreMBB->setHasAddressTaken();
21986 MachineInstrBuilder MIB;
21988 // Transfer the remainder of BB and its successor edges to sinkMBB.
21989 sinkMBB->splice(sinkMBB->begin(), MBB,
21990 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21991 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21994 unsigned PtrStoreOpc = 0;
21995 unsigned LabelReg = 0;
21996 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21997 Reloc::Model RM = MF->getTarget().getRelocationModel();
21998 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21999 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
22001 // Prepare IP either in reg or imm.
22002 if (!UseImmLabel) {
22003 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
22004 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
22005 LabelReg = MRI.createVirtualRegister(PtrRC);
22006 if (Subtarget->is64Bit()) {
22007 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
22011 .addMBB(restoreMBB)
22014 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
22015 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
22016 .addReg(XII->getGlobalBaseReg(MF))
22019 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
22023 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
22025 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
22026 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22027 if (i == X86::AddrDisp)
22028 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
22030 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
22033 MIB.addReg(LabelReg);
22035 MIB.addMBB(restoreMBB);
22036 MIB.setMemRefs(MMOBegin, MMOEnd);
22038 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
22039 .addMBB(restoreMBB);
22041 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22042 MIB.addRegMask(RegInfo->getNoPreservedMask());
22043 thisMBB->addSuccessor(mainMBB);
22044 thisMBB->addSuccessor(restoreMBB);
22048 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
22049 mainMBB->addSuccessor(sinkMBB);
22052 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
22053 TII->get(X86::PHI), DstReg)
22054 .addReg(mainDstReg).addMBB(mainMBB)
22055 .addReg(restoreDstReg).addMBB(restoreMBB);
22058 if (RegInfo->hasBasePointer(*MF)) {
22059 const bool Uses64BitFramePtr =
22060 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
22061 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
22062 X86FI->setRestoreBasePointer(MF);
22063 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
22064 unsigned BasePtr = RegInfo->getBaseRegister();
22065 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
22066 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
22067 FramePtr, true, X86FI->getRestoreBasePointerOffset())
22068 .setMIFlag(MachineInstr::FrameSetup);
22070 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
22071 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
22072 restoreMBB->addSuccessor(sinkMBB);
22074 MI->eraseFromParent();
22078 MachineBasicBlock *
22079 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
22080 MachineBasicBlock *MBB) const {
22081 DebugLoc DL = MI->getDebugLoc();
22082 MachineFunction *MF = MBB->getParent();
22083 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22084 MachineRegisterInfo &MRI = MF->getRegInfo();
22086 // Memory Reference
22087 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22088 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22090 MVT PVT = getPointerTy(MF->getDataLayout());
22091 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22092 "Invalid Pointer Size!");
22094 const TargetRegisterClass *RC =
22095 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
22096 unsigned Tmp = MRI.createVirtualRegister(RC);
22097 // Since FP is only updated here but NOT referenced, it's treated as GPR.
22098 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22099 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
22100 unsigned SP = RegInfo->getStackRegister();
22102 MachineInstrBuilder MIB;
22104 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22105 const int64_t SPOffset = 2 * PVT.getStoreSize();
22107 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
22108 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
22111 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
22112 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
22113 MIB.addOperand(MI->getOperand(i));
22114 MIB.setMemRefs(MMOBegin, MMOEnd);
22116 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
22117 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22118 if (i == X86::AddrDisp)
22119 MIB.addDisp(MI->getOperand(i), LabelOffset);
22121 MIB.addOperand(MI->getOperand(i));
22123 MIB.setMemRefs(MMOBegin, MMOEnd);
22125 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
22126 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22127 if (i == X86::AddrDisp)
22128 MIB.addDisp(MI->getOperand(i), SPOffset);
22130 MIB.addOperand(MI->getOperand(i));
22132 MIB.setMemRefs(MMOBegin, MMOEnd);
22134 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
22136 MI->eraseFromParent();
22140 // Replace 213-type (isel default) FMA3 instructions with 231-type for
22141 // accumulator loops. Writing back to the accumulator allows the coalescer
22142 // to remove extra copies in the loop.
22143 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
22144 MachineBasicBlock *
22145 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
22146 MachineBasicBlock *MBB) const {
22147 MachineOperand &AddendOp = MI->getOperand(3);
22149 // Bail out early if the addend isn't a register - we can't switch these.
22150 if (!AddendOp.isReg())
22153 MachineFunction &MF = *MBB->getParent();
22154 MachineRegisterInfo &MRI = MF.getRegInfo();
22156 // Check whether the addend is defined by a PHI:
22157 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
22158 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
22159 if (!AddendDef.isPHI())
22162 // Look for the following pattern:
22164 // %addend = phi [%entry, 0], [%loop, %result]
22166 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
22170 // %addend = phi [%entry, 0], [%loop, %result]
22172 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
22174 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
22175 assert(AddendDef.getOperand(i).isReg());
22176 MachineOperand PHISrcOp = AddendDef.getOperand(i);
22177 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
22178 if (&PHISrcInst == MI) {
22179 // Found a matching instruction.
22180 unsigned NewFMAOpc = 0;
22181 switch (MI->getOpcode()) {
22182 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
22183 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
22184 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
22185 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
22186 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
22187 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
22188 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
22189 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
22190 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
22191 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
22192 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
22193 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
22194 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
22195 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
22196 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
22197 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
22198 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
22199 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
22200 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
22201 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
22203 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
22204 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
22205 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
22206 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
22207 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
22208 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
22209 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
22210 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
22211 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
22212 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
22213 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
22214 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
22215 default: llvm_unreachable("Unrecognized FMA variant.");
22218 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22219 MachineInstrBuilder MIB =
22220 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
22221 .addOperand(MI->getOperand(0))
22222 .addOperand(MI->getOperand(3))
22223 .addOperand(MI->getOperand(2))
22224 .addOperand(MI->getOperand(1));
22225 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
22226 MI->eraseFromParent();
22233 MachineBasicBlock *
22234 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
22235 MachineBasicBlock *BB) const {
22236 switch (MI->getOpcode()) {
22237 default: llvm_unreachable("Unexpected instr type to insert");
22238 case X86::TAILJMPd64:
22239 case X86::TAILJMPr64:
22240 case X86::TAILJMPm64:
22241 case X86::TAILJMPd64_REX:
22242 case X86::TAILJMPr64_REX:
22243 case X86::TAILJMPm64_REX:
22244 llvm_unreachable("TAILJMP64 would not be touched here.");
22245 case X86::TCRETURNdi64:
22246 case X86::TCRETURNri64:
22247 case X86::TCRETURNmi64:
22249 case X86::WIN_ALLOCA:
22250 return EmitLoweredWinAlloca(MI, BB);
22251 case X86::CATCHRET:
22252 return EmitLoweredCatchRet(MI, BB);
22253 case X86::CATCHPAD:
22254 return EmitLoweredCatchPad(MI, BB);
22255 case X86::SEG_ALLOCA_32:
22256 case X86::SEG_ALLOCA_64:
22257 return EmitLoweredSegAlloca(MI, BB);
22258 case X86::TLSCall_32:
22259 case X86::TLSCall_64:
22260 return EmitLoweredTLSCall(MI, BB);
22261 case X86::CMOV_FR32:
22262 case X86::CMOV_FR64:
22263 case X86::CMOV_FR128:
22264 case X86::CMOV_GR8:
22265 case X86::CMOV_GR16:
22266 case X86::CMOV_GR32:
22267 case X86::CMOV_RFP32:
22268 case X86::CMOV_RFP64:
22269 case X86::CMOV_RFP80:
22270 case X86::CMOV_V2F64:
22271 case X86::CMOV_V2I64:
22272 case X86::CMOV_V4F32:
22273 case X86::CMOV_V4F64:
22274 case X86::CMOV_V4I64:
22275 case X86::CMOV_V16F32:
22276 case X86::CMOV_V8F32:
22277 case X86::CMOV_V8F64:
22278 case X86::CMOV_V8I64:
22279 case X86::CMOV_V8I1:
22280 case X86::CMOV_V16I1:
22281 case X86::CMOV_V32I1:
22282 case X86::CMOV_V64I1:
22283 return EmitLoweredSelect(MI, BB);
22285 case X86::RELEASE_FADD32mr:
22286 case X86::RELEASE_FADD64mr:
22287 return EmitLoweredAtomicFP(MI, BB);
22289 case X86::FP32_TO_INT16_IN_MEM:
22290 case X86::FP32_TO_INT32_IN_MEM:
22291 case X86::FP32_TO_INT64_IN_MEM:
22292 case X86::FP64_TO_INT16_IN_MEM:
22293 case X86::FP64_TO_INT32_IN_MEM:
22294 case X86::FP64_TO_INT64_IN_MEM:
22295 case X86::FP80_TO_INT16_IN_MEM:
22296 case X86::FP80_TO_INT32_IN_MEM:
22297 case X86::FP80_TO_INT64_IN_MEM: {
22298 MachineFunction *F = BB->getParent();
22299 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22300 DebugLoc DL = MI->getDebugLoc();
22302 // Change the floating point control register to use "round towards zero"
22303 // mode when truncating to an integer value.
22304 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
22305 addFrameReference(BuildMI(*BB, MI, DL,
22306 TII->get(X86::FNSTCW16m)), CWFrameIdx);
22308 // Load the old value of the high byte of the control word...
22310 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
22311 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
22314 // Set the high part to be round to zero...
22315 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
22318 // Reload the modified control word now...
22319 addFrameReference(BuildMI(*BB, MI, DL,
22320 TII->get(X86::FLDCW16m)), CWFrameIdx);
22322 // Restore the memory image of control word to original value
22323 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
22326 // Get the X86 opcode to use.
22328 switch (MI->getOpcode()) {
22329 default: llvm_unreachable("illegal opcode!");
22330 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
22331 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22332 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22333 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22334 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22335 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22336 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22337 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22338 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22342 MachineOperand &Op = MI->getOperand(0);
22344 AM.BaseType = X86AddressMode::RegBase;
22345 AM.Base.Reg = Op.getReg();
22347 AM.BaseType = X86AddressMode::FrameIndexBase;
22348 AM.Base.FrameIndex = Op.getIndex();
22350 Op = MI->getOperand(1);
22352 AM.Scale = Op.getImm();
22353 Op = MI->getOperand(2);
22355 AM.IndexReg = Op.getImm();
22356 Op = MI->getOperand(3);
22357 if (Op.isGlobal()) {
22358 AM.GV = Op.getGlobal();
22360 AM.Disp = Op.getImm();
22362 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22363 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22365 // Reload the original control word now.
22366 addFrameReference(BuildMI(*BB, MI, DL,
22367 TII->get(X86::FLDCW16m)), CWFrameIdx);
22369 MI->eraseFromParent(); // The pseudo instruction is gone now.
22372 // String/text processing lowering.
22373 case X86::PCMPISTRM128REG:
22374 case X86::VPCMPISTRM128REG:
22375 case X86::PCMPISTRM128MEM:
22376 case X86::VPCMPISTRM128MEM:
22377 case X86::PCMPESTRM128REG:
22378 case X86::VPCMPESTRM128REG:
22379 case X86::PCMPESTRM128MEM:
22380 case X86::VPCMPESTRM128MEM:
22381 assert(Subtarget->hasSSE42() &&
22382 "Target must have SSE4.2 or AVX features enabled");
22383 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22385 // String/text processing lowering.
22386 case X86::PCMPISTRIREG:
22387 case X86::VPCMPISTRIREG:
22388 case X86::PCMPISTRIMEM:
22389 case X86::VPCMPISTRIMEM:
22390 case X86::PCMPESTRIREG:
22391 case X86::VPCMPESTRIREG:
22392 case X86::PCMPESTRIMEM:
22393 case X86::VPCMPESTRIMEM:
22394 assert(Subtarget->hasSSE42() &&
22395 "Target must have SSE4.2 or AVX features enabled");
22396 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22398 // Thread synchronization.
22400 return EmitMonitor(MI, BB, Subtarget);
22404 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22406 case X86::VASTART_SAVE_XMM_REGS:
22407 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22409 case X86::VAARG_64:
22410 return EmitVAARG64WithCustomInserter(MI, BB);
22412 case X86::EH_SjLj_SetJmp32:
22413 case X86::EH_SjLj_SetJmp64:
22414 return emitEHSjLjSetJmp(MI, BB);
22416 case X86::EH_SjLj_LongJmp32:
22417 case X86::EH_SjLj_LongJmp64:
22418 return emitEHSjLjLongJmp(MI, BB);
22420 case TargetOpcode::STATEPOINT:
22421 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22422 // this point in the process. We diverge later.
22423 return emitPatchPoint(MI, BB);
22425 case TargetOpcode::STACKMAP:
22426 case TargetOpcode::PATCHPOINT:
22427 return emitPatchPoint(MI, BB);
22429 case X86::VFMADDPDr213r:
22430 case X86::VFMADDPSr213r:
22431 case X86::VFMADDSDr213r:
22432 case X86::VFMADDSSr213r:
22433 case X86::VFMSUBPDr213r:
22434 case X86::VFMSUBPSr213r:
22435 case X86::VFMSUBSDr213r:
22436 case X86::VFMSUBSSr213r:
22437 case X86::VFNMADDPDr213r:
22438 case X86::VFNMADDPSr213r:
22439 case X86::VFNMADDSDr213r:
22440 case X86::VFNMADDSSr213r:
22441 case X86::VFNMSUBPDr213r:
22442 case X86::VFNMSUBPSr213r:
22443 case X86::VFNMSUBSDr213r:
22444 case X86::VFNMSUBSSr213r:
22445 case X86::VFMADDSUBPDr213r:
22446 case X86::VFMADDSUBPSr213r:
22447 case X86::VFMSUBADDPDr213r:
22448 case X86::VFMSUBADDPSr213r:
22449 case X86::VFMADDPDr213rY:
22450 case X86::VFMADDPSr213rY:
22451 case X86::VFMSUBPDr213rY:
22452 case X86::VFMSUBPSr213rY:
22453 case X86::VFNMADDPDr213rY:
22454 case X86::VFNMADDPSr213rY:
22455 case X86::VFNMSUBPDr213rY:
22456 case X86::VFNMSUBPSr213rY:
22457 case X86::VFMADDSUBPDr213rY:
22458 case X86::VFMADDSUBPSr213rY:
22459 case X86::VFMSUBADDPDr213rY:
22460 case X86::VFMSUBADDPSr213rY:
22461 return emitFMA3Instr(MI, BB);
22465 //===----------------------------------------------------------------------===//
22466 // X86 Optimization Hooks
22467 //===----------------------------------------------------------------------===//
22469 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22472 const SelectionDAG &DAG,
22473 unsigned Depth) const {
22474 unsigned BitWidth = KnownZero.getBitWidth();
22475 unsigned Opc = Op.getOpcode();
22476 assert((Opc >= ISD::BUILTIN_OP_END ||
22477 Opc == ISD::INTRINSIC_WO_CHAIN ||
22478 Opc == ISD::INTRINSIC_W_CHAIN ||
22479 Opc == ISD::INTRINSIC_VOID) &&
22480 "Should use MaskedValueIsZero if you don't know whether Op"
22481 " is a target node!");
22483 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22497 // These nodes' second result is a boolean.
22498 if (Op.getResNo() == 0)
22501 case X86ISD::SETCC:
22502 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22504 case ISD::INTRINSIC_WO_CHAIN: {
22505 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22506 unsigned NumLoBits = 0;
22509 case Intrinsic::x86_sse_movmsk_ps:
22510 case Intrinsic::x86_avx_movmsk_ps_256:
22511 case Intrinsic::x86_sse2_movmsk_pd:
22512 case Intrinsic::x86_avx_movmsk_pd_256:
22513 case Intrinsic::x86_mmx_pmovmskb:
22514 case Intrinsic::x86_sse2_pmovmskb_128:
22515 case Intrinsic::x86_avx2_pmovmskb: {
22516 // High bits of movmskp{s|d}, pmovmskb are known zero.
22518 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22519 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22520 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22521 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22522 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22523 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22524 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22525 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22527 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22536 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22538 const SelectionDAG &,
22539 unsigned Depth) const {
22540 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22541 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22542 return Op.getValueType().getScalarSizeInBits();
22548 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22549 /// node is a GlobalAddress + offset.
22550 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22551 const GlobalValue* &GA,
22552 int64_t &Offset) const {
22553 if (N->getOpcode() == X86ISD::Wrapper) {
22554 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22555 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22556 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22560 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22563 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22564 /// same as extracting the high 128-bit part of 256-bit vector and then
22565 /// inserting the result into the low part of a new 256-bit vector
22566 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22567 EVT VT = SVOp->getValueType(0);
22568 unsigned NumElems = VT.getVectorNumElements();
22570 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22571 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22572 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22573 SVOp->getMaskElt(j) >= 0)
22579 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22580 /// same as extracting the low 128-bit part of 256-bit vector and then
22581 /// inserting the result into the high part of a new 256-bit vector
22582 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22583 EVT VT = SVOp->getValueType(0);
22584 unsigned NumElems = VT.getVectorNumElements();
22586 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22587 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22588 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22589 SVOp->getMaskElt(j) >= 0)
22595 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22596 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22597 TargetLowering::DAGCombinerInfo &DCI,
22598 const X86Subtarget* Subtarget) {
22600 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22601 SDValue V1 = SVOp->getOperand(0);
22602 SDValue V2 = SVOp->getOperand(1);
22603 MVT VT = SVOp->getSimpleValueType(0);
22604 unsigned NumElems = VT.getVectorNumElements();
22606 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22607 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22611 // V UNDEF BUILD_VECTOR UNDEF
22613 // CONCAT_VECTOR CONCAT_VECTOR
22616 // RESULT: V + zero extended
22618 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22619 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22620 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22623 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22626 // To match the shuffle mask, the first half of the mask should
22627 // be exactly the first vector, and all the rest a splat with the
22628 // first element of the second one.
22629 for (unsigned i = 0; i != NumElems/2; ++i)
22630 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22631 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22634 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22635 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22636 if (Ld->hasNUsesOfValue(1, 0)) {
22637 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22638 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22640 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22642 Ld->getPointerInfo(),
22643 Ld->getAlignment(),
22644 false/*isVolatile*/, true/*ReadMem*/,
22645 false/*WriteMem*/);
22647 // Make sure the newly-created LOAD is in the same position as Ld in
22648 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22649 // and update uses of Ld's output chain to use the TokenFactor.
22650 if (Ld->hasAnyUseOfValue(1)) {
22651 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22652 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22653 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22654 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22655 SDValue(ResNode.getNode(), 1));
22658 return DAG.getBitcast(VT, ResNode);
22662 // Emit a zeroed vector and insert the desired subvector on its
22664 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22665 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22666 return DCI.CombineTo(N, InsV);
22669 //===--------------------------------------------------------------------===//
22670 // Combine some shuffles into subvector extracts and inserts:
22673 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22674 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22675 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22676 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22677 return DCI.CombineTo(N, InsV);
22680 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22681 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22682 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22683 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22684 return DCI.CombineTo(N, InsV);
22690 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22693 /// This is the leaf of the recursive combinine below. When we have found some
22694 /// chain of single-use x86 shuffle instructions and accumulated the combined
22695 /// shuffle mask represented by them, this will try to pattern match that mask
22696 /// into either a single instruction if there is a special purpose instruction
22697 /// for this operation, or into a PSHUFB instruction which is a fully general
22698 /// instruction but should only be used to replace chains over a certain depth.
22699 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22700 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22701 TargetLowering::DAGCombinerInfo &DCI,
22702 const X86Subtarget *Subtarget) {
22703 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22705 // Find the operand that enters the chain. Note that multiple uses are OK
22706 // here, we're not going to remove the operand we find.
22707 SDValue Input = Op.getOperand(0);
22708 while (Input.getOpcode() == ISD::BITCAST)
22709 Input = Input.getOperand(0);
22711 MVT VT = Input.getSimpleValueType();
22712 MVT RootVT = Root.getSimpleValueType();
22715 if (Mask.size() == 1) {
22716 int Index = Mask[0];
22717 assert((Index >= 0 || Index == SM_SentinelUndef ||
22718 Index == SM_SentinelZero) &&
22719 "Invalid shuffle index found!");
22721 // We may end up with an accumulated mask of size 1 as a result of
22722 // widening of shuffle operands (see function canWidenShuffleElements).
22723 // If the only shuffle index is equal to SM_SentinelZero then propagate
22724 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22725 // mask, and therefore the entire chain of shuffles can be folded away.
22726 if (Index == SM_SentinelZero)
22727 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22729 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22734 // Use the float domain if the operand type is a floating point type.
22735 bool FloatDomain = VT.isFloatingPoint();
22737 // For floating point shuffles, we don't have free copies in the shuffle
22738 // instructions or the ability to load as part of the instruction, so
22739 // canonicalize their shuffles to UNPCK or MOV variants.
22741 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22742 // vectors because it can have a load folded into it that UNPCK cannot. This
22743 // doesn't preclude something switching to the shorter encoding post-RA.
22745 // FIXME: Should teach these routines about AVX vector widths.
22746 if (FloatDomain && VT.is128BitVector()) {
22747 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22748 bool Lo = Mask.equals({0, 0});
22751 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22752 // is no slower than UNPCKLPD but has the option to fold the input operand
22753 // into even an unaligned memory load.
22754 if (Lo && Subtarget->hasSSE3()) {
22755 Shuffle = X86ISD::MOVDDUP;
22756 ShuffleVT = MVT::v2f64;
22758 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22759 // than the UNPCK variants.
22760 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22761 ShuffleVT = MVT::v4f32;
22763 if (Depth == 1 && Root->getOpcode() == Shuffle)
22764 return false; // Nothing to do!
22765 Op = DAG.getBitcast(ShuffleVT, Input);
22766 DCI.AddToWorklist(Op.getNode());
22767 if (Shuffle == X86ISD::MOVDDUP)
22768 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22770 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22771 DCI.AddToWorklist(Op.getNode());
22772 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22776 if (Subtarget->hasSSE3() &&
22777 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22778 bool Lo = Mask.equals({0, 0, 2, 2});
22779 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22780 MVT ShuffleVT = MVT::v4f32;
22781 if (Depth == 1 && Root->getOpcode() == Shuffle)
22782 return false; // Nothing to do!
22783 Op = DAG.getBitcast(ShuffleVT, Input);
22784 DCI.AddToWorklist(Op.getNode());
22785 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22786 DCI.AddToWorklist(Op.getNode());
22787 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22791 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22792 bool Lo = Mask.equals({0, 0, 1, 1});
22793 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22794 MVT ShuffleVT = MVT::v4f32;
22795 if (Depth == 1 && Root->getOpcode() == Shuffle)
22796 return false; // Nothing to do!
22797 Op = DAG.getBitcast(ShuffleVT, Input);
22798 DCI.AddToWorklist(Op.getNode());
22799 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22800 DCI.AddToWorklist(Op.getNode());
22801 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22807 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22808 // variants as none of these have single-instruction variants that are
22809 // superior to the UNPCK formulation.
22810 if (!FloatDomain && VT.is128BitVector() &&
22811 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22812 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22813 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22815 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22816 bool Lo = Mask[0] == 0;
22817 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22818 if (Depth == 1 && Root->getOpcode() == Shuffle)
22819 return false; // Nothing to do!
22821 switch (Mask.size()) {
22823 ShuffleVT = MVT::v8i16;
22826 ShuffleVT = MVT::v16i8;
22829 llvm_unreachable("Impossible mask size!");
22831 Op = DAG.getBitcast(ShuffleVT, Input);
22832 DCI.AddToWorklist(Op.getNode());
22833 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22834 DCI.AddToWorklist(Op.getNode());
22835 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22840 // Don't try to re-form single instruction chains under any circumstances now
22841 // that we've done encoding canonicalization for them.
22845 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22846 // can replace them with a single PSHUFB instruction profitably. Intel's
22847 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22848 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22849 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22850 SmallVector<SDValue, 16> PSHUFBMask;
22851 int NumBytes = VT.getSizeInBits() / 8;
22852 int Ratio = NumBytes / Mask.size();
22853 for (int i = 0; i < NumBytes; ++i) {
22854 if (Mask[i / Ratio] == SM_SentinelUndef) {
22855 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22858 int M = Mask[i / Ratio] != SM_SentinelZero
22859 ? Ratio * Mask[i / Ratio] + i % Ratio
22861 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22863 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22864 Op = DAG.getBitcast(ByteVT, Input);
22865 DCI.AddToWorklist(Op.getNode());
22866 SDValue PSHUFBMaskOp =
22867 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22868 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22869 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22870 DCI.AddToWorklist(Op.getNode());
22871 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22876 // Failed to find any combines.
22880 /// \brief Fully generic combining of x86 shuffle instructions.
22882 /// This should be the last combine run over the x86 shuffle instructions. Once
22883 /// they have been fully optimized, this will recursively consider all chains
22884 /// of single-use shuffle instructions, build a generic model of the cumulative
22885 /// shuffle operation, and check for simpler instructions which implement this
22886 /// operation. We use this primarily for two purposes:
22888 /// 1) Collapse generic shuffles to specialized single instructions when
22889 /// equivalent. In most cases, this is just an encoding size win, but
22890 /// sometimes we will collapse multiple generic shuffles into a single
22891 /// special-purpose shuffle.
22892 /// 2) Look for sequences of shuffle instructions with 3 or more total
22893 /// instructions, and replace them with the slightly more expensive SSSE3
22894 /// PSHUFB instruction if available. We do this as the last combining step
22895 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22896 /// a suitable short sequence of other instructions. The PHUFB will either
22897 /// use a register or have to read from memory and so is slightly (but only
22898 /// slightly) more expensive than the other shuffle instructions.
22900 /// Because this is inherently a quadratic operation (for each shuffle in
22901 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22902 /// This should never be an issue in practice as the shuffle lowering doesn't
22903 /// produce sequences of more than 8 instructions.
22905 /// FIXME: We will currently miss some cases where the redundant shuffling
22906 /// would simplify under the threshold for PSHUFB formation because of
22907 /// combine-ordering. To fix this, we should do the redundant instruction
22908 /// combining in this recursive walk.
22909 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22910 ArrayRef<int> RootMask,
22911 int Depth, bool HasPSHUFB,
22913 TargetLowering::DAGCombinerInfo &DCI,
22914 const X86Subtarget *Subtarget) {
22915 // Bound the depth of our recursive combine because this is ultimately
22916 // quadratic in nature.
22920 // Directly rip through bitcasts to find the underlying operand.
22921 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22922 Op = Op.getOperand(0);
22924 MVT VT = Op.getSimpleValueType();
22925 if (!VT.isVector())
22926 return false; // Bail if we hit a non-vector.
22928 assert(Root.getSimpleValueType().isVector() &&
22929 "Shuffles operate on vector types!");
22930 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22931 "Can only combine shuffles of the same vector register size.");
22933 if (!isTargetShuffle(Op.getOpcode()))
22935 SmallVector<int, 16> OpMask;
22937 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22938 // We only can combine unary shuffles which we can decode the mask for.
22939 if (!HaveMask || !IsUnary)
22942 assert(VT.getVectorNumElements() == OpMask.size() &&
22943 "Different mask size from vector size!");
22944 assert(((RootMask.size() > OpMask.size() &&
22945 RootMask.size() % OpMask.size() == 0) ||
22946 (OpMask.size() > RootMask.size() &&
22947 OpMask.size() % RootMask.size() == 0) ||
22948 OpMask.size() == RootMask.size()) &&
22949 "The smaller number of elements must divide the larger.");
22950 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22951 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22952 assert(((RootRatio == 1 && OpRatio == 1) ||
22953 (RootRatio == 1) != (OpRatio == 1)) &&
22954 "Must not have a ratio for both incoming and op masks!");
22956 SmallVector<int, 16> Mask;
22957 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22959 // Merge this shuffle operation's mask into our accumulated mask. Note that
22960 // this shuffle's mask will be the first applied to the input, followed by the
22961 // root mask to get us all the way to the root value arrangement. The reason
22962 // for this order is that we are recursing up the operation chain.
22963 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22964 int RootIdx = i / RootRatio;
22965 if (RootMask[RootIdx] < 0) {
22966 // This is a zero or undef lane, we're done.
22967 Mask.push_back(RootMask[RootIdx]);
22971 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22972 int OpIdx = RootMaskedIdx / OpRatio;
22973 if (OpMask[OpIdx] < 0) {
22974 // The incoming lanes are zero or undef, it doesn't matter which ones we
22976 Mask.push_back(OpMask[OpIdx]);
22980 // Ok, we have non-zero lanes, map them through.
22981 Mask.push_back(OpMask[OpIdx] * OpRatio +
22982 RootMaskedIdx % OpRatio);
22985 // See if we can recurse into the operand to combine more things.
22986 switch (Op.getOpcode()) {
22987 case X86ISD::PSHUFB:
22989 case X86ISD::PSHUFD:
22990 case X86ISD::PSHUFHW:
22991 case X86ISD::PSHUFLW:
22992 if (Op.getOperand(0).hasOneUse() &&
22993 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22994 HasPSHUFB, DAG, DCI, Subtarget))
22998 case X86ISD::UNPCKL:
22999 case X86ISD::UNPCKH:
23000 assert(Op.getOperand(0) == Op.getOperand(1) &&
23001 "We only combine unary shuffles!");
23002 // We can't check for single use, we have to check that this shuffle is the
23004 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
23005 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23006 HasPSHUFB, DAG, DCI, Subtarget))
23011 // Minor canonicalization of the accumulated shuffle mask to make it easier
23012 // to match below. All this does is detect masks with squential pairs of
23013 // elements, and shrink them to the half-width mask. It does this in a loop
23014 // so it will reduce the size of the mask to the minimal width mask which
23015 // performs an equivalent shuffle.
23016 SmallVector<int, 16> WidenedMask;
23017 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
23018 Mask = std::move(WidenedMask);
23019 WidenedMask.clear();
23022 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
23026 /// \brief Get the PSHUF-style mask from PSHUF node.
23028 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
23029 /// PSHUF-style masks that can be reused with such instructions.
23030 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
23031 MVT VT = N.getSimpleValueType();
23032 SmallVector<int, 4> Mask;
23034 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
23038 // If we have more than 128-bits, only the low 128-bits of shuffle mask
23039 // matter. Check that the upper masks are repeats and remove them.
23040 if (VT.getSizeInBits() > 128) {
23041 int LaneElts = 128 / VT.getScalarSizeInBits();
23043 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
23044 for (int j = 0; j < LaneElts; ++j)
23045 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
23046 "Mask doesn't repeat in high 128-bit lanes!");
23048 Mask.resize(LaneElts);
23051 switch (N.getOpcode()) {
23052 case X86ISD::PSHUFD:
23054 case X86ISD::PSHUFLW:
23057 case X86ISD::PSHUFHW:
23058 Mask.erase(Mask.begin(), Mask.begin() + 4);
23059 for (int &M : Mask)
23063 llvm_unreachable("No valid shuffle instruction found!");
23067 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
23069 /// We walk up the chain and look for a combinable shuffle, skipping over
23070 /// shuffles that we could hoist this shuffle's transformation past without
23071 /// altering anything.
23073 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
23075 TargetLowering::DAGCombinerInfo &DCI) {
23076 assert(N.getOpcode() == X86ISD::PSHUFD &&
23077 "Called with something other than an x86 128-bit half shuffle!");
23080 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
23081 // of the shuffles in the chain so that we can form a fresh chain to replace
23083 SmallVector<SDValue, 8> Chain;
23084 SDValue V = N.getOperand(0);
23085 for (; V.hasOneUse(); V = V.getOperand(0)) {
23086 switch (V.getOpcode()) {
23088 return SDValue(); // Nothing combined!
23091 // Skip bitcasts as we always know the type for the target specific
23095 case X86ISD::PSHUFD:
23096 // Found another dword shuffle.
23099 case X86ISD::PSHUFLW:
23100 // Check that the low words (being shuffled) are the identity in the
23101 // dword shuffle, and the high words are self-contained.
23102 if (Mask[0] != 0 || Mask[1] != 1 ||
23103 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
23106 Chain.push_back(V);
23109 case X86ISD::PSHUFHW:
23110 // Check that the high words (being shuffled) are the identity in the
23111 // dword shuffle, and the low words are self-contained.
23112 if (Mask[2] != 2 || Mask[3] != 3 ||
23113 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
23116 Chain.push_back(V);
23119 case X86ISD::UNPCKL:
23120 case X86ISD::UNPCKH:
23121 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
23122 // shuffle into a preceding word shuffle.
23123 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
23124 V.getSimpleValueType().getVectorElementType() != MVT::i16)
23127 // Search for a half-shuffle which we can combine with.
23128 unsigned CombineOp =
23129 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
23130 if (V.getOperand(0) != V.getOperand(1) ||
23131 !V->isOnlyUserOf(V.getOperand(0).getNode()))
23133 Chain.push_back(V);
23134 V = V.getOperand(0);
23136 switch (V.getOpcode()) {
23138 return SDValue(); // Nothing to combine.
23140 case X86ISD::PSHUFLW:
23141 case X86ISD::PSHUFHW:
23142 if (V.getOpcode() == CombineOp)
23145 Chain.push_back(V);
23149 V = V.getOperand(0);
23153 } while (V.hasOneUse());
23156 // Break out of the loop if we break out of the switch.
23160 if (!V.hasOneUse())
23161 // We fell out of the loop without finding a viable combining instruction.
23164 // Merge this node's mask and our incoming mask.
23165 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23166 for (int &M : Mask)
23168 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
23169 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23171 // Rebuild the chain around this new shuffle.
23172 while (!Chain.empty()) {
23173 SDValue W = Chain.pop_back_val();
23175 if (V.getValueType() != W.getOperand(0).getValueType())
23176 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
23178 switch (W.getOpcode()) {
23180 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
23182 case X86ISD::UNPCKL:
23183 case X86ISD::UNPCKH:
23184 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
23187 case X86ISD::PSHUFD:
23188 case X86ISD::PSHUFLW:
23189 case X86ISD::PSHUFHW:
23190 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
23194 if (V.getValueType() != N.getValueType())
23195 V = DAG.getBitcast(N.getValueType(), V);
23197 // Return the new chain to replace N.
23201 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
23204 /// We walk up the chain, skipping shuffles of the other half and looking
23205 /// through shuffles which switch halves trying to find a shuffle of the same
23206 /// pair of dwords.
23207 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
23209 TargetLowering::DAGCombinerInfo &DCI) {
23211 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
23212 "Called with something other than an x86 128-bit half shuffle!");
23214 unsigned CombineOpcode = N.getOpcode();
23216 // Walk up a single-use chain looking for a combinable shuffle.
23217 SDValue V = N.getOperand(0);
23218 for (; V.hasOneUse(); V = V.getOperand(0)) {
23219 switch (V.getOpcode()) {
23221 return false; // Nothing combined!
23224 // Skip bitcasts as we always know the type for the target specific
23228 case X86ISD::PSHUFLW:
23229 case X86ISD::PSHUFHW:
23230 if (V.getOpcode() == CombineOpcode)
23233 // Other-half shuffles are no-ops.
23236 // Break out of the loop if we break out of the switch.
23240 if (!V.hasOneUse())
23241 // We fell out of the loop without finding a viable combining instruction.
23244 // Combine away the bottom node as its shuffle will be accumulated into
23245 // a preceding shuffle.
23246 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23248 // Record the old value.
23251 // Merge this node's mask and our incoming mask (adjusted to account for all
23252 // the pshufd instructions encountered).
23253 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23254 for (int &M : Mask)
23256 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
23257 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23259 // Check that the shuffles didn't cancel each other out. If not, we need to
23260 // combine to the new one.
23262 // Replace the combinable shuffle with the combined one, updating all users
23263 // so that we re-evaluate the chain here.
23264 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
23269 /// \brief Try to combine x86 target specific shuffles.
23270 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
23271 TargetLowering::DAGCombinerInfo &DCI,
23272 const X86Subtarget *Subtarget) {
23274 MVT VT = N.getSimpleValueType();
23275 SmallVector<int, 4> Mask;
23277 switch (N.getOpcode()) {
23278 case X86ISD::PSHUFD:
23279 case X86ISD::PSHUFLW:
23280 case X86ISD::PSHUFHW:
23281 Mask = getPSHUFShuffleMask(N);
23282 assert(Mask.size() == 4);
23284 case X86ISD::UNPCKL: {
23285 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
23286 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
23287 // moves upper half elements into the lower half part. For example:
23289 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
23291 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
23293 // will be combined to:
23295 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
23297 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
23298 // happen due to advanced instructions.
23299 if (!VT.is128BitVector())
23302 auto Op0 = N.getOperand(0);
23303 auto Op1 = N.getOperand(1);
23304 if (Op0.getOpcode() == ISD::UNDEF &&
23305 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
23306 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
23308 unsigned NumElts = VT.getVectorNumElements();
23309 SmallVector<int, 8> ExpectedMask(NumElts, -1);
23310 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
23313 auto ShufOp = Op1.getOperand(0);
23314 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
23315 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
23323 // Nuke no-op shuffles that show up after combining.
23324 if (isNoopShuffleMask(Mask))
23325 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23327 // Look for simplifications involving one or two shuffle instructions.
23328 SDValue V = N.getOperand(0);
23329 switch (N.getOpcode()) {
23332 case X86ISD::PSHUFLW:
23333 case X86ISD::PSHUFHW:
23334 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
23336 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
23337 return SDValue(); // We combined away this shuffle, so we're done.
23339 // See if this reduces to a PSHUFD which is no more expensive and can
23340 // combine with more operations. Note that it has to at least flip the
23341 // dwords as otherwise it would have been removed as a no-op.
23342 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
23343 int DMask[] = {0, 1, 2, 3};
23344 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
23345 DMask[DOffset + 0] = DOffset + 1;
23346 DMask[DOffset + 1] = DOffset + 0;
23347 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
23348 V = DAG.getBitcast(DVT, V);
23349 DCI.AddToWorklist(V.getNode());
23350 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
23351 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
23352 DCI.AddToWorklist(V.getNode());
23353 return DAG.getBitcast(VT, V);
23356 // Look for shuffle patterns which can be implemented as a single unpack.
23357 // FIXME: This doesn't handle the location of the PSHUFD generically, and
23358 // only works when we have a PSHUFD followed by two half-shuffles.
23359 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
23360 (V.getOpcode() == X86ISD::PSHUFLW ||
23361 V.getOpcode() == X86ISD::PSHUFHW) &&
23362 V.getOpcode() != N.getOpcode() &&
23364 SDValue D = V.getOperand(0);
23365 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
23366 D = D.getOperand(0);
23367 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
23368 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23369 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
23370 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23371 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23373 for (int i = 0; i < 4; ++i) {
23374 WordMask[i + NOffset] = Mask[i] + NOffset;
23375 WordMask[i + VOffset] = VMask[i] + VOffset;
23377 // Map the word mask through the DWord mask.
23379 for (int i = 0; i < 8; ++i)
23380 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23381 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23382 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23383 // We can replace all three shuffles with an unpack.
23384 V = DAG.getBitcast(VT, D.getOperand(0));
23385 DCI.AddToWorklist(V.getNode());
23386 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23395 case X86ISD::PSHUFD:
23396 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23405 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23407 /// We combine this directly on the abstract vector shuffle nodes so it is
23408 /// easier to generically match. We also insert dummy vector shuffle nodes for
23409 /// the operands which explicitly discard the lanes which are unused by this
23410 /// operation to try to flow through the rest of the combiner the fact that
23411 /// they're unused.
23412 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23414 EVT VT = N->getValueType(0);
23416 // We only handle target-independent shuffles.
23417 // FIXME: It would be easy and harmless to use the target shuffle mask
23418 // extraction tool to support more.
23419 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23422 auto *SVN = cast<ShuffleVectorSDNode>(N);
23423 SmallVector<int, 8> Mask;
23424 for (int M : SVN->getMask())
23427 SDValue V1 = N->getOperand(0);
23428 SDValue V2 = N->getOperand(1);
23430 // We require the first shuffle operand to be the FSUB node, and the second to
23431 // be the FADD node.
23432 if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
23433 ShuffleVectorSDNode::commuteMask(Mask);
23435 } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
23438 // If there are other uses of these operations we can't fold them.
23439 if (!V1->hasOneUse() || !V2->hasOneUse())
23442 // Ensure that both operations have the same operands. Note that we can
23443 // commute the FADD operands.
23444 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23445 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23446 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23449 // We're looking for blends between FADD and FSUB nodes. We insist on these
23450 // nodes being lined up in a specific expected pattern.
23451 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23452 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23453 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23456 // Only specific types are legal at this point, assert so we notice if and
23457 // when these change.
23458 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23459 VT == MVT::v4f64) &&
23460 "Unknown vector type encountered!");
23462 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23465 /// PerformShuffleCombine - Performs several different shuffle combines.
23466 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23467 TargetLowering::DAGCombinerInfo &DCI,
23468 const X86Subtarget *Subtarget) {
23470 SDValue N0 = N->getOperand(0);
23471 SDValue N1 = N->getOperand(1);
23472 EVT VT = N->getValueType(0);
23474 // Don't create instructions with illegal types after legalize types has run.
23475 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23476 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23479 // If we have legalized the vector types, look for blends of FADD and FSUB
23480 // nodes that we can fuse into an ADDSUB node.
23481 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23482 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23485 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23486 if (TLI.isTypeLegal(VT) && Subtarget->hasFp256() && VT.is256BitVector() &&
23487 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23488 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23490 // During Type Legalization, when promoting illegal vector types,
23491 // the backend might introduce new shuffle dag nodes and bitcasts.
23493 // This code performs the following transformation:
23494 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23495 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23497 // We do this only if both the bitcast and the BINOP dag nodes have
23498 // one use. Also, perform this transformation only if the new binary
23499 // operation is legal. This is to avoid introducing dag nodes that
23500 // potentially need to be further expanded (or custom lowered) into a
23501 // less optimal sequence of dag nodes.
23502 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23503 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23504 N0.getOpcode() == ISD::BITCAST) {
23505 SDValue BC0 = N0.getOperand(0);
23506 EVT SVT = BC0.getValueType();
23507 unsigned Opcode = BC0.getOpcode();
23508 unsigned NumElts = VT.getVectorNumElements();
23510 if (BC0.hasOneUse() && SVT.isVector() &&
23511 SVT.getVectorNumElements() * 2 == NumElts &&
23512 TLI.isOperationLegal(Opcode, VT)) {
23513 bool CanFold = false;
23525 unsigned SVTNumElts = SVT.getVectorNumElements();
23526 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23527 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23528 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23529 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23530 CanFold = SVOp->getMaskElt(i) < 0;
23533 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23534 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23535 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23536 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23541 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23542 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23543 // consecutive, non-overlapping, and in the right order.
23544 SmallVector<SDValue, 16> Elts;
23545 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23546 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23548 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23551 if (isTargetShuffle(N->getOpcode())) {
23553 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23554 if (Shuffle.getNode())
23557 // Try recursively combining arbitrary sequences of x86 shuffle
23558 // instructions into higher-order shuffles. We do this after combining
23559 // specific PSHUF instruction sequences into their minimal form so that we
23560 // can evaluate how many specialized shuffle instructions are involved in
23561 // a particular chain.
23562 SmallVector<int, 1> NonceMask; // Just a placeholder.
23563 NonceMask.push_back(0);
23564 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23565 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23567 return SDValue(); // This routine will use CombineTo to replace N.
23573 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23574 /// specific shuffle of a load can be folded into a single element load.
23575 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23576 /// shuffles have been custom lowered so we need to handle those here.
23577 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23578 TargetLowering::DAGCombinerInfo &DCI) {
23579 if (DCI.isBeforeLegalizeOps())
23582 SDValue InVec = N->getOperand(0);
23583 SDValue EltNo = N->getOperand(1);
23585 if (!isa<ConstantSDNode>(EltNo))
23588 EVT OriginalVT = InVec.getValueType();
23590 if (InVec.getOpcode() == ISD::BITCAST) {
23591 // Don't duplicate a load with other uses.
23592 if (!InVec.hasOneUse())
23594 EVT BCVT = InVec.getOperand(0).getValueType();
23595 if (!BCVT.isVector() ||
23596 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23598 InVec = InVec.getOperand(0);
23601 EVT CurrentVT = InVec.getValueType();
23603 if (!isTargetShuffle(InVec.getOpcode()))
23606 // Don't duplicate a load with other uses.
23607 if (!InVec.hasOneUse())
23610 SmallVector<int, 16> ShuffleMask;
23612 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23613 ShuffleMask, UnaryShuffle))
23616 // Select the input vector, guarding against out of range extract vector.
23617 unsigned NumElems = CurrentVT.getVectorNumElements();
23618 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23619 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23620 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23621 : InVec.getOperand(1);
23623 // If inputs to shuffle are the same for both ops, then allow 2 uses
23624 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23625 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23627 if (LdNode.getOpcode() == ISD::BITCAST) {
23628 // Don't duplicate a load with other uses.
23629 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23632 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23633 LdNode = LdNode.getOperand(0);
23636 if (!ISD::isNormalLoad(LdNode.getNode()))
23639 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23641 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23644 EVT EltVT = N->getValueType(0);
23645 // If there's a bitcast before the shuffle, check if the load type and
23646 // alignment is valid.
23647 unsigned Align = LN0->getAlignment();
23648 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23649 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23650 EltVT.getTypeForEVT(*DAG.getContext()));
23652 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23655 // All checks match so transform back to vector_shuffle so that DAG combiner
23656 // can finish the job
23659 // Create shuffle node taking into account the case that its a unary shuffle
23660 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23661 : InVec.getOperand(1);
23662 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23663 InVec.getOperand(0), Shuffle,
23665 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23666 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23670 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23671 const X86Subtarget *Subtarget) {
23672 SDValue N0 = N->getOperand(0);
23673 EVT VT = N->getValueType(0);
23675 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23676 // special and don't usually play with other vector types, it's better to
23677 // handle them early to be sure we emit efficient code by avoiding
23678 // store-load conversions.
23679 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23680 N0.getValueType() == MVT::v2i32 &&
23681 isNullConstant(N0.getOperand(1))) {
23682 SDValue N00 = N0->getOperand(0);
23683 if (N00.getValueType() == MVT::i32)
23684 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23687 // Convert a bitcasted integer logic operation that has one bitcasted
23688 // floating-point operand and one constant operand into a floating-point
23689 // logic operation. This may create a load of the constant, but that is
23690 // cheaper than materializing the constant in an integer register and
23691 // transferring it to an SSE register or transferring the SSE operand to
23692 // integer register and back.
23694 switch (N0.getOpcode()) {
23695 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23696 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23697 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23698 default: return SDValue();
23700 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23701 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23702 isa<ConstantSDNode>(N0.getOperand(1)) &&
23703 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23704 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23705 SDValue N000 = N0.getOperand(0).getOperand(0);
23706 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23707 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23713 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23714 /// generation and convert it from being a bunch of shuffles and extracts
23715 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23716 /// storing the value and loading scalars back, while for x64 we should
23717 /// use 64-bit extracts and shifts.
23718 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23719 TargetLowering::DAGCombinerInfo &DCI) {
23720 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23723 SDValue InputVector = N->getOperand(0);
23724 SDLoc dl(InputVector);
23725 // Detect mmx to i32 conversion through a v2i32 elt extract.
23726 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23727 N->getValueType(0) == MVT::i32 &&
23728 InputVector.getValueType() == MVT::v2i32) {
23730 // The bitcast source is a direct mmx result.
23731 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23732 if (MMXSrc.getValueType() == MVT::x86mmx)
23733 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23734 N->getValueType(0),
23735 InputVector.getNode()->getOperand(0));
23737 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23738 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23739 MMXSrc.getValueType() == MVT::i64) {
23740 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23741 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23742 MMXSrcOp.getValueType() == MVT::v1i64 &&
23743 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23744 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23745 N->getValueType(0), MMXSrcOp.getOperand(0));
23749 EVT VT = N->getValueType(0);
23751 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
23752 InputVector.getOpcode() == ISD::BITCAST &&
23753 isa<ConstantSDNode>(InputVector.getOperand(0))) {
23754 uint64_t ExtractedElt =
23755 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23756 uint64_t InputValue =
23757 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23758 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23759 return DAG.getConstant(Res, dl, MVT::i1);
23761 // Only operate on vectors of 4 elements, where the alternative shuffling
23762 // gets to be more expensive.
23763 if (InputVector.getValueType() != MVT::v4i32)
23766 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23767 // single use which is a sign-extend or zero-extend, and all elements are
23769 SmallVector<SDNode *, 4> Uses;
23770 unsigned ExtractedElements = 0;
23771 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23772 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23773 if (UI.getUse().getResNo() != InputVector.getResNo())
23776 SDNode *Extract = *UI;
23777 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23780 if (Extract->getValueType(0) != MVT::i32)
23782 if (!Extract->hasOneUse())
23784 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23785 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23787 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23790 // Record which element was extracted.
23791 ExtractedElements |=
23792 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23794 Uses.push_back(Extract);
23797 // If not all the elements were used, this may not be worthwhile.
23798 if (ExtractedElements != 15)
23801 // Ok, we've now decided to do the transformation.
23802 // If 64-bit shifts are legal, use the extract-shift sequence,
23803 // otherwise bounce the vector off the cache.
23804 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23807 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23808 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23809 auto &DL = DAG.getDataLayout();
23810 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23811 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23812 DAG.getConstant(0, dl, VecIdxTy));
23813 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23814 DAG.getConstant(1, dl, VecIdxTy));
23816 SDValue ShAmt = DAG.getConstant(
23817 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23818 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23819 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23820 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23821 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23822 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23823 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23825 // Store the value to a temporary stack slot.
23826 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23827 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23828 MachinePointerInfo(), false, false, 0);
23830 EVT ElementType = InputVector.getValueType().getVectorElementType();
23831 unsigned EltSize = ElementType.getSizeInBits() / 8;
23833 // Replace each use (extract) with a load of the appropriate element.
23834 for (unsigned i = 0; i < 4; ++i) {
23835 uint64_t Offset = EltSize * i;
23836 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23837 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23839 SDValue ScalarAddr =
23840 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23842 // Load the scalar.
23843 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23844 ScalarAddr, MachinePointerInfo(),
23845 false, false, false, 0);
23850 // Replace the extracts
23851 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23852 UE = Uses.end(); UI != UE; ++UI) {
23853 SDNode *Extract = *UI;
23855 SDValue Idx = Extract->getOperand(1);
23856 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23857 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23860 // The replacement was made in place; don't return anything.
23865 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23866 const X86Subtarget *Subtarget) {
23868 SDValue Cond = N->getOperand(0);
23869 SDValue LHS = N->getOperand(1);
23870 SDValue RHS = N->getOperand(2);
23872 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23873 SDValue CondSrc = Cond->getOperand(0);
23874 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23875 Cond = CondSrc->getOperand(0);
23878 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23881 // A vselect where all conditions and data are constants can be optimized into
23882 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23883 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23884 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23887 unsigned MaskValue = 0;
23888 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23891 MVT VT = N->getSimpleValueType(0);
23892 unsigned NumElems = VT.getVectorNumElements();
23893 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23894 for (unsigned i = 0; i < NumElems; ++i) {
23895 // Be sure we emit undef where we can.
23896 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23897 ShuffleMask[i] = -1;
23899 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23902 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23903 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23905 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23908 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23910 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23911 TargetLowering::DAGCombinerInfo &DCI,
23912 const X86Subtarget *Subtarget) {
23914 SDValue Cond = N->getOperand(0);
23915 // Get the LHS/RHS of the select.
23916 SDValue LHS = N->getOperand(1);
23917 SDValue RHS = N->getOperand(2);
23918 EVT VT = LHS.getValueType();
23919 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23921 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23922 // instructions match the semantics of the common C idiom x<y?x:y but not
23923 // x<=y?x:y, because of how they handle negative zero (which can be
23924 // ignored in unsafe-math mode).
23925 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23926 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23927 VT != MVT::f80 && VT != MVT::f128 &&
23928 (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23929 (Subtarget->hasSSE2() ||
23930 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23931 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23933 unsigned Opcode = 0;
23934 // Check for x CC y ? x : y.
23935 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23936 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23940 // Converting this to a min would handle NaNs incorrectly, and swapping
23941 // the operands would cause it to handle comparisons between positive
23942 // and negative zero incorrectly.
23943 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23944 if (!DAG.getTarget().Options.UnsafeFPMath &&
23945 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23947 std::swap(LHS, RHS);
23949 Opcode = X86ISD::FMIN;
23952 // Converting this to a min would handle comparisons between positive
23953 // and negative zero incorrectly.
23954 if (!DAG.getTarget().Options.UnsafeFPMath &&
23955 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23957 Opcode = X86ISD::FMIN;
23960 // Converting this to a min would handle both negative zeros and NaNs
23961 // incorrectly, but we can swap the operands to fix both.
23962 std::swap(LHS, RHS);
23966 Opcode = X86ISD::FMIN;
23970 // Converting this to a max would handle comparisons between positive
23971 // and negative zero incorrectly.
23972 if (!DAG.getTarget().Options.UnsafeFPMath &&
23973 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23975 Opcode = X86ISD::FMAX;
23978 // Converting this to a max would handle NaNs incorrectly, and swapping
23979 // the operands would cause it to handle comparisons between positive
23980 // and negative zero incorrectly.
23981 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23982 if (!DAG.getTarget().Options.UnsafeFPMath &&
23983 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23985 std::swap(LHS, RHS);
23987 Opcode = X86ISD::FMAX;
23990 // Converting this to a max would handle both negative zeros and NaNs
23991 // incorrectly, but we can swap the operands to fix both.
23992 std::swap(LHS, RHS);
23996 Opcode = X86ISD::FMAX;
23999 // Check for x CC y ? y : x -- a min/max with reversed arms.
24000 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
24001 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
24005 // Converting this to a min would handle comparisons between positive
24006 // and negative zero incorrectly, and swapping the operands would
24007 // cause it to handle NaNs incorrectly.
24008 if (!DAG.getTarget().Options.UnsafeFPMath &&
24009 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
24010 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24012 std::swap(LHS, RHS);
24014 Opcode = X86ISD::FMIN;
24017 // Converting this to a min would handle NaNs incorrectly.
24018 if (!DAG.getTarget().Options.UnsafeFPMath &&
24019 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
24021 Opcode = X86ISD::FMIN;
24024 // Converting this to a min would handle both negative zeros and NaNs
24025 // incorrectly, but we can swap the operands to fix both.
24026 std::swap(LHS, RHS);
24030 Opcode = X86ISD::FMIN;
24034 // Converting this to a max would handle NaNs incorrectly.
24035 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24037 Opcode = X86ISD::FMAX;
24040 // Converting this to a max would handle comparisons between positive
24041 // and negative zero incorrectly, and swapping the operands would
24042 // cause it to handle NaNs incorrectly.
24043 if (!DAG.getTarget().Options.UnsafeFPMath &&
24044 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
24045 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24047 std::swap(LHS, RHS);
24049 Opcode = X86ISD::FMAX;
24052 // Converting this to a max would handle both negative zeros and NaNs
24053 // incorrectly, but we can swap the operands to fix both.
24054 std::swap(LHS, RHS);
24058 Opcode = X86ISD::FMAX;
24064 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
24067 EVT CondVT = Cond.getValueType();
24068 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
24069 CondVT.getVectorElementType() == MVT::i1) {
24070 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
24071 // lowering on KNL. In this case we convert it to
24072 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
24073 // The same situation for all 128 and 256-bit vectors of i8 and i16.
24074 // Since SKX these selects have a proper lowering.
24075 EVT OpVT = LHS.getValueType();
24076 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
24077 (OpVT.getVectorElementType() == MVT::i8 ||
24078 OpVT.getVectorElementType() == MVT::i16) &&
24079 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
24080 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
24081 DCI.AddToWorklist(Cond.getNode());
24082 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
24085 // If this is a select between two integer constants, try to do some
24087 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
24088 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
24089 // Don't do this for crazy integer types.
24090 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
24091 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
24092 // so that TrueC (the true value) is larger than FalseC.
24093 bool NeedsCondInvert = false;
24095 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
24096 // Efficiently invertible.
24097 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
24098 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
24099 isa<ConstantSDNode>(Cond.getOperand(1))))) {
24100 NeedsCondInvert = true;
24101 std::swap(TrueC, FalseC);
24104 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
24105 if (FalseC->getAPIntValue() == 0 &&
24106 TrueC->getAPIntValue().isPowerOf2()) {
24107 if (NeedsCondInvert) // Invert the condition if needed.
24108 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24109 DAG.getConstant(1, DL, Cond.getValueType()));
24111 // Zero extend the condition if needed.
24112 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
24114 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24115 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
24116 DAG.getConstant(ShAmt, DL, MVT::i8));
24119 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
24120 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24121 if (NeedsCondInvert) // Invert the condition if needed.
24122 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24123 DAG.getConstant(1, DL, Cond.getValueType()));
24125 // Zero extend the condition if needed.
24126 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24127 FalseC->getValueType(0), Cond);
24128 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24129 SDValue(FalseC, 0));
24132 // Optimize cases that will turn into an LEA instruction. This requires
24133 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24134 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24135 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24136 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24138 bool isFastMultiplier = false;
24140 switch ((unsigned char)Diff) {
24142 case 1: // result = add base, cond
24143 case 2: // result = lea base( , cond*2)
24144 case 3: // result = lea base(cond, cond*2)
24145 case 4: // result = lea base( , cond*4)
24146 case 5: // result = lea base(cond, cond*4)
24147 case 8: // result = lea base( , cond*8)
24148 case 9: // result = lea base(cond, cond*8)
24149 isFastMultiplier = true;
24154 if (isFastMultiplier) {
24155 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24156 if (NeedsCondInvert) // Invert the condition if needed.
24157 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24158 DAG.getConstant(1, DL, Cond.getValueType()));
24160 // Zero extend the condition if needed.
24161 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24163 // Scale the condition by the difference.
24165 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24166 DAG.getConstant(Diff, DL,
24167 Cond.getValueType()));
24169 // Add the base if non-zero.
24170 if (FalseC->getAPIntValue() != 0)
24171 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24172 SDValue(FalseC, 0));
24179 // Canonicalize max and min:
24180 // (x > y) ? x : y -> (x >= y) ? x : y
24181 // (x < y) ? x : y -> (x <= y) ? x : y
24182 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
24183 // the need for an extra compare
24184 // against zero. e.g.
24185 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
24187 // testl %edi, %edi
24189 // cmovgl %edi, %eax
24193 // cmovsl %eax, %edi
24194 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
24195 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24196 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24197 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24202 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
24203 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
24204 Cond.getOperand(0), Cond.getOperand(1), NewCC);
24205 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
24210 // Early exit check
24211 if (!TLI.isTypeLegal(VT))
24214 // Match VSELECTs into subs with unsigned saturation.
24215 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
24216 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
24217 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
24218 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
24219 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24221 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
24222 // left side invert the predicate to simplify logic below.
24224 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
24226 CC = ISD::getSetCCInverse(CC, true);
24227 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
24231 if (Other.getNode() && Other->getNumOperands() == 2 &&
24232 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
24233 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
24234 SDValue CondRHS = Cond->getOperand(1);
24236 // Look for a general sub with unsigned saturation first.
24237 // x >= y ? x-y : 0 --> subus x, y
24238 // x > y ? x-y : 0 --> subus x, y
24239 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
24240 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
24241 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
24243 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
24244 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
24245 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
24246 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
24247 // If the RHS is a constant we have to reverse the const
24248 // canonicalization.
24249 // x > C-1 ? x+-C : 0 --> subus x, C
24250 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
24251 CondRHSConst->getAPIntValue() ==
24252 (-OpRHSConst->getAPIntValue() - 1))
24253 return DAG.getNode(
24254 X86ISD::SUBUS, DL, VT, OpLHS,
24255 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
24257 // Another special case: If C was a sign bit, the sub has been
24258 // canonicalized into a xor.
24259 // FIXME: Would it be better to use computeKnownBits to determine
24260 // whether it's safe to decanonicalize the xor?
24261 // x s< 0 ? x^C : 0 --> subus x, C
24262 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
24263 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
24264 OpRHSConst->getAPIntValue().isSignBit())
24265 // Note that we have to rebuild the RHS constant here to ensure we
24266 // don't rely on particular values of undef lanes.
24267 return DAG.getNode(
24268 X86ISD::SUBUS, DL, VT, OpLHS,
24269 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
24274 // Simplify vector selection if condition value type matches vselect
24276 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
24277 assert(Cond.getValueType().isVector() &&
24278 "vector select expects a vector selector!");
24280 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
24281 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
24283 // Try invert the condition if true value is not all 1s and false value
24285 if (!TValIsAllOnes && !FValIsAllZeros &&
24286 // Check if the selector will be produced by CMPP*/PCMP*
24287 Cond.getOpcode() == ISD::SETCC &&
24288 // Check if SETCC has already been promoted
24289 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
24291 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
24292 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
24294 if (TValIsAllZeros || FValIsAllOnes) {
24295 SDValue CC = Cond.getOperand(2);
24296 ISD::CondCode NewCC =
24297 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
24298 Cond.getOperand(0).getValueType().isInteger());
24299 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
24300 std::swap(LHS, RHS);
24301 TValIsAllOnes = FValIsAllOnes;
24302 FValIsAllZeros = TValIsAllZeros;
24306 if (TValIsAllOnes || FValIsAllZeros) {
24309 if (TValIsAllOnes && FValIsAllZeros)
24311 else if (TValIsAllOnes)
24313 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
24314 else if (FValIsAllZeros)
24315 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24316 DAG.getBitcast(CondVT, LHS));
24318 return DAG.getBitcast(VT, Ret);
24322 // We should generate an X86ISD::BLENDI from a vselect if its argument
24323 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24324 // constants. This specific pattern gets generated when we split a
24325 // selector for a 512 bit vector in a machine without AVX512 (but with
24326 // 256-bit vectors), during legalization:
24328 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24330 // Iff we find this pattern and the build_vectors are built from
24331 // constants, we translate the vselect into a shuffle_vector that we
24332 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24333 if ((N->getOpcode() == ISD::VSELECT ||
24334 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24335 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
24336 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24337 if (Shuffle.getNode())
24341 // If this is a *dynamic* select (non-constant condition) and we can match
24342 // this node with one of the variable blend instructions, restructure the
24343 // condition so that the blends can use the high bit of each element and use
24344 // SimplifyDemandedBits to simplify the condition operand.
24345 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24346 !DCI.isBeforeLegalize() &&
24347 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24348 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
24350 // Don't optimize vector selects that map to mask-registers.
24354 // We can only handle the cases where VSELECT is directly legal on the
24355 // subtarget. We custom lower VSELECT nodes with constant conditions and
24356 // this makes it hard to see whether a dynamic VSELECT will correctly
24357 // lower, so we both check the operation's status and explicitly handle the
24358 // cases where a *dynamic* blend will fail even though a constant-condition
24359 // blend could be custom lowered.
24360 // FIXME: We should find a better way to handle this class of problems.
24361 // Potentially, we should combine constant-condition vselect nodes
24362 // pre-legalization into shuffles and not mark as many types as custom
24364 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
24366 // FIXME: We don't support i16-element blends currently. We could and
24367 // should support them by making *all* the bits in the condition be set
24368 // rather than just the high bit and using an i8-element blend.
24369 if (VT.getVectorElementType() == MVT::i16)
24371 // Dynamic blending was only available from SSE4.1 onward.
24372 if (VT.is128BitVector() && !Subtarget->hasSSE41())
24374 // Byte blends are only available in AVX2
24375 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24378 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24379 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24381 APInt KnownZero, KnownOne;
24382 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24383 DCI.isBeforeLegalizeOps());
24384 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24385 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24387 // If we changed the computation somewhere in the DAG, this change
24388 // will affect all users of Cond.
24389 // Make sure it is fine and update all the nodes so that we do not
24390 // use the generic VSELECT anymore. Otherwise, we may perform
24391 // wrong optimizations as we messed up with the actual expectation
24392 // for the vector boolean values.
24393 if (Cond != TLO.Old) {
24394 // Check all uses of that condition operand to check whether it will be
24395 // consumed by non-BLEND instructions, which may depend on all bits are
24397 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24399 if (I->getOpcode() != ISD::VSELECT)
24400 // TODO: Add other opcodes eventually lowered into BLEND.
24403 // Update all the users of the condition, before committing the change,
24404 // so that the VSELECT optimizations that expect the correct vector
24405 // boolean value will not be triggered.
24406 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24408 DAG.ReplaceAllUsesOfValueWith(
24410 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24411 Cond, I->getOperand(1), I->getOperand(2)));
24412 DCI.CommitTargetLoweringOpt(TLO);
24415 // At this point, only Cond is changed. Change the condition
24416 // just for N to keep the opportunity to optimize all other
24417 // users their own way.
24418 DAG.ReplaceAllUsesOfValueWith(
24420 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24421 TLO.New, N->getOperand(1), N->getOperand(2)));
24429 // Check whether a boolean test is testing a boolean value generated by
24430 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24433 // Simplify the following patterns:
24434 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24435 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24436 // to (Op EFLAGS Cond)
24438 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24439 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24440 // to (Op EFLAGS !Cond)
24442 // where Op could be BRCOND or CMOV.
24444 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24445 // Quit if not CMP and SUB with its value result used.
24446 if (Cmp.getOpcode() != X86ISD::CMP &&
24447 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24450 // Quit if not used as a boolean value.
24451 if (CC != X86::COND_E && CC != X86::COND_NE)
24454 // Check CMP operands. One of them should be 0 or 1 and the other should be
24455 // an SetCC or extended from it.
24456 SDValue Op1 = Cmp.getOperand(0);
24457 SDValue Op2 = Cmp.getOperand(1);
24460 const ConstantSDNode* C = nullptr;
24461 bool needOppositeCond = (CC == X86::COND_E);
24462 bool checkAgainstTrue = false; // Is it a comparison against 1?
24464 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24466 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24468 else // Quit if all operands are not constants.
24471 if (C->getZExtValue() == 1) {
24472 needOppositeCond = !needOppositeCond;
24473 checkAgainstTrue = true;
24474 } else if (C->getZExtValue() != 0)
24475 // Quit if the constant is neither 0 or 1.
24478 bool truncatedToBoolWithAnd = false;
24479 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24480 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24481 SetCC.getOpcode() == ISD::TRUNCATE ||
24482 SetCC.getOpcode() == ISD::AND) {
24483 if (SetCC.getOpcode() == ISD::AND) {
24485 if (isOneConstant(SetCC.getOperand(0)))
24487 if (isOneConstant(SetCC.getOperand(1)))
24491 SetCC = SetCC.getOperand(OpIdx);
24492 truncatedToBoolWithAnd = true;
24494 SetCC = SetCC.getOperand(0);
24497 switch (SetCC.getOpcode()) {
24498 case X86ISD::SETCC_CARRY:
24499 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24500 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24501 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24502 // truncated to i1 using 'and'.
24503 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24505 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24506 "Invalid use of SETCC_CARRY!");
24508 case X86ISD::SETCC:
24509 // Set the condition code or opposite one if necessary.
24510 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24511 if (needOppositeCond)
24512 CC = X86::GetOppositeBranchCondition(CC);
24513 return SetCC.getOperand(1);
24514 case X86ISD::CMOV: {
24515 // Check whether false/true value has canonical one, i.e. 0 or 1.
24516 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24517 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24518 // Quit if true value is not a constant.
24521 // Quit if false value is not a constant.
24523 SDValue Op = SetCC.getOperand(0);
24524 // Skip 'zext' or 'trunc' node.
24525 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24526 Op.getOpcode() == ISD::TRUNCATE)
24527 Op = Op.getOperand(0);
24528 // A special case for rdrand/rdseed, where 0 is set if false cond is
24530 if ((Op.getOpcode() != X86ISD::RDRAND &&
24531 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24534 // Quit if false value is not the constant 0 or 1.
24535 bool FValIsFalse = true;
24536 if (FVal && FVal->getZExtValue() != 0) {
24537 if (FVal->getZExtValue() != 1)
24539 // If FVal is 1, opposite cond is needed.
24540 needOppositeCond = !needOppositeCond;
24541 FValIsFalse = false;
24543 // Quit if TVal is not the constant opposite of FVal.
24544 if (FValIsFalse && TVal->getZExtValue() != 1)
24546 if (!FValIsFalse && TVal->getZExtValue() != 0)
24548 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24549 if (needOppositeCond)
24550 CC = X86::GetOppositeBranchCondition(CC);
24551 return SetCC.getOperand(3);
24558 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24560 /// (X86or (X86setcc) (X86setcc))
24561 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24562 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24563 X86::CondCode &CC1, SDValue &Flags,
24565 if (Cond->getOpcode() == X86ISD::CMP) {
24566 if (!isNullConstant(Cond->getOperand(1)))
24569 Cond = Cond->getOperand(0);
24574 SDValue SetCC0, SetCC1;
24575 switch (Cond->getOpcode()) {
24576 default: return false;
24583 SetCC0 = Cond->getOperand(0);
24584 SetCC1 = Cond->getOperand(1);
24588 // Make sure we have SETCC nodes, using the same flags value.
24589 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24590 SetCC1.getOpcode() != X86ISD::SETCC ||
24591 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24594 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24595 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24596 Flags = SetCC0->getOperand(1);
24600 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24601 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24602 TargetLowering::DAGCombinerInfo &DCI,
24603 const X86Subtarget *Subtarget) {
24606 // If the flag operand isn't dead, don't touch this CMOV.
24607 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24610 SDValue FalseOp = N->getOperand(0);
24611 SDValue TrueOp = N->getOperand(1);
24612 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24613 SDValue Cond = N->getOperand(3);
24615 if (CC == X86::COND_E || CC == X86::COND_NE) {
24616 switch (Cond.getOpcode()) {
24620 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24621 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24622 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24628 Flags = checkBoolTestSetCCCombine(Cond, CC);
24629 if (Flags.getNode() &&
24630 // Extra check as FCMOV only supports a subset of X86 cond.
24631 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24632 SDValue Ops[] = { FalseOp, TrueOp,
24633 DAG.getConstant(CC, DL, MVT::i8), Flags };
24634 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24637 // If this is a select between two integer constants, try to do some
24638 // optimizations. Note that the operands are ordered the opposite of SELECT
24640 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24641 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24642 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24643 // larger than FalseC (the false value).
24644 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24645 CC = X86::GetOppositeBranchCondition(CC);
24646 std::swap(TrueC, FalseC);
24647 std::swap(TrueOp, FalseOp);
24650 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24651 // This is efficient for any integer data type (including i8/i16) and
24653 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24654 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24655 DAG.getConstant(CC, DL, MVT::i8), Cond);
24657 // Zero extend the condition if needed.
24658 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24660 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24661 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24662 DAG.getConstant(ShAmt, DL, MVT::i8));
24663 if (N->getNumValues() == 2) // Dead flag value?
24664 return DCI.CombineTo(N, Cond, SDValue());
24668 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24669 // for any integer data type, including i8/i16.
24670 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24671 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24672 DAG.getConstant(CC, DL, MVT::i8), Cond);
24674 // Zero extend the condition if needed.
24675 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24676 FalseC->getValueType(0), Cond);
24677 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24678 SDValue(FalseC, 0));
24680 if (N->getNumValues() == 2) // Dead flag value?
24681 return DCI.CombineTo(N, Cond, SDValue());
24685 // Optimize cases that will turn into an LEA instruction. This requires
24686 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24687 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24688 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24689 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24691 bool isFastMultiplier = false;
24693 switch ((unsigned char)Diff) {
24695 case 1: // result = add base, cond
24696 case 2: // result = lea base( , cond*2)
24697 case 3: // result = lea base(cond, cond*2)
24698 case 4: // result = lea base( , cond*4)
24699 case 5: // result = lea base(cond, cond*4)
24700 case 8: // result = lea base( , cond*8)
24701 case 9: // result = lea base(cond, cond*8)
24702 isFastMultiplier = true;
24707 if (isFastMultiplier) {
24708 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24709 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24710 DAG.getConstant(CC, DL, MVT::i8), Cond);
24711 // Zero extend the condition if needed.
24712 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24714 // Scale the condition by the difference.
24716 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24717 DAG.getConstant(Diff, DL, Cond.getValueType()));
24719 // Add the base if non-zero.
24720 if (FalseC->getAPIntValue() != 0)
24721 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24722 SDValue(FalseC, 0));
24723 if (N->getNumValues() == 2) // Dead flag value?
24724 return DCI.CombineTo(N, Cond, SDValue());
24731 // Handle these cases:
24732 // (select (x != c), e, c) -> select (x != c), e, x),
24733 // (select (x == c), c, e) -> select (x == c), x, e)
24734 // where the c is an integer constant, and the "select" is the combination
24735 // of CMOV and CMP.
24737 // The rationale for this change is that the conditional-move from a constant
24738 // needs two instructions, however, conditional-move from a register needs
24739 // only one instruction.
24741 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24742 // some instruction-combining opportunities. This opt needs to be
24743 // postponed as late as possible.
24745 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24746 // the DCI.xxxx conditions are provided to postpone the optimization as
24747 // late as possible.
24749 ConstantSDNode *CmpAgainst = nullptr;
24750 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24751 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24752 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24754 if (CC == X86::COND_NE &&
24755 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24756 CC = X86::GetOppositeBranchCondition(CC);
24757 std::swap(TrueOp, FalseOp);
24760 if (CC == X86::COND_E &&
24761 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24762 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24763 DAG.getConstant(CC, DL, MVT::i8), Cond };
24764 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24769 // Fold and/or of setcc's to double CMOV:
24770 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24771 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24773 // This combine lets us generate:
24774 // cmovcc1 (jcc1 if we don't have CMOV)
24780 // cmovne (jne if we don't have CMOV)
24781 // When we can't use the CMOV instruction, it might increase branch
24783 // When we can use CMOV, or when there is no mispredict, this improves
24784 // throughput and reduces register pressure.
24786 if (CC == X86::COND_NE) {
24788 X86::CondCode CC0, CC1;
24790 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24792 std::swap(FalseOp, TrueOp);
24793 CC0 = X86::GetOppositeBranchCondition(CC0);
24794 CC1 = X86::GetOppositeBranchCondition(CC1);
24797 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24799 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24800 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24801 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24802 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24810 /// PerformMulCombine - Optimize a single multiply with constant into two
24811 /// in order to implement it with two cheaper instructions, e.g.
24812 /// LEA + SHL, LEA + LEA.
24813 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24814 TargetLowering::DAGCombinerInfo &DCI) {
24815 // An imul is usually smaller than the alternative sequence.
24816 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24819 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24822 EVT VT = N->getValueType(0);
24823 if (VT != MVT::i64 && VT != MVT::i32)
24826 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24829 uint64_t MulAmt = C->getZExtValue();
24830 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24833 uint64_t MulAmt1 = 0;
24834 uint64_t MulAmt2 = 0;
24835 if ((MulAmt % 9) == 0) {
24837 MulAmt2 = MulAmt / 9;
24838 } else if ((MulAmt % 5) == 0) {
24840 MulAmt2 = MulAmt / 5;
24841 } else if ((MulAmt % 3) == 0) {
24843 MulAmt2 = MulAmt / 3;
24849 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24851 if (isPowerOf2_64(MulAmt2) &&
24852 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24853 // If second multiplifer is pow2, issue it first. We want the multiply by
24854 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24856 std::swap(MulAmt1, MulAmt2);
24858 if (isPowerOf2_64(MulAmt1))
24859 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24860 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24862 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24863 DAG.getConstant(MulAmt1, DL, VT));
24865 if (isPowerOf2_64(MulAmt2))
24866 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24867 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24869 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24870 DAG.getConstant(MulAmt2, DL, VT));
24874 assert(MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX)
24875 && "Both cases that could cause potential overflows should have "
24876 "already been handled.");
24877 if (isPowerOf2_64(MulAmt - 1))
24878 // (mul x, 2^N + 1) => (add (shl x, N), x)
24879 NewMul = DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
24880 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24881 DAG.getConstant(Log2_64(MulAmt - 1), DL,
24884 else if (isPowerOf2_64(MulAmt + 1))
24885 // (mul x, 2^N - 1) => (sub (shl x, N), x)
24886 NewMul = DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::SHL, DL, VT,
24888 DAG.getConstant(Log2_64(MulAmt + 1),
24889 DL, MVT::i8)), N->getOperand(0));
24893 // Do not add new nodes to DAG combiner worklist.
24894 DCI.CombineTo(N, NewMul, false);
24899 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24900 SDValue N0 = N->getOperand(0);
24901 SDValue N1 = N->getOperand(1);
24902 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24903 EVT VT = N0.getValueType();
24905 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24906 // since the result of setcc_c is all zero's or all ones.
24907 if (VT.isInteger() && !VT.isVector() &&
24908 N1C && N0.getOpcode() == ISD::AND &&
24909 N0.getOperand(1).getOpcode() == ISD::Constant) {
24910 SDValue N00 = N0.getOperand(0);
24911 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24912 APInt ShAmt = N1C->getAPIntValue();
24913 Mask = Mask.shl(ShAmt);
24914 bool MaskOK = false;
24915 // We can handle cases concerning bit-widening nodes containing setcc_c if
24916 // we carefully interrogate the mask to make sure we are semantics
24918 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24919 // of the underlying setcc_c operation if the setcc_c was zero extended.
24920 // Consider the following example:
24921 // zext(setcc_c) -> i32 0x0000FFFF
24922 // c1 -> i32 0x0000FFFF
24923 // c2 -> i32 0x00000001
24924 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24925 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24926 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24928 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24929 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24931 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24932 N00.getOpcode() == ISD::ANY_EXTEND) &&
24933 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24934 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24936 if (MaskOK && Mask != 0) {
24938 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24942 // Hardware support for vector shifts is sparse which makes us scalarize the
24943 // vector operations in many cases. Also, on sandybridge ADD is faster than
24945 // (shl V, 1) -> add V,V
24946 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24947 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24948 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24949 // We shift all of the values by one. In many cases we do not have
24950 // hardware support for this operation. This is better expressed as an ADD
24952 if (N1SplatC->getAPIntValue() == 1)
24953 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24959 static SDValue PerformSRACombine(SDNode *N, SelectionDAG &DAG) {
24960 SDValue N0 = N->getOperand(0);
24961 SDValue N1 = N->getOperand(1);
24962 EVT VT = N0.getValueType();
24963 unsigned Size = VT.getSizeInBits();
24965 // fold (ashr (shl, a, [56,48,32,24,16]), SarConst)
24966 // into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or
24967 // into (lshr, (sext (a), SarConst - [56,48,32,24,16]))
24968 // depending on sign of (SarConst - [56,48,32,24,16])
24970 // sexts in X86 are MOVs. The MOVs have the same code size
24971 // as above SHIFTs (only SHIFT on 1 has lower code size).
24972 // However the MOVs have 2 advantages to a SHIFT:
24973 // 1. MOVs can write to a register that differs from source
24974 // 2. MOVs accept memory operands
24976 if (!VT.isInteger() || VT.isVector() || N1.getOpcode() != ISD::Constant ||
24977 N0.getOpcode() != ISD::SHL || !N0.hasOneUse() ||
24978 N0.getOperand(1).getOpcode() != ISD::Constant)
24981 SDValue N00 = N0.getOperand(0);
24982 SDValue N01 = N0.getOperand(1);
24983 APInt ShlConst = (cast<ConstantSDNode>(N01))->getAPIntValue();
24984 APInt SarConst = (cast<ConstantSDNode>(N1))->getAPIntValue();
24985 EVT CVT = N1.getValueType();
24987 if (SarConst.isNegative())
24990 for (MVT SVT : MVT::integer_valuetypes()) {
24991 unsigned ShiftSize = SVT.getSizeInBits();
24992 // skipping types without corresponding sext/zext and
24993 // ShlConst that is not one of [56,48,32,24,16]
24994 if (ShiftSize < 8 || ShiftSize > 64 || ShlConst != Size - ShiftSize)
24998 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, N00, DAG.getValueType(SVT));
24999 SarConst = SarConst - (Size - ShiftSize);
25002 else if (SarConst.isNegative())
25003 return DAG.getNode(ISD::SHL, DL, VT, NN,
25004 DAG.getConstant(-SarConst, DL, CVT));
25006 return DAG.getNode(ISD::SRA, DL, VT, NN,
25007 DAG.getConstant(SarConst, DL, CVT));
25012 /// \brief Returns a vector of 0s if the node in input is a vector logical
25013 /// shift by a constant amount which is known to be bigger than or equal
25014 /// to the vector element size in bits.
25015 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
25016 const X86Subtarget *Subtarget) {
25017 EVT VT = N->getValueType(0);
25019 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
25020 (!Subtarget->hasInt256() ||
25021 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
25024 SDValue Amt = N->getOperand(1);
25026 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
25027 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
25028 APInt ShiftAmt = AmtSplat->getAPIntValue();
25029 unsigned MaxAmount =
25030 VT.getSimpleVT().getVectorElementType().getSizeInBits();
25032 // SSE2/AVX2 logical shifts always return a vector of 0s
25033 // if the shift amount is bigger than or equal to
25034 // the element size. The constant shift amount will be
25035 // encoded as a 8-bit immediate.
25036 if (ShiftAmt.trunc(8).uge(MaxAmount))
25037 return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL);
25043 /// PerformShiftCombine - Combine shifts.
25044 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
25045 TargetLowering::DAGCombinerInfo &DCI,
25046 const X86Subtarget *Subtarget) {
25047 if (N->getOpcode() == ISD::SHL)
25048 if (SDValue V = PerformSHLCombine(N, DAG))
25051 if (N->getOpcode() == ISD::SRA)
25052 if (SDValue V = PerformSRACombine(N, DAG))
25055 // Try to fold this logical shift into a zero vector.
25056 if (N->getOpcode() != ISD::SRA)
25057 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
25063 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
25064 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
25065 // and friends. Likewise for OR -> CMPNEQSS.
25066 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
25067 TargetLowering::DAGCombinerInfo &DCI,
25068 const X86Subtarget *Subtarget) {
25071 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
25072 // we're requiring SSE2 for both.
25073 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
25074 SDValue N0 = N->getOperand(0);
25075 SDValue N1 = N->getOperand(1);
25076 SDValue CMP0 = N0->getOperand(1);
25077 SDValue CMP1 = N1->getOperand(1);
25080 // The SETCCs should both refer to the same CMP.
25081 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
25084 SDValue CMP00 = CMP0->getOperand(0);
25085 SDValue CMP01 = CMP0->getOperand(1);
25086 EVT VT = CMP00.getValueType();
25088 if (VT == MVT::f32 || VT == MVT::f64) {
25089 bool ExpectingFlags = false;
25090 // Check for any users that want flags:
25091 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
25092 !ExpectingFlags && UI != UE; ++UI)
25093 switch (UI->getOpcode()) {
25098 ExpectingFlags = true;
25100 case ISD::CopyToReg:
25101 case ISD::SIGN_EXTEND:
25102 case ISD::ZERO_EXTEND:
25103 case ISD::ANY_EXTEND:
25107 if (!ExpectingFlags) {
25108 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
25109 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
25111 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
25112 X86::CondCode tmp = cc0;
25117 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
25118 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
25119 // FIXME: need symbolic constants for these magic numbers.
25120 // See X86ATTInstPrinter.cpp:printSSECC().
25121 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
25122 if (Subtarget->hasAVX512()) {
25123 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
25125 DAG.getConstant(x86cc, DL, MVT::i8));
25126 if (N->getValueType(0) != MVT::i1)
25127 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
25131 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
25132 CMP00.getValueType(), CMP00, CMP01,
25133 DAG.getConstant(x86cc, DL,
25136 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
25137 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
25139 if (is64BitFP && !Subtarget->is64Bit()) {
25140 // On a 32-bit target, we cannot bitcast the 64-bit float to a
25141 // 64-bit integer, since that's not a legal type. Since
25142 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
25143 // bits, but can do this little dance to extract the lowest 32 bits
25144 // and work with those going forward.
25145 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
25147 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
25148 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
25149 Vector32, DAG.getIntPtrConstant(0, DL));
25153 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
25154 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
25155 DAG.getConstant(1, DL, IntVT));
25156 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
25158 return OneBitOfTruth;
25166 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
25167 /// so it can be folded inside ANDNP.
25168 static bool CanFoldXORWithAllOnes(const SDNode *N) {
25169 EVT VT = N->getValueType(0);
25171 // Match direct AllOnes for 128 and 256-bit vectors
25172 if (ISD::isBuildVectorAllOnes(N))
25175 // Look through a bit convert.
25176 if (N->getOpcode() == ISD::BITCAST)
25177 N = N->getOperand(0).getNode();
25179 // Sometimes the operand may come from a insert_subvector building a 256-bit
25181 if (VT.is256BitVector() &&
25182 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
25183 SDValue V1 = N->getOperand(0);
25184 SDValue V2 = N->getOperand(1);
25186 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
25187 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
25188 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
25189 ISD::isBuildVectorAllOnes(V2.getNode()))
25196 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
25197 // register. In most cases we actually compare or select YMM-sized registers
25198 // and mixing the two types creates horrible code. This method optimizes
25199 // some of the transition sequences.
25200 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
25201 TargetLowering::DAGCombinerInfo &DCI,
25202 const X86Subtarget *Subtarget) {
25203 EVT VT = N->getValueType(0);
25204 if (!VT.is256BitVector())
25207 assert((N->getOpcode() == ISD::ANY_EXTEND ||
25208 N->getOpcode() == ISD::ZERO_EXTEND ||
25209 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
25211 SDValue Narrow = N->getOperand(0);
25212 EVT NarrowVT = Narrow->getValueType(0);
25213 if (!NarrowVT.is128BitVector())
25216 if (Narrow->getOpcode() != ISD::XOR &&
25217 Narrow->getOpcode() != ISD::AND &&
25218 Narrow->getOpcode() != ISD::OR)
25221 SDValue N0 = Narrow->getOperand(0);
25222 SDValue N1 = Narrow->getOperand(1);
25225 // The Left side has to be a trunc.
25226 if (N0.getOpcode() != ISD::TRUNCATE)
25229 // The type of the truncated inputs.
25230 EVT WideVT = N0->getOperand(0)->getValueType(0);
25234 // The right side has to be a 'trunc' or a constant vector.
25235 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
25236 ConstantSDNode *RHSConstSplat = nullptr;
25237 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
25238 RHSConstSplat = RHSBV->getConstantSplatNode();
25239 if (!RHSTrunc && !RHSConstSplat)
25242 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25244 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
25247 // Set N0 and N1 to hold the inputs to the new wide operation.
25248 N0 = N0->getOperand(0);
25249 if (RHSConstSplat) {
25250 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
25251 SDValue(RHSConstSplat, 0));
25252 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
25253 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
25254 } else if (RHSTrunc) {
25255 N1 = N1->getOperand(0);
25258 // Generate the wide operation.
25259 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
25260 unsigned Opcode = N->getOpcode();
25262 case ISD::ANY_EXTEND:
25264 case ISD::ZERO_EXTEND: {
25265 unsigned InBits = NarrowVT.getScalarSizeInBits();
25266 APInt Mask = APInt::getAllOnesValue(InBits);
25267 Mask = Mask.zext(VT.getScalarSizeInBits());
25268 return DAG.getNode(ISD::AND, DL, VT,
25269 Op, DAG.getConstant(Mask, DL, VT));
25271 case ISD::SIGN_EXTEND:
25272 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
25273 Op, DAG.getValueType(NarrowVT));
25275 llvm_unreachable("Unexpected opcode");
25279 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
25280 TargetLowering::DAGCombinerInfo &DCI,
25281 const X86Subtarget *Subtarget) {
25282 SDValue N0 = N->getOperand(0);
25283 SDValue N1 = N->getOperand(1);
25286 // A vector zext_in_reg may be represented as a shuffle,
25287 // feeding into a bitcast (this represents anyext) feeding into
25288 // an and with a mask.
25289 // We'd like to try to combine that into a shuffle with zero
25290 // plus a bitcast, removing the and.
25291 if (N0.getOpcode() != ISD::BITCAST ||
25292 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
25295 // The other side of the AND should be a splat of 2^C, where C
25296 // is the number of bits in the source type.
25297 if (N1.getOpcode() == ISD::BITCAST)
25298 N1 = N1.getOperand(0);
25299 if (N1.getOpcode() != ISD::BUILD_VECTOR)
25301 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
25303 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
25304 EVT SrcType = Shuffle->getValueType(0);
25306 // We expect a single-source shuffle
25307 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
25310 unsigned SrcSize = SrcType.getScalarSizeInBits();
25312 APInt SplatValue, SplatUndef;
25313 unsigned SplatBitSize;
25315 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
25316 SplatBitSize, HasAnyUndefs))
25319 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
25320 // Make sure the splat matches the mask we expect
25321 if (SplatBitSize > ResSize ||
25322 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
25325 // Make sure the input and output size make sense
25326 if (SrcSize >= ResSize || ResSize % SrcSize)
25329 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
25330 // The number of u's between each two values depends on the ratio between
25331 // the source and dest type.
25332 unsigned ZextRatio = ResSize / SrcSize;
25333 bool IsZext = true;
25334 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
25335 if (i % ZextRatio) {
25336 if (Shuffle->getMaskElt(i) > 0) {
25342 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
25343 // Expected element number
25353 // Ok, perform the transformation - replace the shuffle with
25354 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
25355 // (instead of undef) where the k elements come from the zero vector.
25356 SmallVector<int, 8> Mask;
25357 unsigned NumElems = SrcType.getVectorNumElements();
25358 for (unsigned i = 0; i < NumElems; ++i)
25360 Mask.push_back(NumElems);
25362 Mask.push_back(i / ZextRatio);
25364 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
25365 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
25366 return DAG.getBitcast(N0.getValueType(), NewShuffle);
25369 /// If both input operands of a logic op are being cast from floating point
25370 /// types, try to convert this into a floating point logic node to avoid
25371 /// unnecessary moves from SSE to integer registers.
25372 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
25373 const X86Subtarget *Subtarget) {
25374 unsigned FPOpcode = ISD::DELETED_NODE;
25375 if (N->getOpcode() == ISD::AND)
25376 FPOpcode = X86ISD::FAND;
25377 else if (N->getOpcode() == ISD::OR)
25378 FPOpcode = X86ISD::FOR;
25379 else if (N->getOpcode() == ISD::XOR)
25380 FPOpcode = X86ISD::FXOR;
25382 assert(FPOpcode != ISD::DELETED_NODE &&
25383 "Unexpected input node for FP logic conversion");
25385 EVT VT = N->getValueType(0);
25386 SDValue N0 = N->getOperand(0);
25387 SDValue N1 = N->getOperand(1);
25389 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
25390 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
25391 (Subtarget->hasSSE2() && VT == MVT::i64))) {
25392 SDValue N00 = N0.getOperand(0);
25393 SDValue N10 = N1.getOperand(0);
25394 EVT N00Type = N00.getValueType();
25395 EVT N10Type = N10.getValueType();
25396 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
25397 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
25398 return DAG.getBitcast(VT, FPLogic);
25404 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
25405 TargetLowering::DAGCombinerInfo &DCI,
25406 const X86Subtarget *Subtarget) {
25407 if (DCI.isBeforeLegalizeOps())
25410 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
25413 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25416 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25419 EVT VT = N->getValueType(0);
25420 SDValue N0 = N->getOperand(0);
25421 SDValue N1 = N->getOperand(1);
25424 // Create BEXTR instructions
25425 // BEXTR is ((X >> imm) & (2**size-1))
25426 if (VT == MVT::i32 || VT == MVT::i64) {
25427 // Check for BEXTR.
25428 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
25429 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
25430 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
25431 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25432 if (MaskNode && ShiftNode) {
25433 uint64_t Mask = MaskNode->getZExtValue();
25434 uint64_t Shift = ShiftNode->getZExtValue();
25435 if (isMask_64(Mask)) {
25436 uint64_t MaskSize = countPopulation(Mask);
25437 if (Shift + MaskSize <= VT.getSizeInBits())
25438 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25439 DAG.getConstant(Shift | (MaskSize << 8), DL,
25448 // Want to form ANDNP nodes:
25449 // 1) In the hopes of then easily combining them with OR and AND nodes
25450 // to form PBLEND/PSIGN.
25451 // 2) To match ANDN packed intrinsics
25452 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25455 // Check LHS for vnot
25456 if (N0.getOpcode() == ISD::XOR &&
25457 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25458 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25459 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25461 // Check RHS for vnot
25462 if (N1.getOpcode() == ISD::XOR &&
25463 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25464 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25465 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25470 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25471 TargetLowering::DAGCombinerInfo &DCI,
25472 const X86Subtarget *Subtarget) {
25473 if (DCI.isBeforeLegalizeOps())
25476 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25479 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25482 SDValue N0 = N->getOperand(0);
25483 SDValue N1 = N->getOperand(1);
25484 EVT VT = N->getValueType(0);
25486 // look for psign/blend
25487 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25488 if (!Subtarget->hasSSSE3() ||
25489 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25492 // Canonicalize pandn to RHS
25493 if (N0.getOpcode() == X86ISD::ANDNP)
25495 // or (and (m, y), (pandn m, x))
25496 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25497 SDValue Mask = N1.getOperand(0);
25498 SDValue X = N1.getOperand(1);
25500 if (N0.getOperand(0) == Mask)
25501 Y = N0.getOperand(1);
25502 if (N0.getOperand(1) == Mask)
25503 Y = N0.getOperand(0);
25505 // Check to see if the mask appeared in both the AND and ANDNP and
25509 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25510 // Look through mask bitcast.
25511 if (Mask.getOpcode() == ISD::BITCAST)
25512 Mask = Mask.getOperand(0);
25513 if (X.getOpcode() == ISD::BITCAST)
25514 X = X.getOperand(0);
25515 if (Y.getOpcode() == ISD::BITCAST)
25516 Y = Y.getOperand(0);
25518 EVT MaskVT = Mask.getValueType();
25520 // Validate that the Mask operand is a vector sra node.
25521 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25522 // there is no psrai.b
25523 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25524 unsigned SraAmt = ~0;
25525 if (Mask.getOpcode() == ISD::SRA) {
25526 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25527 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25528 SraAmt = AmtConst->getZExtValue();
25529 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25530 SDValue SraC = Mask.getOperand(1);
25531 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25533 if ((SraAmt + 1) != EltBits)
25538 // Now we know we at least have a plendvb with the mask val. See if
25539 // we can form a psignb/w/d.
25540 // psign = x.type == y.type == mask.type && y = sub(0, x);
25541 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25542 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25543 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25544 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25545 "Unsupported VT for PSIGN");
25546 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25547 return DAG.getBitcast(VT, Mask);
25549 // PBLENDVB only available on SSE 4.1
25550 if (!Subtarget->hasSSE41())
25553 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25555 X = DAG.getBitcast(BlendVT, X);
25556 Y = DAG.getBitcast(BlendVT, Y);
25557 Mask = DAG.getBitcast(BlendVT, Mask);
25558 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25559 return DAG.getBitcast(VT, Mask);
25563 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25566 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25567 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25569 // SHLD/SHRD instructions have lower register pressure, but on some
25570 // platforms they have higher latency than the equivalent
25571 // series of shifts/or that would otherwise be generated.
25572 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25573 // have higher latencies and we are not optimizing for size.
25574 if (!OptForSize && Subtarget->isSHLDSlow())
25577 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25579 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25581 if (!N0.hasOneUse() || !N1.hasOneUse())
25584 SDValue ShAmt0 = N0.getOperand(1);
25585 if (ShAmt0.getValueType() != MVT::i8)
25587 SDValue ShAmt1 = N1.getOperand(1);
25588 if (ShAmt1.getValueType() != MVT::i8)
25590 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25591 ShAmt0 = ShAmt0.getOperand(0);
25592 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25593 ShAmt1 = ShAmt1.getOperand(0);
25596 unsigned Opc = X86ISD::SHLD;
25597 SDValue Op0 = N0.getOperand(0);
25598 SDValue Op1 = N1.getOperand(0);
25599 if (ShAmt0.getOpcode() == ISD::SUB) {
25600 Opc = X86ISD::SHRD;
25601 std::swap(Op0, Op1);
25602 std::swap(ShAmt0, ShAmt1);
25605 unsigned Bits = VT.getSizeInBits();
25606 if (ShAmt1.getOpcode() == ISD::SUB) {
25607 SDValue Sum = ShAmt1.getOperand(0);
25608 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25609 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25610 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25611 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25612 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25613 return DAG.getNode(Opc, DL, VT,
25615 DAG.getNode(ISD::TRUNCATE, DL,
25618 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25619 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25621 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25622 return DAG.getNode(Opc, DL, VT,
25623 N0.getOperand(0), N1.getOperand(0),
25624 DAG.getNode(ISD::TRUNCATE, DL,
25631 // Generate NEG and CMOV for integer abs.
25632 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25633 EVT VT = N->getValueType(0);
25635 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25636 // 8-bit integer abs to NEG and CMOV.
25637 if (VT.isInteger() && VT.getSizeInBits() == 8)
25640 SDValue N0 = N->getOperand(0);
25641 SDValue N1 = N->getOperand(1);
25644 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25645 // and change it to SUB and CMOV.
25646 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25647 N0.getOpcode() == ISD::ADD &&
25648 N0.getOperand(1) == N1 &&
25649 N1.getOpcode() == ISD::SRA &&
25650 N1.getOperand(0) == N0.getOperand(0))
25651 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25652 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25653 // Generate SUB & CMOV.
25654 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25655 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25657 SDValue Ops[] = { N0.getOperand(0), Neg,
25658 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25659 SDValue(Neg.getNode(), 1) };
25660 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25665 // Try to turn tests against the signbit in the form of:
25666 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25669 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25670 // This is only worth doing if the output type is i8.
25671 if (N->getValueType(0) != MVT::i8)
25674 SDValue N0 = N->getOperand(0);
25675 SDValue N1 = N->getOperand(1);
25677 // We should be performing an xor against a truncated shift.
25678 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25681 // Make sure we are performing an xor against one.
25682 if (!isOneConstant(N1))
25685 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25686 SDValue Shift = N0.getOperand(0);
25687 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25690 // Make sure we are truncating from one of i16, i32 or i64.
25691 EVT ShiftTy = Shift.getValueType();
25692 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25695 // Make sure the shift amount extracts the sign bit.
25696 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25697 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25700 // Create a greater-than comparison against -1.
25701 // N.B. Using SETGE against 0 works but we want a canonical looking
25702 // comparison, using SETGT matches up with what TranslateX86CC.
25704 SDValue ShiftOp = Shift.getOperand(0);
25705 EVT ShiftOpTy = ShiftOp.getValueType();
25706 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25707 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25711 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25712 TargetLowering::DAGCombinerInfo &DCI,
25713 const X86Subtarget *Subtarget) {
25714 if (DCI.isBeforeLegalizeOps())
25717 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25720 if (Subtarget->hasCMov())
25721 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25724 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25730 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
25731 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
25732 /// X86ISD::AVG instruction.
25733 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
25734 const X86Subtarget *Subtarget, SDLoc DL) {
25735 if (!VT.isVector() || !VT.isSimple())
25737 EVT InVT = In.getValueType();
25738 unsigned NumElems = VT.getVectorNumElements();
25740 EVT ScalarVT = VT.getVectorElementType();
25741 if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
25742 isPowerOf2_32(NumElems)))
25745 // InScalarVT is the intermediate type in AVG pattern and it should be greater
25746 // than the original input type (i8/i16).
25747 EVT InScalarVT = InVT.getVectorElementType();
25748 if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
25751 if (Subtarget->hasAVX512()) {
25752 if (VT.getSizeInBits() > 512)
25754 } else if (Subtarget->hasAVX2()) {
25755 if (VT.getSizeInBits() > 256)
25758 if (VT.getSizeInBits() > 128)
25762 // Detect the following pattern:
25764 // %1 = zext <N x i8> %a to <N x i32>
25765 // %2 = zext <N x i8> %b to <N x i32>
25766 // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
25767 // %4 = add nuw nsw <N x i32> %3, %2
25768 // %5 = lshr <N x i32> %N, <i32 1 x N>
25769 // %6 = trunc <N x i32> %5 to <N x i8>
25771 // In AVX512, the last instruction can also be a trunc store.
25773 if (In.getOpcode() != ISD::SRL)
25776 // A lambda checking the given SDValue is a constant vector and each element
25777 // is in the range [Min, Max].
25778 auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
25779 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
25780 if (!BV || !BV->isConstant())
25782 for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
25783 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
25786 uint64_t Val = C->getZExtValue();
25787 if (Val < Min || Val > Max)
25793 // Check if each element of the vector is left-shifted by one.
25794 auto LHS = In.getOperand(0);
25795 auto RHS = In.getOperand(1);
25796 if (!IsConstVectorInRange(RHS, 1, 1))
25798 if (LHS.getOpcode() != ISD::ADD)
25801 // Detect a pattern of a + b + 1 where the order doesn't matter.
25802 SDValue Operands[3];
25803 Operands[0] = LHS.getOperand(0);
25804 Operands[1] = LHS.getOperand(1);
25806 // Take care of the case when one of the operands is a constant vector whose
25807 // element is in the range [1, 256].
25808 if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
25809 Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
25810 Operands[0].getOperand(0).getValueType() == VT) {
25811 // The pattern is detected. Subtract one from the constant vector, then
25812 // demote it and emit X86ISD::AVG instruction.
25813 SDValue One = DAG.getConstant(1, DL, InScalarVT);
25814 SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
25815 SmallVector<SDValue, 8>(NumElems, One));
25816 Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
25817 Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
25818 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
25822 if (Operands[0].getOpcode() == ISD::ADD)
25823 std::swap(Operands[0], Operands[1]);
25824 else if (Operands[1].getOpcode() != ISD::ADD)
25826 Operands[2] = Operands[1].getOperand(0);
25827 Operands[1] = Operands[1].getOperand(1);
25829 // Now we have three operands of two additions. Check that one of them is a
25830 // constant vector with ones, and the other two are promoted from i8/i16.
25831 for (int i = 0; i < 3; ++i) {
25832 if (!IsConstVectorInRange(Operands[i], 1, 1))
25834 std::swap(Operands[i], Operands[2]);
25836 // Check if Operands[0] and Operands[1] are results of type promotion.
25837 for (int j = 0; j < 2; ++j)
25838 if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
25839 Operands[j].getOperand(0).getValueType() != VT)
25842 // The pattern is detected, emit X86ISD::AVG instruction.
25843 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
25844 Operands[1].getOperand(0));
25850 static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
25851 const X86Subtarget *Subtarget) {
25852 return detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG, Subtarget,
25856 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25857 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25858 TargetLowering::DAGCombinerInfo &DCI,
25859 const X86Subtarget *Subtarget) {
25860 LoadSDNode *Ld = cast<LoadSDNode>(N);
25861 EVT RegVT = Ld->getValueType(0);
25862 EVT MemVT = Ld->getMemoryVT();
25864 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25866 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25867 // into two 16-byte operations.
25868 ISD::LoadExtType Ext = Ld->getExtensionType();
25870 unsigned AddressSpace = Ld->getAddressSpace();
25871 unsigned Alignment = Ld->getAlignment();
25872 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25873 Ext == ISD::NON_EXTLOAD &&
25874 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25875 AddressSpace, Alignment, &Fast) && !Fast) {
25876 unsigned NumElems = RegVT.getVectorNumElements();
25880 SDValue Ptr = Ld->getBasePtr();
25881 SDValue Increment =
25882 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25884 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25886 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25887 Ld->getPointerInfo(), Ld->isVolatile(),
25888 Ld->isNonTemporal(), Ld->isInvariant(),
25890 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25891 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25892 Ld->getPointerInfo(), Ld->isVolatile(),
25893 Ld->isNonTemporal(), Ld->isInvariant(),
25894 std::min(16U, Alignment));
25895 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25897 Load2.getValue(1));
25899 SDValue NewVec = DAG.getUNDEF(RegVT);
25900 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25901 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25902 return DCI.CombineTo(N, NewVec, TF, true);
25908 /// PerformMLOADCombine - Resolve extending loads
25909 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25910 TargetLowering::DAGCombinerInfo &DCI,
25911 const X86Subtarget *Subtarget) {
25912 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25913 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25916 EVT VT = Mld->getValueType(0);
25917 unsigned NumElems = VT.getVectorNumElements();
25918 EVT LdVT = Mld->getMemoryVT();
25921 assert(LdVT != VT && "Cannot extend to the same type");
25922 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25923 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25924 // From, To sizes and ElemCount must be pow of two
25925 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25926 "Unexpected size for extending masked load");
25928 unsigned SizeRatio = ToSz / FromSz;
25929 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25931 // Create a type on which we perform the shuffle
25932 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25933 LdVT.getScalarType(), NumElems*SizeRatio);
25934 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25936 // Convert Src0 value
25937 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25938 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25939 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25940 for (unsigned i = 0; i != NumElems; ++i)
25941 ShuffleVec[i] = i * SizeRatio;
25943 // Can't shuffle using an illegal type.
25944 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25945 "WideVecVT should be legal");
25946 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25947 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25949 // Prepare the new mask
25951 SDValue Mask = Mld->getMask();
25952 if (Mask.getValueType() == VT) {
25953 // Mask and original value have the same type
25954 NewMask = DAG.getBitcast(WideVecVT, Mask);
25955 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25956 for (unsigned i = 0; i != NumElems; ++i)
25957 ShuffleVec[i] = i * SizeRatio;
25958 for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
25959 ShuffleVec[i] = NumElems * SizeRatio;
25960 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25961 DAG.getConstant(0, dl, WideVecVT),
25965 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25966 unsigned WidenNumElts = NumElems*SizeRatio;
25967 unsigned MaskNumElts = VT.getVectorNumElements();
25968 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25971 unsigned NumConcat = WidenNumElts / MaskNumElts;
25972 SmallVector<SDValue, 16> Ops(NumConcat);
25973 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25975 for (unsigned i = 1; i != NumConcat; ++i)
25978 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25981 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25982 Mld->getBasePtr(), NewMask, WideSrc0,
25983 Mld->getMemoryVT(), Mld->getMemOperand(),
25985 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25986 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25988 /// PerformMSTORECombine - Resolve truncating stores
25989 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25990 const X86Subtarget *Subtarget) {
25991 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25992 if (!Mst->isTruncatingStore())
25995 EVT VT = Mst->getValue().getValueType();
25996 unsigned NumElems = VT.getVectorNumElements();
25997 EVT StVT = Mst->getMemoryVT();
26000 assert(StVT != VT && "Cannot truncate to the same type");
26001 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26002 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26004 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26006 // The truncating store is legal in some cases. For example
26007 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26008 // are designated for truncate store.
26009 // In this case we don't need any further transformations.
26010 if (TLI.isTruncStoreLegal(VT, StVT))
26013 // From, To sizes and ElemCount must be pow of two
26014 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26015 "Unexpected size for truncating masked store");
26016 // We are going to use the original vector elt for storing.
26017 // Accumulated smaller vector elements must be a multiple of the store size.
26018 assert (((NumElems * FromSz) % ToSz) == 0 &&
26019 "Unexpected ratio for truncating masked store");
26021 unsigned SizeRatio = FromSz / ToSz;
26022 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26024 // Create a type on which we perform the shuffle
26025 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26026 StVT.getScalarType(), NumElems*SizeRatio);
26028 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26030 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
26031 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26032 for (unsigned i = 0; i != NumElems; ++i)
26033 ShuffleVec[i] = i * SizeRatio;
26035 // Can't shuffle using an illegal type.
26036 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26037 "WideVecVT should be legal");
26039 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26040 DAG.getUNDEF(WideVecVT),
26044 SDValue Mask = Mst->getMask();
26045 if (Mask.getValueType() == VT) {
26046 // Mask and original value have the same type
26047 NewMask = DAG.getBitcast(WideVecVT, Mask);
26048 for (unsigned i = 0; i != NumElems; ++i)
26049 ShuffleVec[i] = i * SizeRatio;
26050 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
26051 ShuffleVec[i] = NumElems*SizeRatio;
26052 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26053 DAG.getConstant(0, dl, WideVecVT),
26057 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26058 unsigned WidenNumElts = NumElems*SizeRatio;
26059 unsigned MaskNumElts = VT.getVectorNumElements();
26060 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26063 unsigned NumConcat = WidenNumElts / MaskNumElts;
26064 SmallVector<SDValue, 16> Ops(NumConcat);
26065 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26067 for (unsigned i = 1; i != NumConcat; ++i)
26070 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26073 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
26074 Mst->getBasePtr(), NewMask, StVT,
26075 Mst->getMemOperand(), false);
26077 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
26078 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
26079 const X86Subtarget *Subtarget) {
26080 StoreSDNode *St = cast<StoreSDNode>(N);
26081 EVT VT = St->getValue().getValueType();
26082 EVT StVT = St->getMemoryVT();
26084 SDValue StoredVal = St->getOperand(1);
26085 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26087 // If we are saving a concatenation of two XMM registers and 32-byte stores
26088 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
26090 unsigned AddressSpace = St->getAddressSpace();
26091 unsigned Alignment = St->getAlignment();
26092 if (VT.is256BitVector() && StVT == VT &&
26093 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
26094 AddressSpace, Alignment, &Fast) && !Fast) {
26095 unsigned NumElems = VT.getVectorNumElements();
26099 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
26100 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
26103 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26104 SDValue Ptr0 = St->getBasePtr();
26105 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
26107 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
26108 St->getPointerInfo(), St->isVolatile(),
26109 St->isNonTemporal(), Alignment);
26110 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
26111 St->getPointerInfo(), St->isVolatile(),
26112 St->isNonTemporal(),
26113 std::min(16U, Alignment));
26114 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
26117 // Optimize trunc store (of multiple scalars) to shuffle and store.
26118 // First, pack all of the elements in one place. Next, store to memory
26119 // in fewer chunks.
26120 if (St->isTruncatingStore() && VT.isVector()) {
26121 // Check if we can detect an AVG pattern from the truncation. If yes,
26122 // replace the trunc store by a normal store with the result of X86ISD::AVG
26125 detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
26127 return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
26128 St->getPointerInfo(), St->isVolatile(),
26129 St->isNonTemporal(), St->getAlignment());
26131 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26132 unsigned NumElems = VT.getVectorNumElements();
26133 assert(StVT != VT && "Cannot truncate to the same type");
26134 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26135 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26137 // The truncating store is legal in some cases. For example
26138 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26139 // are designated for truncate store.
26140 // In this case we don't need any further transformations.
26141 if (TLI.isTruncStoreLegal(VT, StVT))
26144 // From, To sizes and ElemCount must be pow of two
26145 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
26146 // We are going to use the original vector elt for storing.
26147 // Accumulated smaller vector elements must be a multiple of the store size.
26148 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
26150 unsigned SizeRatio = FromSz / ToSz;
26152 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26154 // Create a type on which we perform the shuffle
26155 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26156 StVT.getScalarType(), NumElems*SizeRatio);
26158 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26160 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
26161 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
26162 for (unsigned i = 0; i != NumElems; ++i)
26163 ShuffleVec[i] = i * SizeRatio;
26165 // Can't shuffle using an illegal type.
26166 if (!TLI.isTypeLegal(WideVecVT))
26169 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26170 DAG.getUNDEF(WideVecVT),
26172 // At this point all of the data is stored at the bottom of the
26173 // register. We now need to save it to mem.
26175 // Find the largest store unit
26176 MVT StoreType = MVT::i8;
26177 for (MVT Tp : MVT::integer_valuetypes()) {
26178 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
26182 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
26183 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
26184 (64 <= NumElems * ToSz))
26185 StoreType = MVT::f64;
26187 // Bitcast the original vector into a vector of store-size units
26188 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
26189 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
26190 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
26191 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
26192 SmallVector<SDValue, 8> Chains;
26193 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
26194 TLI.getPointerTy(DAG.getDataLayout()));
26195 SDValue Ptr = St->getBasePtr();
26197 // Perform one or more big stores into memory.
26198 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
26199 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
26200 StoreType, ShuffWide,
26201 DAG.getIntPtrConstant(i, dl));
26202 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
26203 St->getPointerInfo(), St->isVolatile(),
26204 St->isNonTemporal(), St->getAlignment());
26205 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26206 Chains.push_back(Ch);
26209 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
26212 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
26213 // the FP state in cases where an emms may be missing.
26214 // A preferable solution to the general problem is to figure out the right
26215 // places to insert EMMS. This qualifies as a quick hack.
26217 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
26218 if (VT.getSizeInBits() != 64)
26221 const Function *F = DAG.getMachineFunction().getFunction();
26222 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
26224 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
26225 if ((VT.isVector() ||
26226 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
26227 isa<LoadSDNode>(St->getValue()) &&
26228 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
26229 St->getChain().hasOneUse() && !St->isVolatile()) {
26230 SDNode* LdVal = St->getValue().getNode();
26231 LoadSDNode *Ld = nullptr;
26232 int TokenFactorIndex = -1;
26233 SmallVector<SDValue, 8> Ops;
26234 SDNode* ChainVal = St->getChain().getNode();
26235 // Must be a store of a load. We currently handle two cases: the load
26236 // is a direct child, and it's under an intervening TokenFactor. It is
26237 // possible to dig deeper under nested TokenFactors.
26238 if (ChainVal == LdVal)
26239 Ld = cast<LoadSDNode>(St->getChain());
26240 else if (St->getValue().hasOneUse() &&
26241 ChainVal->getOpcode() == ISD::TokenFactor) {
26242 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
26243 if (ChainVal->getOperand(i).getNode() == LdVal) {
26244 TokenFactorIndex = i;
26245 Ld = cast<LoadSDNode>(St->getValue());
26247 Ops.push_back(ChainVal->getOperand(i));
26251 if (!Ld || !ISD::isNormalLoad(Ld))
26254 // If this is not the MMX case, i.e. we are just turning i64 load/store
26255 // into f64 load/store, avoid the transformation if there are multiple
26256 // uses of the loaded value.
26257 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
26262 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
26263 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
26265 if (Subtarget->is64Bit() || F64IsLegal) {
26266 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
26267 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
26268 Ld->getPointerInfo(), Ld->isVolatile(),
26269 Ld->isNonTemporal(), Ld->isInvariant(),
26270 Ld->getAlignment());
26271 SDValue NewChain = NewLd.getValue(1);
26272 if (TokenFactorIndex != -1) {
26273 Ops.push_back(NewChain);
26274 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26276 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
26277 St->getPointerInfo(),
26278 St->isVolatile(), St->isNonTemporal(),
26279 St->getAlignment());
26282 // Otherwise, lower to two pairs of 32-bit loads / stores.
26283 SDValue LoAddr = Ld->getBasePtr();
26284 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
26285 DAG.getConstant(4, LdDL, MVT::i32));
26287 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
26288 Ld->getPointerInfo(),
26289 Ld->isVolatile(), Ld->isNonTemporal(),
26290 Ld->isInvariant(), Ld->getAlignment());
26291 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
26292 Ld->getPointerInfo().getWithOffset(4),
26293 Ld->isVolatile(), Ld->isNonTemporal(),
26295 MinAlign(Ld->getAlignment(), 4));
26297 SDValue NewChain = LoLd.getValue(1);
26298 if (TokenFactorIndex != -1) {
26299 Ops.push_back(LoLd);
26300 Ops.push_back(HiLd);
26301 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26304 LoAddr = St->getBasePtr();
26305 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
26306 DAG.getConstant(4, StDL, MVT::i32));
26308 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
26309 St->getPointerInfo(),
26310 St->isVolatile(), St->isNonTemporal(),
26311 St->getAlignment());
26312 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
26313 St->getPointerInfo().getWithOffset(4),
26315 St->isNonTemporal(),
26316 MinAlign(St->getAlignment(), 4));
26317 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
26320 // This is similar to the above case, but here we handle a scalar 64-bit
26321 // integer store that is extracted from a vector on a 32-bit target.
26322 // If we have SSE2, then we can treat it like a floating-point double
26323 // to get past legalization. The execution dependencies fixup pass will
26324 // choose the optimal machine instruction for the store if this really is
26325 // an integer or v2f32 rather than an f64.
26326 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
26327 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
26328 SDValue OldExtract = St->getOperand(1);
26329 SDValue ExtOp0 = OldExtract.getOperand(0);
26330 unsigned VecSize = ExtOp0.getValueSizeInBits();
26331 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
26332 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
26333 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
26334 BitCast, OldExtract.getOperand(1));
26335 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
26336 St->getPointerInfo(), St->isVolatile(),
26337 St->isNonTemporal(), St->getAlignment());
26343 /// Return 'true' if this vector operation is "horizontal"
26344 /// and return the operands for the horizontal operation in LHS and RHS. A
26345 /// horizontal operation performs the binary operation on successive elements
26346 /// of its first operand, then on successive elements of its second operand,
26347 /// returning the resulting values in a vector. For example, if
26348 /// A = < float a0, float a1, float a2, float a3 >
26350 /// B = < float b0, float b1, float b2, float b3 >
26351 /// then the result of doing a horizontal operation on A and B is
26352 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
26353 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
26354 /// A horizontal-op B, for some already available A and B, and if so then LHS is
26355 /// set to A, RHS to B, and the routine returns 'true'.
26356 /// Note that the binary operation should have the property that if one of the
26357 /// operands is UNDEF then the result is UNDEF.
26358 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
26359 // Look for the following pattern: if
26360 // A = < float a0, float a1, float a2, float a3 >
26361 // B = < float b0, float b1, float b2, float b3 >
26363 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
26364 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
26365 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
26366 // which is A horizontal-op B.
26368 // At least one of the operands should be a vector shuffle.
26369 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
26370 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
26373 MVT VT = LHS.getSimpleValueType();
26375 assert((VT.is128BitVector() || VT.is256BitVector()) &&
26376 "Unsupported vector type for horizontal add/sub");
26378 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
26379 // operate independently on 128-bit lanes.
26380 unsigned NumElts = VT.getVectorNumElements();
26381 unsigned NumLanes = VT.getSizeInBits()/128;
26382 unsigned NumLaneElts = NumElts / NumLanes;
26383 assert((NumLaneElts % 2 == 0) &&
26384 "Vector type should have an even number of elements in each lane");
26385 unsigned HalfLaneElts = NumLaneElts/2;
26387 // View LHS in the form
26388 // LHS = VECTOR_SHUFFLE A, B, LMask
26389 // If LHS is not a shuffle then pretend it is the shuffle
26390 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
26391 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
26394 SmallVector<int, 16> LMask(NumElts);
26395 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26396 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
26397 A = LHS.getOperand(0);
26398 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
26399 B = LHS.getOperand(1);
26400 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
26401 std::copy(Mask.begin(), Mask.end(), LMask.begin());
26403 if (LHS.getOpcode() != ISD::UNDEF)
26405 for (unsigned i = 0; i != NumElts; ++i)
26409 // Likewise, view RHS in the form
26410 // RHS = VECTOR_SHUFFLE C, D, RMask
26412 SmallVector<int, 16> RMask(NumElts);
26413 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26414 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
26415 C = RHS.getOperand(0);
26416 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
26417 D = RHS.getOperand(1);
26418 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
26419 std::copy(Mask.begin(), Mask.end(), RMask.begin());
26421 if (RHS.getOpcode() != ISD::UNDEF)
26423 for (unsigned i = 0; i != NumElts; ++i)
26427 // Check that the shuffles are both shuffling the same vectors.
26428 if (!(A == C && B == D) && !(A == D && B == C))
26431 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
26432 if (!A.getNode() && !B.getNode())
26435 // If A and B occur in reverse order in RHS, then "swap" them (which means
26436 // rewriting the mask).
26438 ShuffleVectorSDNode::commuteMask(RMask);
26440 // At this point LHS and RHS are equivalent to
26441 // LHS = VECTOR_SHUFFLE A, B, LMask
26442 // RHS = VECTOR_SHUFFLE A, B, RMask
26443 // Check that the masks correspond to performing a horizontal operation.
26444 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
26445 for (unsigned i = 0; i != NumLaneElts; ++i) {
26446 int LIdx = LMask[i+l], RIdx = RMask[i+l];
26448 // Ignore any UNDEF components.
26449 if (LIdx < 0 || RIdx < 0 ||
26450 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
26451 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
26454 // Check that successive elements are being operated on. If not, this is
26455 // not a horizontal operation.
26456 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
26457 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
26458 if (!(LIdx == Index && RIdx == Index + 1) &&
26459 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
26464 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
26465 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
26469 /// Do target-specific dag combines on floating point adds.
26470 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
26471 const X86Subtarget *Subtarget) {
26472 EVT VT = N->getValueType(0);
26473 SDValue LHS = N->getOperand(0);
26474 SDValue RHS = N->getOperand(1);
26476 // Try to synthesize horizontal adds from adds of shuffles.
26477 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26478 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26479 isHorizontalBinOp(LHS, RHS, true))
26480 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
26484 /// Do target-specific dag combines on floating point subs.
26485 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
26486 const X86Subtarget *Subtarget) {
26487 EVT VT = N->getValueType(0);
26488 SDValue LHS = N->getOperand(0);
26489 SDValue RHS = N->getOperand(1);
26491 // Try to synthesize horizontal subs from subs of shuffles.
26492 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26493 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26494 isHorizontalBinOp(LHS, RHS, false))
26495 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
26499 /// Do target-specific dag combines on floating point negations.
26500 static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
26501 const X86Subtarget *Subtarget) {
26502 EVT VT = N->getValueType(0);
26503 EVT SVT = VT.getScalarType();
26504 SDValue Arg = N->getOperand(0);
26507 // Let legalize expand this if it isn't a legal type yet.
26508 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26511 // If we're negating a FMUL node on a target with FMA, then we can avoid the
26512 // use of a constant by performing (-0 - A*B) instead.
26513 // FIXME: Check rounding control flags as well once it becomes available.
26514 if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) &&
26515 Arg->getFlags()->hasNoSignedZeros() && Subtarget->hasAnyFMA()) {
26516 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
26517 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26518 Arg.getOperand(1), Zero);
26521 // If we're negating a FMA node, then we can adjust the
26522 // instruction to include the extra negation.
26523 if (Arg.hasOneUse()) {
26524 switch (Arg.getOpcode()) {
26525 case X86ISD::FMADD:
26526 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26527 Arg.getOperand(1), Arg.getOperand(2));
26528 case X86ISD::FMSUB:
26529 return DAG.getNode(X86ISD::FNMADD, DL, VT, Arg.getOperand(0),
26530 Arg.getOperand(1), Arg.getOperand(2));
26531 case X86ISD::FNMADD:
26532 return DAG.getNode(X86ISD::FMSUB, DL, VT, Arg.getOperand(0),
26533 Arg.getOperand(1), Arg.getOperand(2));
26534 case X86ISD::FNMSUB:
26535 return DAG.getNode(X86ISD::FMADD, DL, VT, Arg.getOperand(0),
26536 Arg.getOperand(1), Arg.getOperand(2));
26542 static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
26543 const X86Subtarget *Subtarget) {
26544 EVT VT = N->getValueType(0);
26545 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
26546 // VXORPS, VORPS, VANDPS, VANDNPS are supported only under DQ extention.
26547 // These logic operations may be executed in the integer domain.
26549 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
26550 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
26552 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
26553 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
26554 unsigned IntOpcode = 0;
26555 switch (N->getOpcode()) {
26556 default: llvm_unreachable("Unexpected FP logic op");
26557 case X86ISD::FOR: IntOpcode = ISD::OR; break;
26558 case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
26559 case X86ISD::FAND: IntOpcode = ISD::AND; break;
26560 case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
26562 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
26563 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
26567 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
26568 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
26569 const X86Subtarget *Subtarget) {
26570 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
26572 // F[X]OR(0.0, x) -> x
26573 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26574 if (C->getValueAPF().isPosZero())
26575 return N->getOperand(1);
26577 // F[X]OR(x, 0.0) -> x
26578 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26579 if (C->getValueAPF().isPosZero())
26580 return N->getOperand(0);
26582 return lowerX86FPLogicOp(N, DAG, Subtarget);
26585 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
26586 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
26587 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
26589 // Only perform optimizations if UnsafeMath is used.
26590 if (!DAG.getTarget().Options.UnsafeFPMath)
26593 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
26594 // into FMINC and FMAXC, which are Commutative operations.
26595 unsigned NewOp = 0;
26596 switch (N->getOpcode()) {
26597 default: llvm_unreachable("unknown opcode");
26598 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
26599 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
26602 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
26603 N->getOperand(0), N->getOperand(1));
26606 static SDValue performFMaxNumCombine(SDNode *N, SelectionDAG &DAG,
26607 const X86Subtarget *Subtarget) {
26608 // This takes at least 3 instructions, so favor a library call when
26609 // minimizing code size.
26610 if (DAG.getMachineFunction().getFunction()->optForMinSize())
26613 EVT VT = N->getValueType(0);
26615 // TODO: Check for global or instruction-level "nnan". In that case, we
26616 // should be able to lower to FMAX/FMIN alone.
26617 // TODO: If an operand is already known to be a NaN or not a NaN, this
26618 // should be an optional swap and FMAX/FMIN.
26619 // TODO: Allow f64, vectors, and fminnum.
26621 if (VT != MVT::f32 || !Subtarget->hasSSE1() || Subtarget->useSoftFloat())
26624 SDValue Op0 = N->getOperand(0);
26625 SDValue Op1 = N->getOperand(1);
26627 EVT SetCCType = DAG.getTargetLoweringInfo().getSetCCResultType(
26628 DAG.getDataLayout(), *DAG.getContext(), VT);
26630 // There are 4 possibilities involving NaN inputs, and these are the required
26634 // ----------------
26635 // Num | Max | Op0 |
26636 // Op0 ----------------
26637 // NaN | Op1 | NaN |
26638 // ----------------
26640 // The SSE FP max/min instructions were not designed for this case, but rather
26642 // Max = Op1 > Op0 ? Op1 : Op0
26644 // So they always return Op0 if either input is a NaN. However, we can still
26645 // use those instructions for fmaxnum by selecting away a NaN input.
26647 // If either operand is NaN, the 2nd source operand (Op0) is passed through.
26648 SDValue Max = DAG.getNode(X86ISD::FMAX, DL, VT, Op1, Op0);
26649 SDValue IsOp0Nan = DAG.getSetCC(DL, SetCCType , Op0, Op0, ISD::SETUO);
26651 // If Op0 is a NaN, select Op1. Otherwise, select the max. If both operands
26652 // are NaN, the NaN value of Op1 is the result.
26653 return DAG.getNode(ISD::SELECT, DL, VT, IsOp0Nan, Op1, Max);
26656 /// Do target-specific dag combines on X86ISD::FAND nodes.
26657 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG,
26658 const X86Subtarget *Subtarget) {
26659 // FAND(0.0, x) -> 0.0
26660 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26661 if (C->getValueAPF().isPosZero())
26662 return N->getOperand(0);
26664 // FAND(x, 0.0) -> 0.0
26665 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26666 if (C->getValueAPF().isPosZero())
26667 return N->getOperand(1);
26669 return lowerX86FPLogicOp(N, DAG, Subtarget);
26672 /// Do target-specific dag combines on X86ISD::FANDN nodes
26673 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG,
26674 const X86Subtarget *Subtarget) {
26675 // FANDN(0.0, x) -> x
26676 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26677 if (C->getValueAPF().isPosZero())
26678 return N->getOperand(1);
26680 // FANDN(x, 0.0) -> 0.0
26681 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26682 if (C->getValueAPF().isPosZero())
26683 return N->getOperand(1);
26685 return lowerX86FPLogicOp(N, DAG, Subtarget);
26688 static SDValue PerformBTCombine(SDNode *N,
26690 TargetLowering::DAGCombinerInfo &DCI) {
26691 // BT ignores high bits in the bit index operand.
26692 SDValue Op1 = N->getOperand(1);
26693 if (Op1.hasOneUse()) {
26694 unsigned BitWidth = Op1.getValueSizeInBits();
26695 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
26696 APInt KnownZero, KnownOne;
26697 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
26698 !DCI.isBeforeLegalizeOps());
26699 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26700 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
26701 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
26702 DCI.CommitTargetLoweringOpt(TLO);
26707 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
26708 SDValue Op = N->getOperand(0);
26709 if (Op.getOpcode() == ISD::BITCAST)
26710 Op = Op.getOperand(0);
26711 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
26712 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
26713 VT.getVectorElementType().getSizeInBits() ==
26714 OpVT.getVectorElementType().getSizeInBits()) {
26715 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
26720 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
26721 const X86Subtarget *Subtarget) {
26722 EVT VT = N->getValueType(0);
26723 if (!VT.isVector())
26726 SDValue N0 = N->getOperand(0);
26727 SDValue N1 = N->getOperand(1);
26728 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
26731 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
26732 // both SSE and AVX2 since there is no sign-extended shift right
26733 // operation on a vector with 64-bit elements.
26734 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
26735 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
26736 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
26737 N0.getOpcode() == ISD::SIGN_EXTEND)) {
26738 SDValue N00 = N0.getOperand(0);
26740 // EXTLOAD has a better solution on AVX2,
26741 // it may be replaced with X86ISD::VSEXT node.
26742 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
26743 if (!ISD::isNormalLoad(N00.getNode()))
26746 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
26747 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
26749 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
26755 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
26756 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
26757 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
26758 /// eliminate extend, add, and shift instructions.
26759 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
26760 const X86Subtarget *Subtarget) {
26761 // TODO: This should be valid for other integer types.
26762 EVT VT = Sext->getValueType(0);
26763 if (VT != MVT::i64)
26766 // We need an 'add nsw' feeding into the 'sext'.
26767 SDValue Add = Sext->getOperand(0);
26768 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
26771 // Having a constant operand to the 'add' ensures that we are not increasing
26772 // the instruction count because the constant is extended for free below.
26773 // A constant operand can also become the displacement field of an LEA.
26774 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
26778 // Don't make the 'add' bigger if there's no hope of combining it with some
26779 // other 'add' or 'shl' instruction.
26780 // TODO: It may be profitable to generate simpler LEA instructions in place
26781 // of single 'add' instructions, but the cost model for selecting an LEA
26782 // currently has a high threshold.
26783 bool HasLEAPotential = false;
26784 for (auto *User : Sext->uses()) {
26785 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
26786 HasLEAPotential = true;
26790 if (!HasLEAPotential)
26793 // Everything looks good, so pull the 'sext' ahead of the 'add'.
26794 int64_t AddConstant = AddOp1->getSExtValue();
26795 SDValue AddOp0 = Add.getOperand(0);
26796 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
26797 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
26799 // The wider add is guaranteed to not wrap because both operands are
26802 Flags.setNoSignedWrap(true);
26803 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
26806 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
26807 TargetLowering::DAGCombinerInfo &DCI,
26808 const X86Subtarget *Subtarget) {
26809 SDValue N0 = N->getOperand(0);
26810 EVT VT = N->getValueType(0);
26811 EVT SVT = VT.getScalarType();
26812 EVT InVT = N0.getValueType();
26813 EVT InSVT = InVT.getScalarType();
26816 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26817 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26818 // This exposes the sext to the sdivrem lowering, so that it directly extends
26819 // from AH (which we otherwise need to do contortions to access).
26820 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26821 InVT == MVT::i8 && VT == MVT::i32) {
26822 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26823 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26824 N0.getOperand(0), N0.getOperand(1));
26825 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26826 return R.getValue(1);
26829 if (!DCI.isBeforeLegalizeOps()) {
26830 if (InVT == MVT::i1) {
26831 SDValue Zero = DAG.getConstant(0, DL, VT);
26833 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26834 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26839 if (VT.isVector() && Subtarget->hasSSE2()) {
26840 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26841 EVT InVT = N.getValueType();
26842 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26843 Size / InVT.getScalarSizeInBits());
26844 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26845 DAG.getUNDEF(InVT));
26847 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26850 // If target-size is less than 128-bits, extend to a type that would extend
26851 // to 128 bits, extend that and extract the original target vector.
26852 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26853 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26854 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26855 unsigned Scale = 128 / VT.getSizeInBits();
26857 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26858 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26859 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26860 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26861 DAG.getIntPtrConstant(0, DL));
26864 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26865 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26866 if (VT.getSizeInBits() == 128 &&
26867 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26868 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26869 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26870 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26873 // On pre-AVX2 targets, split into 128-bit nodes of
26874 // ISD::SIGN_EXTEND_VECTOR_INREG.
26875 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26876 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26877 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26878 unsigned NumVecs = VT.getSizeInBits() / 128;
26879 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26880 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26881 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26883 SmallVector<SDValue, 8> Opnds;
26884 for (unsigned i = 0, Offset = 0; i != NumVecs;
26885 ++i, Offset += NumSubElts) {
26886 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26887 DAG.getIntPtrConstant(Offset, DL));
26888 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26889 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26890 Opnds.push_back(SrcVec);
26892 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26896 if (Subtarget->hasAVX() && VT.is256BitVector())
26897 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26900 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26906 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26907 const X86Subtarget* Subtarget) {
26909 EVT VT = N->getValueType(0);
26911 // Let legalize expand this if it isn't a legal type yet.
26912 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26915 EVT ScalarVT = VT.getScalarType();
26916 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
26919 SDValue A = N->getOperand(0);
26920 SDValue B = N->getOperand(1);
26921 SDValue C = N->getOperand(2);
26923 bool NegA = (A.getOpcode() == ISD::FNEG);
26924 bool NegB = (B.getOpcode() == ISD::FNEG);
26925 bool NegC = (C.getOpcode() == ISD::FNEG);
26927 // Negative multiplication when NegA xor NegB
26928 bool NegMul = (NegA != NegB);
26930 A = A.getOperand(0);
26932 B = B.getOperand(0);
26934 C = C.getOperand(0);
26938 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26940 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26942 return DAG.getNode(Opcode, dl, VT, A, B, C);
26945 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26946 TargetLowering::DAGCombinerInfo &DCI,
26947 const X86Subtarget *Subtarget) {
26948 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26949 // (and (i32 x86isd::setcc_carry), 1)
26950 // This eliminates the zext. This transformation is necessary because
26951 // ISD::SETCC is always legalized to i8.
26953 SDValue N0 = N->getOperand(0);
26954 EVT VT = N->getValueType(0);
26956 if (N0.getOpcode() == ISD::AND &&
26958 N0.getOperand(0).hasOneUse()) {
26959 SDValue N00 = N0.getOperand(0);
26960 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26961 if (!isOneConstant(N0.getOperand(1)))
26963 return DAG.getNode(ISD::AND, dl, VT,
26964 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26965 N00.getOperand(0), N00.getOperand(1)),
26966 DAG.getConstant(1, dl, VT));
26970 if (N0.getOpcode() == ISD::TRUNCATE &&
26972 N0.getOperand(0).hasOneUse()) {
26973 SDValue N00 = N0.getOperand(0);
26974 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26975 return DAG.getNode(ISD::AND, dl, VT,
26976 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26977 N00.getOperand(0), N00.getOperand(1)),
26978 DAG.getConstant(1, dl, VT));
26982 if (VT.is256BitVector())
26983 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26986 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26987 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26988 // This exposes the zext to the udivrem lowering, so that it directly extends
26989 // from AH (which we otherwise need to do contortions to access).
26990 if (N0.getOpcode() == ISD::UDIVREM &&
26991 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26992 (VT == MVT::i32 || VT == MVT::i64)) {
26993 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26994 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26995 N0.getOperand(0), N0.getOperand(1));
26996 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26997 return R.getValue(1);
27003 // Optimize x == -y --> x+y == 0
27004 // x != -y --> x+y != 0
27005 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
27006 const X86Subtarget* Subtarget) {
27007 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
27008 SDValue LHS = N->getOperand(0);
27009 SDValue RHS = N->getOperand(1);
27010 EVT VT = N->getValueType(0);
27013 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
27014 if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
27015 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
27016 LHS.getOperand(1));
27017 return DAG.getSetCC(DL, N->getValueType(0), addV,
27018 DAG.getConstant(0, DL, addV.getValueType()), CC);
27020 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
27021 if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
27022 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
27023 RHS.getOperand(1));
27024 return DAG.getSetCC(DL, N->getValueType(0), addV,
27025 DAG.getConstant(0, DL, addV.getValueType()), CC);
27028 if (VT.getScalarType() == MVT::i1 &&
27029 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
27031 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27032 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27033 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27035 if (!IsSEXT0 || !IsVZero1) {
27036 // Swap the operands and update the condition code.
27037 std::swap(LHS, RHS);
27038 CC = ISD::getSetCCSwappedOperands(CC);
27040 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27041 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27042 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27045 if (IsSEXT0 && IsVZero1) {
27046 assert(VT == LHS.getOperand(0).getValueType() &&
27047 "Uexpected operand type");
27048 if (CC == ISD::SETGT)
27049 return DAG.getConstant(0, DL, VT);
27050 if (CC == ISD::SETLE)
27051 return DAG.getConstant(1, DL, VT);
27052 if (CC == ISD::SETEQ || CC == ISD::SETGE)
27053 return DAG.getNOT(DL, LHS.getOperand(0), VT);
27055 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
27056 "Unexpected condition code!");
27057 return LHS.getOperand(0);
27064 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
27065 SDValue V0 = N->getOperand(0);
27066 SDValue V1 = N->getOperand(1);
27068 EVT VT = N->getValueType(0);
27070 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
27071 // operands and changing the mask to 1. This saves us a bunch of
27072 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
27073 // x86InstrInfo knows how to commute this back after instruction selection
27074 // if it would help register allocation.
27076 // TODO: If optimizing for size or a processor that doesn't suffer from
27077 // partial register update stalls, this should be transformed into a MOVSD
27078 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
27080 if (VT == MVT::v2f64)
27081 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
27082 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
27083 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
27084 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
27090 static SDValue PerformGatherScatterCombine(SDNode *N, SelectionDAG &DAG) {
27092 // Gather and Scatter instructions use k-registers for masks. The type of
27093 // the masks is v*i1. So the mask will be truncated anyway.
27094 // The SIGN_EXTEND_INREG my be dropped.
27095 SDValue Mask = N->getOperand(2);
27096 if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) {
27097 SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end());
27098 NewOps[2] = Mask.getOperand(0);
27099 DAG.UpdateNodeOperands(N, NewOps);
27104 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
27105 // as "sbb reg,reg", since it can be extended without zext and produces
27106 // an all-ones bit which is more useful than 0/1 in some cases.
27107 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
27110 return DAG.getNode(ISD::AND, DL, VT,
27111 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27112 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27114 DAG.getConstant(1, DL, VT));
27115 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
27116 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
27117 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27118 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27122 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
27123 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
27124 TargetLowering::DAGCombinerInfo &DCI,
27125 const X86Subtarget *Subtarget) {
27127 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
27128 SDValue EFLAGS = N->getOperand(1);
27130 if (CC == X86::COND_A) {
27131 // Try to convert COND_A into COND_B in an attempt to facilitate
27132 // materializing "setb reg".
27134 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
27135 // cannot take an immediate as its first operand.
27137 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
27138 EFLAGS.getValueType().isInteger() &&
27139 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
27140 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
27141 EFLAGS.getNode()->getVTList(),
27142 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
27143 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
27144 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
27148 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
27149 // a zext and produces an all-ones bit which is more useful than 0/1 in some
27151 if (CC == X86::COND_B)
27152 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
27154 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27155 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27156 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
27162 // Optimize branch condition evaluation.
27164 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
27165 TargetLowering::DAGCombinerInfo &DCI,
27166 const X86Subtarget *Subtarget) {
27168 SDValue Chain = N->getOperand(0);
27169 SDValue Dest = N->getOperand(1);
27170 SDValue EFLAGS = N->getOperand(3);
27171 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
27173 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27174 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27175 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
27182 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
27183 SelectionDAG &DAG) {
27184 // Take advantage of vector comparisons producing 0 or -1 in each lane to
27185 // optimize away operation when it's from a constant.
27187 // The general transformation is:
27188 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
27189 // AND(VECTOR_CMP(x,y), constant2)
27190 // constant2 = UNARYOP(constant)
27192 // Early exit if this isn't a vector operation, the operand of the
27193 // unary operation isn't a bitwise AND, or if the sizes of the operations
27194 // aren't the same.
27195 EVT VT = N->getValueType(0);
27196 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
27197 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
27198 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
27201 // Now check that the other operand of the AND is a constant. We could
27202 // make the transformation for non-constant splats as well, but it's unclear
27203 // that would be a benefit as it would not eliminate any operations, just
27204 // perform one more step in scalar code before moving to the vector unit.
27205 if (BuildVectorSDNode *BV =
27206 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
27207 // Bail out if the vector isn't a constant.
27208 if (!BV->isConstant())
27211 // Everything checks out. Build up the new and improved node.
27213 EVT IntVT = BV->getValueType(0);
27214 // Create a new constant of the appropriate type for the transformed
27216 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
27217 // The AND node needs bitcasts to/from an integer vector type around it.
27218 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
27219 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
27220 N->getOperand(0)->getOperand(0), MaskConst);
27221 SDValue Res = DAG.getBitcast(VT, NewAnd);
27228 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27229 const X86Subtarget *Subtarget) {
27230 SDValue Op0 = N->getOperand(0);
27231 EVT VT = N->getValueType(0);
27232 EVT InVT = Op0.getValueType();
27233 EVT InSVT = InVT.getScalarType();
27234 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27236 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
27237 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
27238 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27240 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27241 InVT.getVectorNumElements());
27242 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
27244 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
27245 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
27247 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27253 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27254 const X86Subtarget *Subtarget) {
27255 // First try to optimize away the conversion entirely when it's
27256 // conditionally from a constant. Vectors only.
27257 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
27260 // Now move on to more general possibilities.
27261 SDValue Op0 = N->getOperand(0);
27262 EVT VT = N->getValueType(0);
27263 EVT InVT = Op0.getValueType();
27264 EVT InSVT = InVT.getScalarType();
27266 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
27267 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
27268 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27270 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27271 InVT.getVectorNumElements());
27272 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
27273 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27276 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
27277 // a 32-bit target where SSE doesn't support i64->FP operations.
27278 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
27279 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
27280 EVT LdVT = Ld->getValueType(0);
27282 // This transformation is not supported if the result type is f16
27283 if (VT == MVT::f16)
27286 if (!Ld->isVolatile() && !VT.isVector() &&
27287 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
27288 !Subtarget->is64Bit() && LdVT == MVT::i64) {
27289 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
27290 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
27291 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
27298 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
27299 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
27300 X86TargetLowering::DAGCombinerInfo &DCI) {
27301 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
27302 // the result is either zero or one (depending on the input carry bit).
27303 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
27304 if (X86::isZeroNode(N->getOperand(0)) &&
27305 X86::isZeroNode(N->getOperand(1)) &&
27306 // We don't have a good way to replace an EFLAGS use, so only do this when
27308 SDValue(N, 1).use_empty()) {
27310 EVT VT = N->getValueType(0);
27311 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
27312 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
27313 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
27314 DAG.getConstant(X86::COND_B, DL,
27317 DAG.getConstant(1, DL, VT));
27318 return DCI.CombineTo(N, Res1, CarryOut);
27324 // fold (add Y, (sete X, 0)) -> adc 0, Y
27325 // (add Y, (setne X, 0)) -> sbb -1, Y
27326 // (sub (sete X, 0), Y) -> sbb 0, Y
27327 // (sub (setne X, 0), Y) -> adc -1, Y
27328 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
27331 // Look through ZExts.
27332 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
27333 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
27336 SDValue SetCC = Ext.getOperand(0);
27337 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
27340 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
27341 if (CC != X86::COND_E && CC != X86::COND_NE)
27344 SDValue Cmp = SetCC.getOperand(1);
27345 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
27346 !X86::isZeroNode(Cmp.getOperand(1)) ||
27347 !Cmp.getOperand(0).getValueType().isInteger())
27350 SDValue CmpOp0 = Cmp.getOperand(0);
27351 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
27352 DAG.getConstant(1, DL, CmpOp0.getValueType()));
27354 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
27355 if (CC == X86::COND_NE)
27356 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
27357 DL, OtherVal.getValueType(), OtherVal,
27358 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
27360 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
27361 DL, OtherVal.getValueType(), OtherVal,
27362 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
27365 /// PerformADDCombine - Do target-specific dag combines on integer adds.
27366 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
27367 const X86Subtarget *Subtarget) {
27368 EVT VT = N->getValueType(0);
27369 SDValue Op0 = N->getOperand(0);
27370 SDValue Op1 = N->getOperand(1);
27372 // Try to synthesize horizontal adds from adds of shuffles.
27373 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27374 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27375 isHorizontalBinOp(Op0, Op1, true))
27376 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
27378 return OptimizeConditionalInDecrement(N, DAG);
27381 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
27382 const X86Subtarget *Subtarget) {
27383 SDValue Op0 = N->getOperand(0);
27384 SDValue Op1 = N->getOperand(1);
27386 // X86 can't encode an immediate LHS of a sub. See if we can push the
27387 // negation into a preceding instruction.
27388 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
27389 // If the RHS of the sub is a XOR with one use and a constant, invert the
27390 // immediate. Then add one to the LHS of the sub so we can turn
27391 // X-Y -> X+~Y+1, saving one register.
27392 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
27393 isa<ConstantSDNode>(Op1.getOperand(1))) {
27394 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
27395 EVT VT = Op0.getValueType();
27396 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
27398 DAG.getConstant(~XorC, SDLoc(Op1), VT));
27399 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
27400 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
27404 // Try to synthesize horizontal adds from adds of shuffles.
27405 EVT VT = N->getValueType(0);
27406 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27407 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27408 isHorizontalBinOp(Op0, Op1, true))
27409 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
27411 return OptimizeConditionalInDecrement(N, DAG);
27414 /// performVZEXTCombine - Performs build vector combines
27415 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
27416 TargetLowering::DAGCombinerInfo &DCI,
27417 const X86Subtarget *Subtarget) {
27419 MVT VT = N->getSimpleValueType(0);
27420 SDValue Op = N->getOperand(0);
27421 MVT OpVT = Op.getSimpleValueType();
27422 MVT OpEltVT = OpVT.getVectorElementType();
27423 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
27425 // (vzext (bitcast (vzext (x)) -> (vzext x)
27427 while (V.getOpcode() == ISD::BITCAST)
27428 V = V.getOperand(0);
27430 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
27431 MVT InnerVT = V.getSimpleValueType();
27432 MVT InnerEltVT = InnerVT.getVectorElementType();
27434 // If the element sizes match exactly, we can just do one larger vzext. This
27435 // is always an exact type match as vzext operates on integer types.
27436 if (OpEltVT == InnerEltVT) {
27437 assert(OpVT == InnerVT && "Types must match for vzext!");
27438 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
27441 // The only other way we can combine them is if only a single element of the
27442 // inner vzext is used in the input to the outer vzext.
27443 if (InnerEltVT.getSizeInBits() < InputBits)
27446 // In this case, the inner vzext is completely dead because we're going to
27447 // only look at bits inside of the low element. Just do the outer vzext on
27448 // a bitcast of the input to the inner.
27449 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
27452 // Check if we can bypass extracting and re-inserting an element of an input
27453 // vector. Essentially:
27454 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
27455 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
27456 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
27457 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
27458 SDValue ExtractedV = V.getOperand(0);
27459 SDValue OrigV = ExtractedV.getOperand(0);
27460 if (isNullConstant(ExtractedV.getOperand(1))) {
27461 MVT OrigVT = OrigV.getSimpleValueType();
27462 // Extract a subvector if necessary...
27463 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
27464 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
27465 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
27466 OrigVT.getVectorNumElements() / Ratio);
27467 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
27468 DAG.getIntPtrConstant(0, DL));
27470 Op = DAG.getBitcast(OpVT, OrigV);
27471 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
27478 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
27479 DAGCombinerInfo &DCI) const {
27480 SelectionDAG &DAG = DCI.DAG;
27481 switch (N->getOpcode()) {
27483 case ISD::EXTRACT_VECTOR_ELT:
27484 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
27487 case X86ISD::SHRUNKBLEND:
27488 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
27489 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
27490 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
27491 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
27492 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
27493 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
27494 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
27497 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
27498 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
27499 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
27500 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
27501 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
27502 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
27503 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
27504 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
27505 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
27506 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
27507 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
27508 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
27509 case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
27510 case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
27512 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
27514 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
27515 case ISD::FMAXNUM: return performFMaxNumCombine(N, DAG, Subtarget);
27516 case X86ISD::FAND: return PerformFANDCombine(N, DAG, Subtarget);
27517 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG, Subtarget);
27518 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
27519 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
27520 case ISD::ANY_EXTEND:
27521 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
27522 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
27523 case ISD::SIGN_EXTEND_INREG:
27524 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
27525 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
27526 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
27527 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
27528 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
27529 case X86ISD::SHUFP: // Handle all target specific shuffles
27530 case X86ISD::PALIGNR:
27531 case X86ISD::UNPCKH:
27532 case X86ISD::UNPCKL:
27533 case X86ISD::MOVHLPS:
27534 case X86ISD::MOVLHPS:
27535 case X86ISD::PSHUFB:
27536 case X86ISD::PSHUFD:
27537 case X86ISD::PSHUFHW:
27538 case X86ISD::PSHUFLW:
27539 case X86ISD::MOVSS:
27540 case X86ISD::MOVSD:
27541 case X86ISD::VPERMILPI:
27542 case X86ISD::VPERM2X128:
27543 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
27544 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
27545 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
27547 case ISD::MSCATTER: return PerformGatherScatterCombine(N, DAG);
27553 /// isTypeDesirableForOp - Return true if the target has native support for
27554 /// the specified value type and it is 'desirable' to use the type for the
27555 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
27556 /// instruction encodings are longer and some i16 instructions are slow.
27557 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
27558 if (!isTypeLegal(VT))
27560 if (VT != MVT::i16)
27567 case ISD::SIGN_EXTEND:
27568 case ISD::ZERO_EXTEND:
27569 case ISD::ANY_EXTEND:
27582 /// IsDesirableToPromoteOp - This method query the target whether it is
27583 /// beneficial for dag combiner to promote the specified node. If true, it
27584 /// should return the desired promotion type by reference.
27585 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
27586 EVT VT = Op.getValueType();
27587 if (VT != MVT::i16)
27590 bool Promote = false;
27591 bool Commute = false;
27592 switch (Op.getOpcode()) {
27595 LoadSDNode *LD = cast<LoadSDNode>(Op);
27596 // If the non-extending load has a single use and it's not live out, then it
27597 // might be folded.
27598 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
27599 Op.hasOneUse()*/) {
27600 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
27601 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
27602 // The only case where we'd want to promote LOAD (rather then it being
27603 // promoted as an operand is when it's only use is liveout.
27604 if (UI->getOpcode() != ISD::CopyToReg)
27611 case ISD::SIGN_EXTEND:
27612 case ISD::ZERO_EXTEND:
27613 case ISD::ANY_EXTEND:
27618 SDValue N0 = Op.getOperand(0);
27619 // Look out for (store (shl (load), x)).
27620 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
27633 SDValue N0 = Op.getOperand(0);
27634 SDValue N1 = Op.getOperand(1);
27635 if (!Commute && MayFoldLoad(N1))
27637 // Avoid disabling potential load folding opportunities.
27638 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
27640 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
27650 //===----------------------------------------------------------------------===//
27651 // X86 Inline Assembly Support
27652 //===----------------------------------------------------------------------===//
27654 // Helper to match a string separated by whitespace.
27655 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
27656 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
27658 for (StringRef Piece : Pieces) {
27659 if (!S.startswith(Piece)) // Check if the piece matches.
27662 S = S.substr(Piece.size());
27663 StringRef::size_type Pos = S.find_first_not_of(" \t");
27664 if (Pos == 0) // We matched a prefix.
27673 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
27675 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
27676 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
27677 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
27678 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
27680 if (AsmPieces.size() == 3)
27682 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
27689 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
27690 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
27692 std::string AsmStr = IA->getAsmString();
27694 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
27695 if (!Ty || Ty->getBitWidth() % 16 != 0)
27698 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
27699 SmallVector<StringRef, 4> AsmPieces;
27700 SplitString(AsmStr, AsmPieces, ";\n");
27702 switch (AsmPieces.size()) {
27703 default: return false;
27705 // FIXME: this should verify that we are targeting a 486 or better. If not,
27706 // we will turn this bswap into something that will be lowered to logical
27707 // ops instead of emitting the bswap asm. For now, we don't support 486 or
27708 // lower so don't worry about this.
27710 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
27711 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
27712 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
27713 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
27714 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
27715 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
27716 // No need to check constraints, nothing other than the equivalent of
27717 // "=r,0" would be valid here.
27718 return IntrinsicLowering::LowerToByteSwap(CI);
27721 // rorw $$8, ${0:w} --> llvm.bswap.i16
27722 if (CI->getType()->isIntegerTy(16) &&
27723 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27724 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
27725 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
27727 StringRef ConstraintsStr = IA->getConstraintString();
27728 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27729 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27730 if (clobbersFlagRegisters(AsmPieces))
27731 return IntrinsicLowering::LowerToByteSwap(CI);
27735 if (CI->getType()->isIntegerTy(32) &&
27736 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27737 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
27738 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
27739 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
27741 StringRef ConstraintsStr = IA->getConstraintString();
27742 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27743 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27744 if (clobbersFlagRegisters(AsmPieces))
27745 return IntrinsicLowering::LowerToByteSwap(CI);
27748 if (CI->getType()->isIntegerTy(64)) {
27749 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
27750 if (Constraints.size() >= 2 &&
27751 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
27752 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
27753 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
27754 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
27755 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
27756 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
27757 return IntrinsicLowering::LowerToByteSwap(CI);
27765 /// getConstraintType - Given a constraint letter, return the type of
27766 /// constraint it is for this target.
27767 X86TargetLowering::ConstraintType
27768 X86TargetLowering::getConstraintType(StringRef Constraint) const {
27769 if (Constraint.size() == 1) {
27770 switch (Constraint[0]) {
27781 return C_RegisterClass;
27805 return TargetLowering::getConstraintType(Constraint);
27808 /// Examine constraint type and operand type and determine a weight value.
27809 /// This object must already have been set up with the operand type
27810 /// and the current alternative constraint selected.
27811 TargetLowering::ConstraintWeight
27812 X86TargetLowering::getSingleConstraintMatchWeight(
27813 AsmOperandInfo &info, const char *constraint) const {
27814 ConstraintWeight weight = CW_Invalid;
27815 Value *CallOperandVal = info.CallOperandVal;
27816 // If we don't have a value, we can't do a match,
27817 // but allow it at the lowest weight.
27818 if (!CallOperandVal)
27820 Type *type = CallOperandVal->getType();
27821 // Look at the constraint type.
27822 switch (*constraint) {
27824 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27835 if (CallOperandVal->getType()->isIntegerTy())
27836 weight = CW_SpecificReg;
27841 if (type->isFloatingPointTy())
27842 weight = CW_SpecificReg;
27845 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27846 weight = CW_SpecificReg;
27850 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27851 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27852 weight = CW_Register;
27855 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27856 if (C->getZExtValue() <= 31)
27857 weight = CW_Constant;
27861 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27862 if (C->getZExtValue() <= 63)
27863 weight = CW_Constant;
27867 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27868 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27869 weight = CW_Constant;
27873 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27874 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27875 weight = CW_Constant;
27879 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27880 if (C->getZExtValue() <= 3)
27881 weight = CW_Constant;
27885 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27886 if (C->getZExtValue() <= 0xff)
27887 weight = CW_Constant;
27892 if (isa<ConstantFP>(CallOperandVal)) {
27893 weight = CW_Constant;
27897 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27898 if ((C->getSExtValue() >= -0x80000000LL) &&
27899 (C->getSExtValue() <= 0x7fffffffLL))
27900 weight = CW_Constant;
27904 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27905 if (C->getZExtValue() <= 0xffffffff)
27906 weight = CW_Constant;
27913 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27914 /// with another that has more specific requirements based on the type of the
27915 /// corresponding operand.
27916 const char *X86TargetLowering::
27917 LowerXConstraint(EVT ConstraintVT) const {
27918 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27919 // 'f' like normal targets.
27920 if (ConstraintVT.isFloatingPoint()) {
27921 if (Subtarget->hasSSE2())
27923 if (Subtarget->hasSSE1())
27927 return TargetLowering::LowerXConstraint(ConstraintVT);
27930 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27931 /// vector. If it is invalid, don't add anything to Ops.
27932 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27933 std::string &Constraint,
27934 std::vector<SDValue>&Ops,
27935 SelectionDAG &DAG) const {
27938 // Only support length 1 constraints for now.
27939 if (Constraint.length() > 1) return;
27941 char ConstraintLetter = Constraint[0];
27942 switch (ConstraintLetter) {
27945 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27946 if (C->getZExtValue() <= 31) {
27947 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27948 Op.getValueType());
27954 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27955 if (C->getZExtValue() <= 63) {
27956 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27957 Op.getValueType());
27963 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27964 if (isInt<8>(C->getSExtValue())) {
27965 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27966 Op.getValueType());
27972 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27973 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27974 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27975 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27976 Op.getValueType());
27982 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27983 if (C->getZExtValue() <= 3) {
27984 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27985 Op.getValueType());
27991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27992 if (C->getZExtValue() <= 255) {
27993 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27994 Op.getValueType());
28000 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28001 if (C->getZExtValue() <= 127) {
28002 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28003 Op.getValueType());
28009 // 32-bit signed value
28010 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28011 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28012 C->getSExtValue())) {
28013 // Widen to 64 bits here to get it sign extended.
28014 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
28017 // FIXME gcc accepts some relocatable values here too, but only in certain
28018 // memory models; it's complicated.
28023 // 32-bit unsigned value
28024 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28025 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28026 C->getZExtValue())) {
28027 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28028 Op.getValueType());
28032 // FIXME gcc accepts some relocatable values here too, but only in certain
28033 // memory models; it's complicated.
28037 // Literal immediates are always ok.
28038 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
28039 // Widen to 64 bits here to get it sign extended.
28040 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
28044 // In any sort of PIC mode addresses need to be computed at runtime by
28045 // adding in a register or some sort of table lookup. These can't
28046 // be used as immediates.
28047 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
28050 // If we are in non-pic codegen mode, we allow the address of a global (with
28051 // an optional displacement) to be used with 'i'.
28052 GlobalAddressSDNode *GA = nullptr;
28053 int64_t Offset = 0;
28055 // Match either (GA), (GA+C), (GA+C1+C2), etc.
28057 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
28058 Offset += GA->getOffset();
28060 } else if (Op.getOpcode() == ISD::ADD) {
28061 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28062 Offset += C->getZExtValue();
28063 Op = Op.getOperand(0);
28066 } else if (Op.getOpcode() == ISD::SUB) {
28067 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28068 Offset += -C->getZExtValue();
28069 Op = Op.getOperand(0);
28074 // Otherwise, this isn't something we can handle, reject it.
28078 const GlobalValue *GV = GA->getGlobal();
28079 // If we require an extra load to get this address, as in PIC mode, we
28080 // can't accept it.
28081 if (isGlobalStubReference(
28082 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
28085 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
28086 GA->getValueType(0), Offset);
28091 if (Result.getNode()) {
28092 Ops.push_back(Result);
28095 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
28098 std::pair<unsigned, const TargetRegisterClass *>
28099 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
28100 StringRef Constraint,
28102 // First, see if this is a constraint that directly corresponds to an LLVM
28104 if (Constraint.size() == 1) {
28105 // GCC Constraint Letters
28106 switch (Constraint[0]) {
28108 // TODO: Slight differences here in allocation order and leaving
28109 // RIP in the class. Do they matter any more here than they do
28110 // in the normal allocation?
28111 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
28112 if (Subtarget->is64Bit()) {
28113 if (VT == MVT::i32 || VT == MVT::f32)
28114 return std::make_pair(0U, &X86::GR32RegClass);
28115 if (VT == MVT::i16)
28116 return std::make_pair(0U, &X86::GR16RegClass);
28117 if (VT == MVT::i8 || VT == MVT::i1)
28118 return std::make_pair(0U, &X86::GR8RegClass);
28119 if (VT == MVT::i64 || VT == MVT::f64)
28120 return std::make_pair(0U, &X86::GR64RegClass);
28123 // 32-bit fallthrough
28124 case 'Q': // Q_REGS
28125 if (VT == MVT::i32 || VT == MVT::f32)
28126 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
28127 if (VT == MVT::i16)
28128 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
28129 if (VT == MVT::i8 || VT == MVT::i1)
28130 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
28131 if (VT == MVT::i64)
28132 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
28134 case 'r': // GENERAL_REGS
28135 case 'l': // INDEX_REGS
28136 if (VT == MVT::i8 || VT == MVT::i1)
28137 return std::make_pair(0U, &X86::GR8RegClass);
28138 if (VT == MVT::i16)
28139 return std::make_pair(0U, &X86::GR16RegClass);
28140 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
28141 return std::make_pair(0U, &X86::GR32RegClass);
28142 return std::make_pair(0U, &X86::GR64RegClass);
28143 case 'R': // LEGACY_REGS
28144 if (VT == MVT::i8 || VT == MVT::i1)
28145 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
28146 if (VT == MVT::i16)
28147 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
28148 if (VT == MVT::i32 || !Subtarget->is64Bit())
28149 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
28150 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
28151 case 'f': // FP Stack registers.
28152 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
28153 // value to the correct fpstack register class.
28154 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
28155 return std::make_pair(0U, &X86::RFP32RegClass);
28156 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
28157 return std::make_pair(0U, &X86::RFP64RegClass);
28158 return std::make_pair(0U, &X86::RFP80RegClass);
28159 case 'y': // MMX_REGS if MMX allowed.
28160 if (!Subtarget->hasMMX()) break;
28161 return std::make_pair(0U, &X86::VR64RegClass);
28162 case 'Y': // SSE_REGS if SSE2 allowed
28163 if (!Subtarget->hasSSE2()) break;
28165 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
28166 if (!Subtarget->hasSSE1()) break;
28168 switch (VT.SimpleTy) {
28170 // Scalar SSE types.
28173 return std::make_pair(0U, &X86::FR32RegClass);
28176 return std::make_pair(0U, &X86::FR64RegClass);
28177 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28185 return std::make_pair(0U, &X86::VR128RegClass);
28193 return std::make_pair(0U, &X86::VR256RegClass);
28198 return std::make_pair(0U, &X86::VR512RegClass);
28204 // Use the default implementation in TargetLowering to convert the register
28205 // constraint into a member of a register class.
28206 std::pair<unsigned, const TargetRegisterClass*> Res;
28207 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
28209 // Not found as a standard register?
28211 // Map st(0) -> st(7) -> ST0
28212 if (Constraint.size() == 7 && Constraint[0] == '{' &&
28213 tolower(Constraint[1]) == 's' &&
28214 tolower(Constraint[2]) == 't' &&
28215 Constraint[3] == '(' &&
28216 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
28217 Constraint[5] == ')' &&
28218 Constraint[6] == '}') {
28220 Res.first = X86::FP0+Constraint[4]-'0';
28221 Res.second = &X86::RFP80RegClass;
28225 // GCC allows "st(0)" to be called just plain "st".
28226 if (StringRef("{st}").equals_lower(Constraint)) {
28227 Res.first = X86::FP0;
28228 Res.second = &X86::RFP80RegClass;
28233 if (StringRef("{flags}").equals_lower(Constraint)) {
28234 Res.first = X86::EFLAGS;
28235 Res.second = &X86::CCRRegClass;
28239 // 'A' means EAX + EDX.
28240 if (Constraint == "A") {
28241 Res.first = X86::EAX;
28242 Res.second = &X86::GR32_ADRegClass;
28248 // Otherwise, check to see if this is a register class of the wrong value
28249 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
28250 // turn into {ax},{dx}.
28251 // MVT::Other is used to specify clobber names.
28252 if (Res.second->hasType(VT) || VT == MVT::Other)
28253 return Res; // Correct type already, nothing to do.
28255 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
28256 // return "eax". This should even work for things like getting 64bit integer
28257 // registers when given an f64 type.
28258 const TargetRegisterClass *Class = Res.second;
28259 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
28260 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
28261 unsigned Size = VT.getSizeInBits();
28262 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
28263 : Size == 16 ? MVT::i16
28264 : Size == 32 ? MVT::i32
28265 : Size == 64 ? MVT::i64
28267 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
28269 Res.first = DestReg;
28270 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
28271 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
28272 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
28273 : &X86::GR64RegClass;
28274 assert(Res.second->contains(Res.first) && "Register in register class");
28276 // No register found/type mismatch.
28278 Res.second = nullptr;
28280 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
28281 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
28282 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
28283 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
28284 Class == &X86::VR512RegClass) {
28285 // Handle references to XMM physical registers that got mapped into the
28286 // wrong class. This can happen with constraints like {xmm0} where the
28287 // target independent register mapper will just pick the first match it can
28288 // find, ignoring the required type.
28290 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28291 if (VT == MVT::f32 || VT == MVT::i32)
28292 Res.second = &X86::FR32RegClass;
28293 else if (VT == MVT::f64 || VT == MVT::i64)
28294 Res.second = &X86::FR64RegClass;
28295 else if (X86::VR128RegClass.hasType(VT))
28296 Res.second = &X86::VR128RegClass;
28297 else if (X86::VR256RegClass.hasType(VT))
28298 Res.second = &X86::VR256RegClass;
28299 else if (X86::VR512RegClass.hasType(VT))
28300 Res.second = &X86::VR512RegClass;
28302 // Type mismatch and not a clobber: Return an error;
28304 Res.second = nullptr;
28311 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
28312 const AddrMode &AM, Type *Ty,
28313 unsigned AS) const {
28314 // Scaling factors are not free at all.
28315 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
28316 // will take 2 allocations in the out of order engine instead of 1
28317 // for plain addressing mode, i.e. inst (reg1).
28319 // vaddps (%rsi,%drx), %ymm0, %ymm1
28320 // Requires two allocations (one for the load, one for the computation)
28322 // vaddps (%rsi), %ymm0, %ymm1
28323 // Requires just 1 allocation, i.e., freeing allocations for other operations
28324 // and having less micro operations to execute.
28326 // For some X86 architectures, this is even worse because for instance for
28327 // stores, the complex addressing mode forces the instruction to use the
28328 // "load" ports instead of the dedicated "store" port.
28329 // E.g., on Haswell:
28330 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
28331 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
28332 if (isLegalAddressingMode(DL, AM, Ty, AS))
28333 // Scale represents reg2 * scale, thus account for 1
28334 // as soon as we use a second register.
28335 return AM.Scale != 0;
28339 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
28340 // Integer division on x86 is expensive. However, when aggressively optimizing
28341 // for code size, we prefer to use a div instruction, as it is usually smaller
28342 // than the alternative sequence.
28343 // The exception to this is vector division. Since x86 doesn't have vector
28344 // integer division, leaving the division as-is is a loss even in terms of
28345 // size, because it will have to be scalarized, while the alternative code
28346 // sequence can be performed in vector form.
28347 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
28348 Attribute::MinSize);
28349 return OptSize && !VT.isVector();
28352 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
28353 TargetLowering::ArgListTy& Args) const {
28354 // The MCU psABI requires some arguments to be passed in-register.
28355 // For regular calls, the inreg arguments are marked by the front-end.
28356 // However, for compiler generated library calls, we have to patch this
28358 if (!Subtarget->isTargetMCU() || !Args.size())
28361 unsigned FreeRegs = 3;
28362 for (auto &Arg : Args) {
28363 // For library functions, we do not expect any fancy types.
28364 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
28365 unsigned SizeInRegs = (Size + 31) / 32;
28366 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
28369 Arg.isInReg = true;
28370 FreeRegs -= SizeInRegs;