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/ADT/VariadicFunction.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/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 static cl::opt<bool> ExperimentalVectorShuffleLowering(
71 "x86-experimental-vector-shuffle-lowering", cl::init(true),
72 cl::desc("Enable an experimental vector shuffle lowering code path."),
75 static cl::opt<bool> ExperimentalVectorShuffleLegality(
76 "x86-experimental-vector-shuffle-legality", cl::init(false),
77 cl::desc("Enable experimental shuffle legality based on the experimental "
78 "shuffle lowering. Should only be used with the experimental "
82 static cl::opt<int> ReciprocalEstimateRefinementSteps(
83 "x86-recip-refinement-steps", cl::init(1),
84 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
85 "result of the hardware reciprocal estimate instruction."),
88 // Forward declarations.
89 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
92 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
93 SelectionDAG &DAG, SDLoc dl,
94 unsigned vectorWidth) {
95 assert((vectorWidth == 128 || vectorWidth == 256) &&
96 "Unsupported vector width");
97 EVT VT = Vec.getValueType();
98 EVT ElVT = VT.getVectorElementType();
99 unsigned Factor = VT.getSizeInBits()/vectorWidth;
100 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
101 VT.getVectorNumElements()/Factor);
103 // Extract from UNDEF is UNDEF.
104 if (Vec.getOpcode() == ISD::UNDEF)
105 return DAG.getUNDEF(ResultVT);
107 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
108 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
110 // This is the index of the first element of the vectorWidth-bit chunk
112 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
115 // If the input is a buildvector just emit a smaller one.
116 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
117 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
118 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
121 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
122 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
125 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
126 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
127 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
128 /// instructions or a simple subregister reference. Idx is an index in the
129 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
130 /// lowering EXTRACT_VECTOR_ELT operations easier.
131 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
132 SelectionDAG &DAG, SDLoc dl) {
133 assert((Vec.getValueType().is256BitVector() ||
134 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
135 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
138 /// Generate a DAG to grab 256-bits from a 512-bit vector.
139 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
140 SelectionDAG &DAG, SDLoc dl) {
141 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
142 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
145 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
146 unsigned IdxVal, SelectionDAG &DAG,
147 SDLoc dl, unsigned vectorWidth) {
148 assert((vectorWidth == 128 || vectorWidth == 256) &&
149 "Unsupported vector width");
150 // Inserting UNDEF is Result
151 if (Vec.getOpcode() == ISD::UNDEF)
153 EVT VT = Vec.getValueType();
154 EVT ElVT = VT.getVectorElementType();
155 EVT ResultVT = Result.getValueType();
157 // Insert the relevant vectorWidth bits.
158 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
160 // This is the index of the first element of the vectorWidth-bit chunk
162 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
165 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
166 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
169 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
170 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
171 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
172 /// simple superregister reference. Idx is an index in the 128 bits
173 /// we want. It need not be aligned to a 128-bit boundary. That makes
174 /// lowering INSERT_VECTOR_ELT operations easier.
175 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
176 SelectionDAG &DAG,SDLoc dl) {
177 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
178 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
181 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
182 SelectionDAG &DAG, SDLoc dl) {
183 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
184 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
187 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
188 /// instructions. This is used because creating CONCAT_VECTOR nodes of
189 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
190 /// large BUILD_VECTORS.
191 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
192 unsigned NumElems, SelectionDAG &DAG,
194 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
195 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
198 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
199 unsigned NumElems, SelectionDAG &DAG,
201 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
202 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
205 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
206 const X86Subtarget &STI)
207 : TargetLowering(TM), Subtarget(&STI) {
208 X86ScalarSSEf64 = Subtarget->hasSSE2();
209 X86ScalarSSEf32 = Subtarget->hasSSE1();
210 TD = getDataLayout();
212 // Set up the TargetLowering object.
213 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
215 // X86 is weird. It always uses i8 for shift amounts and setcc results.
216 setBooleanContents(ZeroOrOneBooleanContent);
217 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
218 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
220 // For 64-bit, since we have so many registers, use the ILP scheduler.
221 // For 32-bit, use the register pressure specific scheduling.
222 // For Atom, always use ILP scheduling.
223 if (Subtarget->isAtom())
224 setSchedulingPreference(Sched::ILP);
225 else if (Subtarget->is64Bit())
226 setSchedulingPreference(Sched::ILP);
228 setSchedulingPreference(Sched::RegPressure);
229 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
230 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
232 // Bypass expensive divides on Atom when compiling with O2.
233 if (TM.getOptLevel() >= CodeGenOpt::Default) {
234 if (Subtarget->hasSlowDivide32())
235 addBypassSlowDiv(32, 8);
236 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
237 addBypassSlowDiv(64, 16);
240 if (Subtarget->isTargetKnownWindowsMSVC()) {
241 // Setup Windows compiler runtime calls.
242 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
243 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
244 setLibcallName(RTLIB::SREM_I64, "_allrem");
245 setLibcallName(RTLIB::UREM_I64, "_aullrem");
246 setLibcallName(RTLIB::MUL_I64, "_allmul");
247 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
248 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
249 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
250 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
251 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
253 // The _ftol2 runtime function has an unusual calling conv, which
254 // is modeled by a special pseudo-instruction.
255 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
256 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
257 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
258 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
261 if (Subtarget->isTargetDarwin()) {
262 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
263 setUseUnderscoreSetJmp(false);
264 setUseUnderscoreLongJmp(false);
265 } else if (Subtarget->isTargetWindowsGNU()) {
266 // MS runtime is weird: it exports _setjmp, but longjmp!
267 setUseUnderscoreSetJmp(true);
268 setUseUnderscoreLongJmp(false);
270 setUseUnderscoreSetJmp(true);
271 setUseUnderscoreLongJmp(true);
274 // Set up the register classes.
275 addRegisterClass(MVT::i8, &X86::GR8RegClass);
276 addRegisterClass(MVT::i16, &X86::GR16RegClass);
277 addRegisterClass(MVT::i32, &X86::GR32RegClass);
278 if (Subtarget->is64Bit())
279 addRegisterClass(MVT::i64, &X86::GR64RegClass);
281 for (MVT VT : MVT::integer_valuetypes())
282 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
284 // We don't accept any truncstore of integer registers.
285 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
286 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
287 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
288 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
289 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
290 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
292 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
294 // SETOEQ and SETUNE require checking two conditions.
295 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
296 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
297 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
298 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
299 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
300 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
302 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
304 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
305 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
306 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
308 if (Subtarget->is64Bit()) {
309 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
310 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
311 } else if (!TM.Options.UseSoftFloat) {
312 // We have an algorithm for SSE2->double, and we turn this into a
313 // 64-bit FILD followed by conditional FADD for other targets.
314 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
315 // We have an algorithm for SSE2, and we turn this into a 64-bit
316 // FILD for other targets.
317 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
320 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
322 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
323 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
325 if (!TM.Options.UseSoftFloat) {
326 // SSE has no i16 to fp conversion, only i32
327 if (X86ScalarSSEf32) {
328 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
329 // f32 and f64 cases are Legal, f80 case is not
330 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
332 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
333 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
336 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
337 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
340 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
341 // are Legal, f80 is custom lowered.
342 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
343 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
345 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
347 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
348 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
350 if (X86ScalarSSEf32) {
351 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
352 // f32 and f64 cases are Legal, f80 case is not
353 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
355 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
356 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
359 // Handle FP_TO_UINT by promoting the destination to a larger signed
361 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
362 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
363 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
365 if (Subtarget->is64Bit()) {
366 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
367 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
368 } else if (!TM.Options.UseSoftFloat) {
369 // Since AVX is a superset of SSE3, only check for SSE here.
370 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
371 // Expand FP_TO_UINT into a select.
372 // FIXME: We would like to use a Custom expander here eventually to do
373 // the optimal thing for SSE vs. the default expansion in the legalizer.
374 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
376 // With SSE3 we can use fisttpll to convert to a signed i64; without
377 // SSE, we're stuck with a fistpll.
378 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
381 if (isTargetFTOL()) {
382 // Use the _ftol2 runtime function, which has a pseudo-instruction
383 // to handle its weird calling convention.
384 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
387 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
388 if (!X86ScalarSSEf64) {
389 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
390 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
391 if (Subtarget->is64Bit()) {
392 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
393 // Without SSE, i64->f64 goes through memory.
394 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
398 // Scalar integer divide and remainder are lowered to use operations that
399 // produce two results, to match the available instructions. This exposes
400 // the two-result form to trivial CSE, which is able to combine x/y and x%y
401 // into a single instruction.
403 // Scalar integer multiply-high is also lowered to use two-result
404 // operations, to match the available instructions. However, plain multiply
405 // (low) operations are left as Legal, as there are single-result
406 // instructions for this in x86. Using the two-result multiply instructions
407 // when both high and low results are needed must be arranged by dagcombine.
408 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
410 setOperationAction(ISD::MULHS, VT, Expand);
411 setOperationAction(ISD::MULHU, VT, Expand);
412 setOperationAction(ISD::SDIV, VT, Expand);
413 setOperationAction(ISD::UDIV, VT, Expand);
414 setOperationAction(ISD::SREM, VT, Expand);
415 setOperationAction(ISD::UREM, VT, Expand);
417 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
418 setOperationAction(ISD::ADDC, VT, Custom);
419 setOperationAction(ISD::ADDE, VT, Custom);
420 setOperationAction(ISD::SUBC, VT, Custom);
421 setOperationAction(ISD::SUBE, VT, Custom);
424 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
425 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
426 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
427 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
428 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
429 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
430 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
431 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
432 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
433 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
434 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
435 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
436 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
437 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
438 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
439 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
440 if (Subtarget->is64Bit())
441 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
442 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
443 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
444 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
445 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
446 setOperationAction(ISD::FREM , MVT::f32 , Expand);
447 setOperationAction(ISD::FREM , MVT::f64 , Expand);
448 setOperationAction(ISD::FREM , MVT::f80 , Expand);
449 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
451 // Promote the i8 variants and force them on up to i32 which has a shorter
453 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
454 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
455 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
456 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
457 if (Subtarget->hasBMI()) {
458 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
459 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
460 if (Subtarget->is64Bit())
461 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
463 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
464 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
465 if (Subtarget->is64Bit())
466 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
469 if (Subtarget->hasLZCNT()) {
470 // When promoting the i8 variants, force them to i32 for a shorter
472 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
473 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
474 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
475 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
476 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
477 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
478 if (Subtarget->is64Bit())
479 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
481 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
482 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
483 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
484 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
485 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
486 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
487 if (Subtarget->is64Bit()) {
488 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
489 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
493 // Special handling for half-precision floating point conversions.
494 // If we don't have F16C support, then lower half float conversions
495 // into library calls.
496 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
497 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
498 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
501 // There's never any support for operations beyond MVT::f32.
502 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
503 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
504 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
505 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
507 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
508 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
509 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
510 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
511 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
512 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
514 if (Subtarget->hasPOPCNT()) {
515 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
517 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
518 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
519 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
520 if (Subtarget->is64Bit())
521 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
524 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
526 if (!Subtarget->hasMOVBE())
527 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
529 // These should be promoted to a larger select which is supported.
530 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
531 // X86 wants to expand cmov itself.
532 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
533 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
534 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
535 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
536 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
537 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
538 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
539 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
540 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
541 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
542 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
543 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
544 if (Subtarget->is64Bit()) {
545 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
546 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
548 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
549 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
550 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
551 // support continuation, user-level threading, and etc.. As a result, no
552 // other SjLj exception interfaces are implemented and please don't build
553 // your own exception handling based on them.
554 // LLVM/Clang supports zero-cost DWARF exception handling.
555 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
556 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
559 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
560 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
561 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
562 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
563 if (Subtarget->is64Bit())
564 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
565 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
566 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
567 if (Subtarget->is64Bit()) {
568 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
569 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
570 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
571 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
572 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
574 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
575 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
576 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
577 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
578 if (Subtarget->is64Bit()) {
579 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
580 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
581 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
584 if (Subtarget->hasSSE1())
585 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
587 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
589 // Expand certain atomics
590 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
592 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
593 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
594 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
597 if (Subtarget->hasCmpxchg16b()) {
598 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
601 // FIXME - use subtarget debug flags
602 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
603 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
604 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
607 if (Subtarget->is64Bit()) {
608 setExceptionPointerRegister(X86::RAX);
609 setExceptionSelectorRegister(X86::RDX);
611 setExceptionPointerRegister(X86::EAX);
612 setExceptionSelectorRegister(X86::EDX);
614 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
615 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
617 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
618 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
620 setOperationAction(ISD::TRAP, MVT::Other, Legal);
621 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
623 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
624 setOperationAction(ISD::VASTART , MVT::Other, Custom);
625 setOperationAction(ISD::VAEND , MVT::Other, Expand);
626 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
627 // TargetInfo::X86_64ABIBuiltinVaList
628 setOperationAction(ISD::VAARG , MVT::Other, Custom);
629 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
631 // TargetInfo::CharPtrBuiltinVaList
632 setOperationAction(ISD::VAARG , MVT::Other, Expand);
633 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
636 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
637 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
639 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
641 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
642 // f32 and f64 use SSE.
643 // Set up the FP register classes.
644 addRegisterClass(MVT::f32, &X86::FR32RegClass);
645 addRegisterClass(MVT::f64, &X86::FR64RegClass);
647 // Use ANDPD to simulate FABS.
648 setOperationAction(ISD::FABS , MVT::f64, Custom);
649 setOperationAction(ISD::FABS , MVT::f32, Custom);
651 // Use XORP to simulate FNEG.
652 setOperationAction(ISD::FNEG , MVT::f64, Custom);
653 setOperationAction(ISD::FNEG , MVT::f32, Custom);
655 // Use ANDPD and ORPD to simulate FCOPYSIGN.
656 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
657 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
659 // Lower this to FGETSIGNx86 plus an AND.
660 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
661 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
663 // We don't support sin/cos/fmod
664 setOperationAction(ISD::FSIN , MVT::f64, Expand);
665 setOperationAction(ISD::FCOS , MVT::f64, Expand);
666 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
667 setOperationAction(ISD::FSIN , MVT::f32, Expand);
668 setOperationAction(ISD::FCOS , MVT::f32, Expand);
669 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
671 // Expand FP immediates into loads from the stack, except for the special
673 addLegalFPImmediate(APFloat(+0.0)); // xorpd
674 addLegalFPImmediate(APFloat(+0.0f)); // xorps
675 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
676 // Use SSE for f32, x87 for f64.
677 // Set up the FP register classes.
678 addRegisterClass(MVT::f32, &X86::FR32RegClass);
679 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
681 // Use ANDPS to simulate FABS.
682 setOperationAction(ISD::FABS , MVT::f32, Custom);
684 // Use XORP to simulate FNEG.
685 setOperationAction(ISD::FNEG , MVT::f32, Custom);
687 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
689 // Use ANDPS and ORPS to simulate FCOPYSIGN.
690 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
691 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
693 // We don't support sin/cos/fmod
694 setOperationAction(ISD::FSIN , MVT::f32, Expand);
695 setOperationAction(ISD::FCOS , MVT::f32, Expand);
696 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
698 // Special cases we handle for FP constants.
699 addLegalFPImmediate(APFloat(+0.0f)); // xorps
700 addLegalFPImmediate(APFloat(+0.0)); // FLD0
701 addLegalFPImmediate(APFloat(+1.0)); // FLD1
702 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
703 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
705 if (!TM.Options.UnsafeFPMath) {
706 setOperationAction(ISD::FSIN , MVT::f64, Expand);
707 setOperationAction(ISD::FCOS , MVT::f64, Expand);
708 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
710 } else if (!TM.Options.UseSoftFloat) {
711 // f32 and f64 in x87.
712 // Set up the FP register classes.
713 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
714 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
716 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
717 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
718 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
719 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
721 if (!TM.Options.UnsafeFPMath) {
722 setOperationAction(ISD::FSIN , MVT::f64, Expand);
723 setOperationAction(ISD::FSIN , MVT::f32, Expand);
724 setOperationAction(ISD::FCOS , MVT::f64, Expand);
725 setOperationAction(ISD::FCOS , MVT::f32, Expand);
726 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
727 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
729 addLegalFPImmediate(APFloat(+0.0)); // FLD0
730 addLegalFPImmediate(APFloat(+1.0)); // FLD1
731 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
732 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
733 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
734 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
735 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
736 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
739 // We don't support FMA.
740 setOperationAction(ISD::FMA, MVT::f64, Expand);
741 setOperationAction(ISD::FMA, MVT::f32, Expand);
743 // Long double always uses X87.
744 if (!TM.Options.UseSoftFloat) {
745 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
746 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
747 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
749 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
750 addLegalFPImmediate(TmpFlt); // FLD0
752 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
755 APFloat TmpFlt2(+1.0);
756 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
758 addLegalFPImmediate(TmpFlt2); // FLD1
759 TmpFlt2.changeSign();
760 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
763 if (!TM.Options.UnsafeFPMath) {
764 setOperationAction(ISD::FSIN , MVT::f80, Expand);
765 setOperationAction(ISD::FCOS , MVT::f80, Expand);
766 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
769 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
770 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
771 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
772 setOperationAction(ISD::FRINT, MVT::f80, Expand);
773 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
774 setOperationAction(ISD::FMA, MVT::f80, Expand);
777 // Always use a library call for pow.
778 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
779 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
780 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
782 setOperationAction(ISD::FLOG, MVT::f80, Expand);
783 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
784 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
785 setOperationAction(ISD::FEXP, MVT::f80, Expand);
786 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
787 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
788 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
790 // First set operation action for all vector types to either promote
791 // (for widening) or expand (for scalarization). Then we will selectively
792 // turn on ones that can be effectively codegen'd.
793 for (MVT VT : MVT::vector_valuetypes()) {
794 setOperationAction(ISD::ADD , VT, Expand);
795 setOperationAction(ISD::SUB , VT, Expand);
796 setOperationAction(ISD::FADD, VT, Expand);
797 setOperationAction(ISD::FNEG, VT, Expand);
798 setOperationAction(ISD::FSUB, VT, Expand);
799 setOperationAction(ISD::MUL , VT, Expand);
800 setOperationAction(ISD::FMUL, VT, Expand);
801 setOperationAction(ISD::SDIV, VT, Expand);
802 setOperationAction(ISD::UDIV, VT, Expand);
803 setOperationAction(ISD::FDIV, VT, Expand);
804 setOperationAction(ISD::SREM, VT, Expand);
805 setOperationAction(ISD::UREM, VT, Expand);
806 setOperationAction(ISD::LOAD, VT, Expand);
807 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
808 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
809 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
810 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
811 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
812 setOperationAction(ISD::FABS, VT, Expand);
813 setOperationAction(ISD::FSIN, VT, Expand);
814 setOperationAction(ISD::FSINCOS, VT, Expand);
815 setOperationAction(ISD::FCOS, VT, Expand);
816 setOperationAction(ISD::FSINCOS, VT, Expand);
817 setOperationAction(ISD::FREM, VT, Expand);
818 setOperationAction(ISD::FMA, VT, Expand);
819 setOperationAction(ISD::FPOWI, VT, Expand);
820 setOperationAction(ISD::FSQRT, VT, Expand);
821 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
822 setOperationAction(ISD::FFLOOR, VT, Expand);
823 setOperationAction(ISD::FCEIL, VT, Expand);
824 setOperationAction(ISD::FTRUNC, VT, Expand);
825 setOperationAction(ISD::FRINT, VT, Expand);
826 setOperationAction(ISD::FNEARBYINT, VT, Expand);
827 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
828 setOperationAction(ISD::MULHS, VT, Expand);
829 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
830 setOperationAction(ISD::MULHU, VT, Expand);
831 setOperationAction(ISD::SDIVREM, VT, Expand);
832 setOperationAction(ISD::UDIVREM, VT, Expand);
833 setOperationAction(ISD::FPOW, VT, Expand);
834 setOperationAction(ISD::CTPOP, VT, Expand);
835 setOperationAction(ISD::CTTZ, VT, Expand);
836 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
837 setOperationAction(ISD::CTLZ, VT, Expand);
838 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
839 setOperationAction(ISD::SHL, VT, Expand);
840 setOperationAction(ISD::SRA, VT, Expand);
841 setOperationAction(ISD::SRL, VT, Expand);
842 setOperationAction(ISD::ROTL, VT, Expand);
843 setOperationAction(ISD::ROTR, VT, Expand);
844 setOperationAction(ISD::BSWAP, VT, Expand);
845 setOperationAction(ISD::SETCC, VT, Expand);
846 setOperationAction(ISD::FLOG, VT, Expand);
847 setOperationAction(ISD::FLOG2, VT, Expand);
848 setOperationAction(ISD::FLOG10, VT, Expand);
849 setOperationAction(ISD::FEXP, VT, Expand);
850 setOperationAction(ISD::FEXP2, VT, Expand);
851 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
852 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
853 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
854 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
855 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
856 setOperationAction(ISD::TRUNCATE, VT, Expand);
857 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
858 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
859 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
860 setOperationAction(ISD::VSELECT, VT, Expand);
861 setOperationAction(ISD::SELECT_CC, VT, Expand);
862 for (MVT InnerVT : MVT::vector_valuetypes()) {
863 setTruncStoreAction(InnerVT, VT, Expand);
865 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
866 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
868 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
869 // types, we have to deal with them whether we ask for Expansion or not.
870 // Setting Expand causes its own optimisation problems though, so leave
872 if (VT.getVectorElementType() == MVT::i1)
873 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
877 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
878 // with -msoft-float, disable use of MMX as well.
879 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
880 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
881 // No operations on x86mmx supported, everything uses intrinsics.
884 // MMX-sized vectors (other than x86mmx) are expected to be expanded
885 // into smaller operations.
886 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
887 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
888 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
889 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
890 setOperationAction(ISD::AND, MVT::v8i8, Expand);
891 setOperationAction(ISD::AND, MVT::v4i16, Expand);
892 setOperationAction(ISD::AND, MVT::v2i32, Expand);
893 setOperationAction(ISD::AND, MVT::v1i64, Expand);
894 setOperationAction(ISD::OR, MVT::v8i8, Expand);
895 setOperationAction(ISD::OR, MVT::v4i16, Expand);
896 setOperationAction(ISD::OR, MVT::v2i32, Expand);
897 setOperationAction(ISD::OR, MVT::v1i64, Expand);
898 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
899 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
900 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
901 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
902 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
903 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
904 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
905 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
907 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
908 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
909 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
910 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
911 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
912 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
913 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
914 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
916 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
917 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
919 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
920 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
921 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
922 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
923 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
924 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
925 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
926 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
927 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
928 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
929 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
930 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
931 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
932 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
935 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
936 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
938 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
939 // registers cannot be used even for integer operations.
940 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
941 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
942 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
943 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
945 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
946 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
947 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
948 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
949 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
950 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
951 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
952 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
953 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
954 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
955 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
956 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
957 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
958 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
959 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
960 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
961 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
962 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
963 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
964 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
965 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
966 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
968 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
969 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
970 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
971 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
973 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
974 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
975 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
976 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
977 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
979 // Only provide customized ctpop vector bit twiddling for vector types we
980 // know to perform better than using the popcnt instructions on each vector
981 // element. If popcnt isn't supported, always provide the custom version.
982 if (!Subtarget->hasPOPCNT()) {
983 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
984 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
987 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
988 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
989 MVT VT = (MVT::SimpleValueType)i;
990 // Do not attempt to custom lower non-power-of-2 vectors
991 if (!isPowerOf2_32(VT.getVectorNumElements()))
993 // Do not attempt to custom lower non-128-bit vectors
994 if (!VT.is128BitVector())
996 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
997 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
998 setOperationAction(ISD::VSELECT, VT, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1002 // We support custom legalizing of sext and anyext loads for specific
1003 // memory vector types which we can load as a scalar (or sequence of
1004 // scalars) and extend in-register to a legal 128-bit vector type. For sext
1005 // loads these must work with a single scalar load.
1006 for (MVT VT : MVT::integer_vector_valuetypes()) {
1007 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
1008 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
1009 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
1010 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
1011 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
1012 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
1013 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
1014 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
1015 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
1018 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1019 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1020 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1021 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1022 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1023 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1024 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1025 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1027 if (Subtarget->is64Bit()) {
1028 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1029 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1032 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1033 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1034 MVT VT = (MVT::SimpleValueType)i;
1036 // Do not attempt to promote non-128-bit vectors
1037 if (!VT.is128BitVector())
1040 setOperationAction(ISD::AND, VT, Promote);
1041 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1042 setOperationAction(ISD::OR, VT, Promote);
1043 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1044 setOperationAction(ISD::XOR, VT, Promote);
1045 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1046 setOperationAction(ISD::LOAD, VT, Promote);
1047 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1048 setOperationAction(ISD::SELECT, VT, Promote);
1049 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1052 // Custom lower v2i64 and v2f64 selects.
1053 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1054 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1055 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1056 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1058 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1059 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1061 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1062 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1063 // As there is no 64-bit GPR available, we need build a special custom
1064 // sequence to convert from v2i32 to v2f32.
1065 if (!Subtarget->is64Bit())
1066 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1068 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1069 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1071 for (MVT VT : MVT::fp_vector_valuetypes())
1072 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
1074 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1075 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1076 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1079 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1080 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1081 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1082 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1083 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1084 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1085 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1086 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1087 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1088 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1089 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1091 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1092 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1093 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1094 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1095 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1096 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1097 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1098 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1099 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1100 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1102 // FIXME: Do we need to handle scalar-to-vector here?
1103 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1105 // We directly match byte blends in the backend as they match the VSELECT
1107 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1109 // SSE41 brings specific instructions for doing vector sign extend even in
1110 // cases where we don't have SRA.
1111 for (MVT VT : MVT::integer_vector_valuetypes()) {
1112 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
1113 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
1114 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
1117 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
1118 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1119 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1120 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1121 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1122 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1123 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1125 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1126 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1127 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1128 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1129 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1130 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1132 // i8 and i16 vectors are custom because the source register and source
1133 // source memory operand types are not the same width. f32 vectors are
1134 // custom since the immediate controlling the insert encodes additional
1136 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1137 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1138 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1139 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1141 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1142 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1143 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1144 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1146 // FIXME: these should be Legal, but that's only for the case where
1147 // the index is constant. For now custom expand to deal with that.
1148 if (Subtarget->is64Bit()) {
1149 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1150 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1154 if (Subtarget->hasSSE2()) {
1155 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1156 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1158 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1159 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1161 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1162 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1164 // In the customized shift lowering, the legal cases in AVX2 will be
1166 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1167 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1169 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1170 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1172 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1175 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1176 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1177 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1178 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1179 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1180 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1181 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1183 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1184 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1185 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1187 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1188 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1189 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1190 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1191 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1192 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1193 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1194 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1195 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1196 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1197 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1198 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1200 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1201 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1202 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1203 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1204 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1205 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1206 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1207 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1208 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1209 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1210 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1211 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1213 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1214 // even though v8i16 is a legal type.
1215 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1216 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1217 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1219 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1220 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1221 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1223 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1224 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1226 for (MVT VT : MVT::fp_vector_valuetypes())
1227 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1229 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1230 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1232 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1235 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1236 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1238 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1239 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1240 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1241 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1243 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1244 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1245 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1247 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1248 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1249 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1250 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1251 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1252 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1253 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1254 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1255 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1256 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1257 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1258 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1260 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1261 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1262 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1263 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1264 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1265 setOperationAction(ISD::FMA, MVT::f32, Legal);
1266 setOperationAction(ISD::FMA, MVT::f64, Legal);
1269 if (Subtarget->hasInt256()) {
1270 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1271 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1272 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1273 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1275 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1276 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1277 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1278 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1280 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1281 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1282 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1283 // Don't lower v32i8 because there is no 128-bit byte mul
1285 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1286 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1287 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1288 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1290 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1291 // when we have a 256bit-wide blend with immediate.
1292 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1294 // Only provide customized ctpop vector bit twiddling for vector types we
1295 // know to perform better than using the popcnt instructions on each
1296 // vector element. If popcnt isn't supported, always provide the custom
1298 if (!Subtarget->hasPOPCNT())
1299 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1301 // Custom CTPOP always performs better on natively supported v8i32
1302 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1304 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1305 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1306 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1307 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1308 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1309 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1310 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1312 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1313 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1314 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1315 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1316 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1317 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1319 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1320 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1321 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1322 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1324 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1325 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1326 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1327 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1329 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1330 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1331 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1332 // Don't lower v32i8 because there is no 128-bit byte mul
1335 // In the customized shift lowering, the legal cases in AVX2 will be
1337 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1338 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1340 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1341 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1343 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1345 // Custom lower several nodes for 256-bit types.
1346 for (MVT VT : MVT::vector_valuetypes()) {
1347 if (VT.getScalarSizeInBits() >= 32) {
1348 setOperationAction(ISD::MLOAD, VT, Legal);
1349 setOperationAction(ISD::MSTORE, VT, Legal);
1351 // Extract subvector is special because the value type
1352 // (result) is 128-bit but the source is 256-bit wide.
1353 if (VT.is128BitVector()) {
1354 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1356 // Do not attempt to custom lower other non-256-bit vectors
1357 if (!VT.is256BitVector())
1360 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1361 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1362 setOperationAction(ISD::VSELECT, VT, Custom);
1363 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1364 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1365 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1366 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1367 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1370 if (Subtarget->hasInt256())
1371 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1374 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1375 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1376 MVT VT = (MVT::SimpleValueType)i;
1378 // Do not attempt to promote non-256-bit vectors
1379 if (!VT.is256BitVector())
1382 setOperationAction(ISD::AND, VT, Promote);
1383 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1384 setOperationAction(ISD::OR, VT, Promote);
1385 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1386 setOperationAction(ISD::XOR, VT, Promote);
1387 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1388 setOperationAction(ISD::LOAD, VT, Promote);
1389 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1390 setOperationAction(ISD::SELECT, VT, Promote);
1391 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1395 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1396 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1397 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1398 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1399 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1401 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1402 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1403 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1405 for (MVT VT : MVT::fp_vector_valuetypes())
1406 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1408 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1409 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1410 setOperationAction(ISD::XOR, MVT::i1, Legal);
1411 setOperationAction(ISD::OR, MVT::i1, Legal);
1412 setOperationAction(ISD::AND, MVT::i1, Legal);
1413 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1414 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1415 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1416 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1417 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1419 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1420 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1421 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1422 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1423 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1424 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1426 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1427 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1428 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1429 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1430 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1431 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1432 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1433 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1436 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1438 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1439 if (Subtarget->is64Bit()) {
1440 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1441 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1442 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1443 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1445 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1446 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1447 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1448 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1449 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1450 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1451 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1452 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1453 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1454 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1455 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1456 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1457 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1458 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1460 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1461 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1462 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1463 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1464 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1465 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1466 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1467 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1468 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1469 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1470 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1471 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1472 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1474 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1475 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1476 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1477 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1478 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1479 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1480 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1481 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1482 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1483 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1485 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1486 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1487 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1488 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1489 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1490 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1492 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1493 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1495 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1497 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1498 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1499 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1500 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1501 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1502 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1503 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1504 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1505 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1507 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1508 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1510 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1511 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1513 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1515 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1516 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1518 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1519 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1521 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1522 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1524 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1525 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1526 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1527 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1528 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1529 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1531 if (Subtarget->hasCDI()) {
1532 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1533 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1536 // Custom lower several nodes.
1537 for (MVT VT : MVT::vector_valuetypes()) {
1538 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1539 // Extract subvector is special because the value type
1540 // (result) is 256/128-bit but the source is 512-bit wide.
1541 if (VT.is128BitVector() || VT.is256BitVector()) {
1542 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1544 if (VT.getVectorElementType() == MVT::i1)
1545 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1547 // Do not attempt to custom lower other non-512-bit vectors
1548 if (!VT.is512BitVector())
1551 if ( EltSize >= 32) {
1552 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1553 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1554 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1555 setOperationAction(ISD::VSELECT, VT, Legal);
1556 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1557 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1558 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1559 setOperationAction(ISD::MLOAD, VT, Legal);
1560 setOperationAction(ISD::MSTORE, VT, Legal);
1563 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1564 MVT VT = (MVT::SimpleValueType)i;
1566 // Do not attempt to promote non-512-bit vectors.
1567 if (!VT.is512BitVector())
1570 setOperationAction(ISD::SELECT, VT, Promote);
1571 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1575 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1576 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1577 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1579 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1580 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1582 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1583 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1584 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1585 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1586 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1587 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1588 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1589 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1590 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1592 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1593 const MVT VT = (MVT::SimpleValueType)i;
1595 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1597 // Do not attempt to promote non-512-bit vectors.
1598 if (!VT.is512BitVector())
1602 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1603 setOperationAction(ISD::VSELECT, VT, Legal);
1608 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1609 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1610 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1612 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1613 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1614 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal);
1616 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1617 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1618 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1619 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1620 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1621 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1624 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1625 // of this type with custom code.
1626 for (MVT VT : MVT::vector_valuetypes())
1627 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
1629 // We want to custom lower some of our intrinsics.
1630 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1631 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1632 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1633 if (!Subtarget->is64Bit())
1634 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1636 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1637 // handle type legalization for these operations here.
1639 // FIXME: We really should do custom legalization for addition and
1640 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1641 // than generic legalization for 64-bit multiplication-with-overflow, though.
1642 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1643 // Add/Sub/Mul with overflow operations are custom lowered.
1645 setOperationAction(ISD::SADDO, VT, Custom);
1646 setOperationAction(ISD::UADDO, VT, Custom);
1647 setOperationAction(ISD::SSUBO, VT, Custom);
1648 setOperationAction(ISD::USUBO, VT, Custom);
1649 setOperationAction(ISD::SMULO, VT, Custom);
1650 setOperationAction(ISD::UMULO, VT, Custom);
1654 if (!Subtarget->is64Bit()) {
1655 // These libcalls are not available in 32-bit.
1656 setLibcallName(RTLIB::SHL_I128, nullptr);
1657 setLibcallName(RTLIB::SRL_I128, nullptr);
1658 setLibcallName(RTLIB::SRA_I128, nullptr);
1661 // Combine sin / cos into one node or libcall if possible.
1662 if (Subtarget->hasSinCos()) {
1663 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1664 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1665 if (Subtarget->isTargetDarwin()) {
1666 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1667 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1668 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1669 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1673 if (Subtarget->isTargetWin64()) {
1674 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1675 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1676 setOperationAction(ISD::SREM, MVT::i128, Custom);
1677 setOperationAction(ISD::UREM, MVT::i128, Custom);
1678 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1679 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1682 // We have target-specific dag combine patterns for the following nodes:
1683 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1684 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1685 setTargetDAGCombine(ISD::BITCAST);
1686 setTargetDAGCombine(ISD::VSELECT);
1687 setTargetDAGCombine(ISD::SELECT);
1688 setTargetDAGCombine(ISD::SHL);
1689 setTargetDAGCombine(ISD::SRA);
1690 setTargetDAGCombine(ISD::SRL);
1691 setTargetDAGCombine(ISD::OR);
1692 setTargetDAGCombine(ISD::AND);
1693 setTargetDAGCombine(ISD::ADD);
1694 setTargetDAGCombine(ISD::FADD);
1695 setTargetDAGCombine(ISD::FSUB);
1696 setTargetDAGCombine(ISD::FMA);
1697 setTargetDAGCombine(ISD::SUB);
1698 setTargetDAGCombine(ISD::LOAD);
1699 setTargetDAGCombine(ISD::MLOAD);
1700 setTargetDAGCombine(ISD::STORE);
1701 setTargetDAGCombine(ISD::MSTORE);
1702 setTargetDAGCombine(ISD::ZERO_EXTEND);
1703 setTargetDAGCombine(ISD::ANY_EXTEND);
1704 setTargetDAGCombine(ISD::SIGN_EXTEND);
1705 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1706 setTargetDAGCombine(ISD::TRUNCATE);
1707 setTargetDAGCombine(ISD::SINT_TO_FP);
1708 setTargetDAGCombine(ISD::SETCC);
1709 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1710 setTargetDAGCombine(ISD::BUILD_VECTOR);
1711 setTargetDAGCombine(ISD::MUL);
1712 setTargetDAGCombine(ISD::XOR);
1714 computeRegisterProperties();
1716 // On Darwin, -Os means optimize for size without hurting performance,
1717 // do not reduce the limit.
1718 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1719 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1720 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1721 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1722 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1723 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1724 setPrefLoopAlignment(4); // 2^4 bytes.
1726 // Predictable cmov don't hurt on atom because it's in-order.
1727 PredictableSelectIsExpensive = !Subtarget->isAtom();
1728 EnableExtLdPromotion = true;
1729 setPrefFunctionAlignment(4); // 2^4 bytes.
1731 verifyIntrinsicTables();
1734 // This has so far only been implemented for 64-bit MachO.
1735 bool X86TargetLowering::useLoadStackGuardNode() const {
1736 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1739 TargetLoweringBase::LegalizeTypeAction
1740 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1741 if (ExperimentalVectorWideningLegalization &&
1742 VT.getVectorNumElements() != 1 &&
1743 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1744 return TypeWidenVector;
1746 return TargetLoweringBase::getPreferredVectorAction(VT);
1749 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1751 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1753 const unsigned NumElts = VT.getVectorNumElements();
1754 const EVT EltVT = VT.getVectorElementType();
1755 if (VT.is512BitVector()) {
1756 if (Subtarget->hasAVX512())
1757 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1758 EltVT == MVT::f32 || EltVT == MVT::f64)
1760 case 8: return MVT::v8i1;
1761 case 16: return MVT::v16i1;
1763 if (Subtarget->hasBWI())
1764 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1766 case 32: return MVT::v32i1;
1767 case 64: return MVT::v64i1;
1771 if (VT.is256BitVector() || VT.is128BitVector()) {
1772 if (Subtarget->hasVLX())
1773 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1774 EltVT == MVT::f32 || EltVT == MVT::f64)
1776 case 2: return MVT::v2i1;
1777 case 4: return MVT::v4i1;
1778 case 8: return MVT::v8i1;
1780 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1781 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1783 case 8: return MVT::v8i1;
1784 case 16: return MVT::v16i1;
1785 case 32: return MVT::v32i1;
1789 return VT.changeVectorElementTypeToInteger();
1792 /// Helper for getByValTypeAlignment to determine
1793 /// the desired ByVal argument alignment.
1794 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1797 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1798 if (VTy->getBitWidth() == 128)
1800 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1801 unsigned EltAlign = 0;
1802 getMaxByValAlign(ATy->getElementType(), EltAlign);
1803 if (EltAlign > MaxAlign)
1804 MaxAlign = EltAlign;
1805 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1806 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1807 unsigned EltAlign = 0;
1808 getMaxByValAlign(STy->getElementType(i), EltAlign);
1809 if (EltAlign > MaxAlign)
1810 MaxAlign = EltAlign;
1817 /// Return the desired alignment for ByVal aggregate
1818 /// function arguments in the caller parameter area. For X86, aggregates
1819 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1820 /// are at 4-byte boundaries.
1821 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1822 if (Subtarget->is64Bit()) {
1823 // Max of 8 and alignment of type.
1824 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1831 if (Subtarget->hasSSE1())
1832 getMaxByValAlign(Ty, Align);
1836 /// Returns the target specific optimal type for load
1837 /// and store operations as a result of memset, memcpy, and memmove
1838 /// lowering. If DstAlign is zero that means it's safe to destination
1839 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1840 /// means there isn't a need to check it against alignment requirement,
1841 /// probably because the source does not need to be loaded. If 'IsMemset' is
1842 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1843 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1844 /// source is constant so it does not need to be loaded.
1845 /// It returns EVT::Other if the type should be determined using generic
1846 /// target-independent logic.
1848 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1849 unsigned DstAlign, unsigned SrcAlign,
1850 bool IsMemset, bool ZeroMemset,
1852 MachineFunction &MF) const {
1853 const Function *F = MF.getFunction();
1854 if ((!IsMemset || ZeroMemset) &&
1855 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1857 (Subtarget->isUnalignedMemAccessFast() ||
1858 ((DstAlign == 0 || DstAlign >= 16) &&
1859 (SrcAlign == 0 || SrcAlign >= 16)))) {
1861 if (Subtarget->hasInt256())
1863 if (Subtarget->hasFp256())
1866 if (Subtarget->hasSSE2())
1868 if (Subtarget->hasSSE1())
1870 } else if (!MemcpyStrSrc && Size >= 8 &&
1871 !Subtarget->is64Bit() &&
1872 Subtarget->hasSSE2()) {
1873 // Do not use f64 to lower memcpy if source is string constant. It's
1874 // better to use i32 to avoid the loads.
1878 if (Subtarget->is64Bit() && Size >= 8)
1883 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1885 return X86ScalarSSEf32;
1886 else if (VT == MVT::f64)
1887 return X86ScalarSSEf64;
1892 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1897 *Fast = Subtarget->isUnalignedMemAccessFast();
1901 /// Return the entry encoding for a jump table in the
1902 /// current function. The returned value is a member of the
1903 /// MachineJumpTableInfo::JTEntryKind enum.
1904 unsigned X86TargetLowering::getJumpTableEncoding() const {
1905 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1907 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1908 Subtarget->isPICStyleGOT())
1909 return MachineJumpTableInfo::EK_Custom32;
1911 // Otherwise, use the normal jump table encoding heuristics.
1912 return TargetLowering::getJumpTableEncoding();
1916 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1917 const MachineBasicBlock *MBB,
1918 unsigned uid,MCContext &Ctx) const{
1919 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1920 Subtarget->isPICStyleGOT());
1921 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1923 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1924 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1927 /// Returns relocation base for the given PIC jumptable.
1928 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1929 SelectionDAG &DAG) const {
1930 if (!Subtarget->is64Bit())
1931 // This doesn't have SDLoc associated with it, but is not really the
1932 // same as a Register.
1933 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1937 /// This returns the relocation base for the given PIC jumptable,
1938 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1939 const MCExpr *X86TargetLowering::
1940 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1941 MCContext &Ctx) const {
1942 // X86-64 uses RIP relative addressing based on the jump table label.
1943 if (Subtarget->isPICStyleRIPRel())
1944 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1946 // Otherwise, the reference is relative to the PIC base.
1947 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1950 // FIXME: Why this routine is here? Move to RegInfo!
1951 std::pair<const TargetRegisterClass*, uint8_t>
1952 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1953 const TargetRegisterClass *RRC = nullptr;
1955 switch (VT.SimpleTy) {
1957 return TargetLowering::findRepresentativeClass(VT);
1958 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1959 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1962 RRC = &X86::VR64RegClass;
1964 case MVT::f32: case MVT::f64:
1965 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1966 case MVT::v4f32: case MVT::v2f64:
1967 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1969 RRC = &X86::VR128RegClass;
1972 return std::make_pair(RRC, Cost);
1975 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1976 unsigned &Offset) const {
1977 if (!Subtarget->isTargetLinux())
1980 if (Subtarget->is64Bit()) {
1981 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1983 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1995 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1996 unsigned DestAS) const {
1997 assert(SrcAS != DestAS && "Expected different address spaces!");
1999 return SrcAS < 256 && DestAS < 256;
2002 //===----------------------------------------------------------------------===//
2003 // Return Value Calling Convention Implementation
2004 //===----------------------------------------------------------------------===//
2006 #include "X86GenCallingConv.inc"
2009 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2010 MachineFunction &MF, bool isVarArg,
2011 const SmallVectorImpl<ISD::OutputArg> &Outs,
2012 LLVMContext &Context) const {
2013 SmallVector<CCValAssign, 16> RVLocs;
2014 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2015 return CCInfo.CheckReturn(Outs, RetCC_X86);
2018 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2019 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2024 X86TargetLowering::LowerReturn(SDValue Chain,
2025 CallingConv::ID CallConv, bool isVarArg,
2026 const SmallVectorImpl<ISD::OutputArg> &Outs,
2027 const SmallVectorImpl<SDValue> &OutVals,
2028 SDLoc dl, SelectionDAG &DAG) const {
2029 MachineFunction &MF = DAG.getMachineFunction();
2030 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2032 SmallVector<CCValAssign, 16> RVLocs;
2033 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2034 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2037 SmallVector<SDValue, 6> RetOps;
2038 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2039 // Operand #1 = Bytes To Pop
2040 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
2043 // Copy the result values into the output registers.
2044 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2045 CCValAssign &VA = RVLocs[i];
2046 assert(VA.isRegLoc() && "Can only return in registers!");
2047 SDValue ValToCopy = OutVals[i];
2048 EVT ValVT = ValToCopy.getValueType();
2050 // Promote values to the appropriate types.
2051 if (VA.getLocInfo() == CCValAssign::SExt)
2052 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2053 else if (VA.getLocInfo() == CCValAssign::ZExt)
2054 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2055 else if (VA.getLocInfo() == CCValAssign::AExt)
2056 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2057 else if (VA.getLocInfo() == CCValAssign::BCvt)
2058 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2060 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2061 "Unexpected FP-extend for return value.");
2063 // If this is x86-64, and we disabled SSE, we can't return FP values,
2064 // or SSE or MMX vectors.
2065 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2066 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2067 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2068 report_fatal_error("SSE register return with SSE disabled");
2070 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2071 // llvm-gcc has never done it right and no one has noticed, so this
2072 // should be OK for now.
2073 if (ValVT == MVT::f64 &&
2074 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2075 report_fatal_error("SSE2 register return with SSE2 disabled");
2077 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2078 // the RET instruction and handled by the FP Stackifier.
2079 if (VA.getLocReg() == X86::FP0 ||
2080 VA.getLocReg() == X86::FP1) {
2081 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2082 // change the value to the FP stack register class.
2083 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2084 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2085 RetOps.push_back(ValToCopy);
2086 // Don't emit a copytoreg.
2090 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2091 // which is returned in RAX / RDX.
2092 if (Subtarget->is64Bit()) {
2093 if (ValVT == MVT::x86mmx) {
2094 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2095 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2096 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2098 // If we don't have SSE2 available, convert to v4f32 so the generated
2099 // register is legal.
2100 if (!Subtarget->hasSSE2())
2101 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2106 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2107 Flag = Chain.getValue(1);
2108 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2111 // The x86-64 ABIs require that for returning structs by value we copy
2112 // the sret argument into %rax/%eax (depending on ABI) for the return.
2113 // Win32 requires us to put the sret argument to %eax as well.
2114 // We saved the argument into a virtual register in the entry block,
2115 // so now we copy the value out and into %rax/%eax.
2117 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2118 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2119 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2120 // either case FuncInfo->setSRetReturnReg() will have been called.
2121 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2122 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
2123 "No need for an sret register");
2124 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2127 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2128 X86::RAX : X86::EAX;
2129 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2130 Flag = Chain.getValue(1);
2132 // RAX/EAX now acts like a return value.
2133 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2136 RetOps[0] = Chain; // Update chain.
2138 // Add the flag if we have it.
2140 RetOps.push_back(Flag);
2142 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2145 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2146 if (N->getNumValues() != 1)
2148 if (!N->hasNUsesOfValue(1, 0))
2151 SDValue TCChain = Chain;
2152 SDNode *Copy = *N->use_begin();
2153 if (Copy->getOpcode() == ISD::CopyToReg) {
2154 // If the copy has a glue operand, we conservatively assume it isn't safe to
2155 // perform a tail call.
2156 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2158 TCChain = Copy->getOperand(0);
2159 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2162 bool HasRet = false;
2163 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2165 if (UI->getOpcode() != X86ISD::RET_FLAG)
2167 // If we are returning more than one value, we can definitely
2168 // not make a tail call see PR19530
2169 if (UI->getNumOperands() > 4)
2171 if (UI->getNumOperands() == 4 &&
2172 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2185 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2186 ISD::NodeType ExtendKind) const {
2188 // TODO: Is this also valid on 32-bit?
2189 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2190 ReturnMVT = MVT::i8;
2192 ReturnMVT = MVT::i32;
2194 EVT MinVT = getRegisterType(Context, ReturnMVT);
2195 return VT.bitsLT(MinVT) ? MinVT : VT;
2198 /// Lower the result values of a call into the
2199 /// appropriate copies out of appropriate physical registers.
2202 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2203 CallingConv::ID CallConv, bool isVarArg,
2204 const SmallVectorImpl<ISD::InputArg> &Ins,
2205 SDLoc dl, SelectionDAG &DAG,
2206 SmallVectorImpl<SDValue> &InVals) const {
2208 // Assign locations to each value returned by this call.
2209 SmallVector<CCValAssign, 16> RVLocs;
2210 bool Is64Bit = Subtarget->is64Bit();
2211 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2213 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2215 // Copy all of the result registers out of their specified physreg.
2216 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2217 CCValAssign &VA = RVLocs[i];
2218 EVT CopyVT = VA.getValVT();
2220 // If this is x86-64, and we disabled SSE, we can't return FP values
2221 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2222 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2223 report_fatal_error("SSE register return with SSE disabled");
2226 // If we prefer to use the value in xmm registers, copy it out as f80 and
2227 // use a truncate to move it from fp stack reg to xmm reg.
2228 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2229 isScalarFPTypeInSSEReg(VA.getValVT()))
2232 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2233 CopyVT, InFlag).getValue(1);
2234 SDValue Val = Chain.getValue(0);
2236 if (CopyVT != VA.getValVT())
2237 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2238 // This truncation won't change the value.
2239 DAG.getIntPtrConstant(1));
2241 InFlag = Chain.getValue(2);
2242 InVals.push_back(Val);
2248 //===----------------------------------------------------------------------===//
2249 // C & StdCall & Fast Calling Convention implementation
2250 //===----------------------------------------------------------------------===//
2251 // StdCall calling convention seems to be standard for many Windows' API
2252 // routines and around. It differs from C calling convention just a little:
2253 // callee should clean up the stack, not caller. Symbols should be also
2254 // decorated in some fancy way :) It doesn't support any vector arguments.
2255 // For info on fast calling convention see Fast Calling Convention (tail call)
2256 // implementation LowerX86_32FastCCCallTo.
2258 /// CallIsStructReturn - Determines whether a call uses struct return
2260 enum StructReturnType {
2265 static StructReturnType
2266 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2268 return NotStructReturn;
2270 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2271 if (!Flags.isSRet())
2272 return NotStructReturn;
2273 if (Flags.isInReg())
2274 return RegStructReturn;
2275 return StackStructReturn;
2278 /// Determines whether a function uses struct return semantics.
2279 static StructReturnType
2280 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2282 return NotStructReturn;
2284 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2285 if (!Flags.isSRet())
2286 return NotStructReturn;
2287 if (Flags.isInReg())
2288 return RegStructReturn;
2289 return StackStructReturn;
2292 /// Make a copy of an aggregate at address specified by "Src" to address
2293 /// "Dst" with size and alignment information specified by the specific
2294 /// parameter attribute. The copy will be passed as a byval function parameter.
2296 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2297 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2299 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2301 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2302 /*isVolatile*/false, /*AlwaysInline=*/true,
2303 MachinePointerInfo(), MachinePointerInfo());
2306 /// Return true if the calling convention is one that
2307 /// supports tail call optimization.
2308 static bool IsTailCallConvention(CallingConv::ID CC) {
2309 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2310 CC == CallingConv::HiPE);
2313 /// \brief Return true if the calling convention is a C calling convention.
2314 static bool IsCCallConvention(CallingConv::ID CC) {
2315 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2316 CC == CallingConv::X86_64_SysV);
2319 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2320 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2324 CallingConv::ID CalleeCC = CS.getCallingConv();
2325 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2331 /// Return true if the function is being made into
2332 /// a tailcall target by changing its ABI.
2333 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2334 bool GuaranteedTailCallOpt) {
2335 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2339 X86TargetLowering::LowerMemArgument(SDValue Chain,
2340 CallingConv::ID CallConv,
2341 const SmallVectorImpl<ISD::InputArg> &Ins,
2342 SDLoc dl, SelectionDAG &DAG,
2343 const CCValAssign &VA,
2344 MachineFrameInfo *MFI,
2346 // Create the nodes corresponding to a load from this parameter slot.
2347 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2348 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2349 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2350 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2353 // If value is passed by pointer we have address passed instead of the value
2355 if (VA.getLocInfo() == CCValAssign::Indirect)
2356 ValVT = VA.getLocVT();
2358 ValVT = VA.getValVT();
2360 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2361 // changed with more analysis.
2362 // In case of tail call optimization mark all arguments mutable. Since they
2363 // could be overwritten by lowering of arguments in case of a tail call.
2364 if (Flags.isByVal()) {
2365 unsigned Bytes = Flags.getByValSize();
2366 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2367 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2368 return DAG.getFrameIndex(FI, getPointerTy());
2370 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2371 VA.getLocMemOffset(), isImmutable);
2372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2373 return DAG.getLoad(ValVT, dl, Chain, FIN,
2374 MachinePointerInfo::getFixedStack(FI),
2375 false, false, false, 0);
2379 // FIXME: Get this from tablegen.
2380 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2381 const X86Subtarget *Subtarget) {
2382 assert(Subtarget->is64Bit());
2384 if (Subtarget->isCallingConvWin64(CallConv)) {
2385 static const MCPhysReg GPR64ArgRegsWin64[] = {
2386 X86::RCX, X86::RDX, X86::R8, X86::R9
2388 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2391 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2392 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2394 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2397 // FIXME: Get this from tablegen.
2398 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2399 CallingConv::ID CallConv,
2400 const X86Subtarget *Subtarget) {
2401 assert(Subtarget->is64Bit());
2402 if (Subtarget->isCallingConvWin64(CallConv)) {
2403 // The XMM registers which might contain var arg parameters are shadowed
2404 // in their paired GPR. So we only need to save the GPR to their home
2406 // TODO: __vectorcall will change this.
2410 const Function *Fn = MF.getFunction();
2411 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2412 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2413 "SSE register cannot be used when SSE is disabled!");
2414 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2415 !Subtarget->hasSSE1())
2416 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2420 static const MCPhysReg XMMArgRegs64Bit[] = {
2421 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2422 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2424 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2428 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2429 CallingConv::ID CallConv,
2431 const SmallVectorImpl<ISD::InputArg> &Ins,
2434 SmallVectorImpl<SDValue> &InVals)
2436 MachineFunction &MF = DAG.getMachineFunction();
2437 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2439 const Function* Fn = MF.getFunction();
2440 if (Fn->hasExternalLinkage() &&
2441 Subtarget->isTargetCygMing() &&
2442 Fn->getName() == "main")
2443 FuncInfo->setForceFramePointer(true);
2445 MachineFrameInfo *MFI = MF.getFrameInfo();
2446 bool Is64Bit = Subtarget->is64Bit();
2447 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2449 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2450 "Var args not supported with calling convention fastcc, ghc or hipe");
2452 // Assign locations to all of the incoming arguments.
2453 SmallVector<CCValAssign, 16> ArgLocs;
2454 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2456 // Allocate shadow area for Win64
2458 CCInfo.AllocateStack(32, 8);
2460 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2462 unsigned LastVal = ~0U;
2464 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2465 CCValAssign &VA = ArgLocs[i];
2466 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2468 assert(VA.getValNo() != LastVal &&
2469 "Don't support value assigned to multiple locs yet");
2471 LastVal = VA.getValNo();
2473 if (VA.isRegLoc()) {
2474 EVT RegVT = VA.getLocVT();
2475 const TargetRegisterClass *RC;
2476 if (RegVT == MVT::i32)
2477 RC = &X86::GR32RegClass;
2478 else if (Is64Bit && RegVT == MVT::i64)
2479 RC = &X86::GR64RegClass;
2480 else if (RegVT == MVT::f32)
2481 RC = &X86::FR32RegClass;
2482 else if (RegVT == MVT::f64)
2483 RC = &X86::FR64RegClass;
2484 else if (RegVT.is512BitVector())
2485 RC = &X86::VR512RegClass;
2486 else if (RegVT.is256BitVector())
2487 RC = &X86::VR256RegClass;
2488 else if (RegVT.is128BitVector())
2489 RC = &X86::VR128RegClass;
2490 else if (RegVT == MVT::x86mmx)
2491 RC = &X86::VR64RegClass;
2492 else if (RegVT == MVT::i1)
2493 RC = &X86::VK1RegClass;
2494 else if (RegVT == MVT::v8i1)
2495 RC = &X86::VK8RegClass;
2496 else if (RegVT == MVT::v16i1)
2497 RC = &X86::VK16RegClass;
2498 else if (RegVT == MVT::v32i1)
2499 RC = &X86::VK32RegClass;
2500 else if (RegVT == MVT::v64i1)
2501 RC = &X86::VK64RegClass;
2503 llvm_unreachable("Unknown argument type!");
2505 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2506 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2508 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2509 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2511 if (VA.getLocInfo() == CCValAssign::SExt)
2512 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2513 DAG.getValueType(VA.getValVT()));
2514 else if (VA.getLocInfo() == CCValAssign::ZExt)
2515 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2516 DAG.getValueType(VA.getValVT()));
2517 else if (VA.getLocInfo() == CCValAssign::BCvt)
2518 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2520 if (VA.isExtInLoc()) {
2521 // Handle MMX values passed in XMM regs.
2522 if (RegVT.isVector())
2523 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2525 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2528 assert(VA.isMemLoc());
2529 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2532 // If value is passed via pointer - do a load.
2533 if (VA.getLocInfo() == CCValAssign::Indirect)
2534 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2535 MachinePointerInfo(), false, false, false, 0);
2537 InVals.push_back(ArgValue);
2540 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2541 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2542 // The x86-64 ABIs require that for returning structs by value we copy
2543 // the sret argument into %rax/%eax (depending on ABI) for the return.
2544 // Win32 requires us to put the sret argument to %eax as well.
2545 // Save the argument into a virtual register so that we can access it
2546 // from the return points.
2547 if (Ins[i].Flags.isSRet()) {
2548 unsigned Reg = FuncInfo->getSRetReturnReg();
2550 MVT PtrTy = getPointerTy();
2551 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2552 FuncInfo->setSRetReturnReg(Reg);
2554 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2555 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2561 unsigned StackSize = CCInfo.getNextStackOffset();
2562 // Align stack specially for tail calls.
2563 if (FuncIsMadeTailCallSafe(CallConv,
2564 MF.getTarget().Options.GuaranteedTailCallOpt))
2565 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2567 // If the function takes variable number of arguments, make a frame index for
2568 // the start of the first vararg value... for expansion of llvm.va_start. We
2569 // can skip this if there are no va_start calls.
2570 if (MFI->hasVAStart() &&
2571 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2572 CallConv != CallingConv::X86_ThisCall))) {
2573 FuncInfo->setVarArgsFrameIndex(
2574 MFI->CreateFixedObject(1, StackSize, true));
2577 // Figure out if XMM registers are in use.
2578 assert(!(MF.getTarget().Options.UseSoftFloat &&
2579 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2580 "SSE register cannot be used when SSE is disabled!");
2582 // 64-bit calling conventions support varargs and register parameters, so we
2583 // have to do extra work to spill them in the prologue.
2584 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2585 // Find the first unallocated argument registers.
2586 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2587 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2588 unsigned NumIntRegs =
2589 CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size());
2590 unsigned NumXMMRegs =
2591 CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size());
2592 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2593 "SSE register cannot be used when SSE is disabled!");
2595 // Gather all the live in physical registers.
2596 SmallVector<SDValue, 6> LiveGPRs;
2597 SmallVector<SDValue, 8> LiveXMMRegs;
2599 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2600 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2602 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2604 if (!ArgXMMs.empty()) {
2605 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2606 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2607 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2608 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2609 LiveXMMRegs.push_back(
2610 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2615 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2616 // Get to the caller-allocated home save location. Add 8 to account
2617 // for the return address.
2618 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2619 FuncInfo->setRegSaveFrameIndex(
2620 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2621 // Fixup to set vararg frame on shadow area (4 x i64).
2623 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2625 // For X86-64, if there are vararg parameters that are passed via
2626 // registers, then we must store them to their spots on the stack so
2627 // they may be loaded by deferencing the result of va_next.
2628 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2629 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2630 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2631 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2634 // Store the integer parameter registers.
2635 SmallVector<SDValue, 8> MemOps;
2636 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2638 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2639 for (SDValue Val : LiveGPRs) {
2640 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2641 DAG.getIntPtrConstant(Offset));
2643 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2644 MachinePointerInfo::getFixedStack(
2645 FuncInfo->getRegSaveFrameIndex(), Offset),
2647 MemOps.push_back(Store);
2651 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2652 // Now store the XMM (fp + vector) parameter registers.
2653 SmallVector<SDValue, 12> SaveXMMOps;
2654 SaveXMMOps.push_back(Chain);
2655 SaveXMMOps.push_back(ALVal);
2656 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2657 FuncInfo->getRegSaveFrameIndex()));
2658 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2659 FuncInfo->getVarArgsFPOffset()));
2660 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2662 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2663 MVT::Other, SaveXMMOps));
2666 if (!MemOps.empty())
2667 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2670 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2671 // Find the largest legal vector type.
2672 MVT VecVT = MVT::Other;
2673 // FIXME: Only some x86_32 calling conventions support AVX512.
2674 if (Subtarget->hasAVX512() &&
2675 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2676 CallConv == CallingConv::Intel_OCL_BI)))
2677 VecVT = MVT::v16f32;
2678 else if (Subtarget->hasAVX())
2680 else if (Subtarget->hasSSE2())
2683 // We forward some GPRs and some vector types.
2684 SmallVector<MVT, 2> RegParmTypes;
2685 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2686 RegParmTypes.push_back(IntVT);
2687 if (VecVT != MVT::Other)
2688 RegParmTypes.push_back(VecVT);
2690 // Compute the set of forwarded registers. The rest are scratch.
2691 SmallVectorImpl<ForwardedRegister> &Forwards =
2692 FuncInfo->getForwardedMustTailRegParms();
2693 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2695 // Conservatively forward AL on x86_64, since it might be used for varargs.
2696 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2697 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2698 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2701 // Copy all forwards from physical to virtual registers.
2702 for (ForwardedRegister &F : Forwards) {
2703 // FIXME: Can we use a less constrained schedule?
2704 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2705 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2706 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2710 // Some CCs need callee pop.
2711 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2712 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2713 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2715 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2716 // If this is an sret function, the return should pop the hidden pointer.
2717 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2718 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2719 argsAreStructReturn(Ins) == StackStructReturn)
2720 FuncInfo->setBytesToPopOnReturn(4);
2724 // RegSaveFrameIndex is X86-64 only.
2725 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2726 if (CallConv == CallingConv::X86_FastCall ||
2727 CallConv == CallingConv::X86_ThisCall)
2728 // fastcc functions can't have varargs.
2729 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2732 FuncInfo->setArgumentStackSize(StackSize);
2738 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2739 SDValue StackPtr, SDValue Arg,
2740 SDLoc dl, SelectionDAG &DAG,
2741 const CCValAssign &VA,
2742 ISD::ArgFlagsTy Flags) const {
2743 unsigned LocMemOffset = VA.getLocMemOffset();
2744 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2745 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2746 if (Flags.isByVal())
2747 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2749 return DAG.getStore(Chain, dl, Arg, PtrOff,
2750 MachinePointerInfo::getStack(LocMemOffset),
2754 /// Emit a load of return address if tail call
2755 /// optimization is performed and it is required.
2757 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2758 SDValue &OutRetAddr, SDValue Chain,
2759 bool IsTailCall, bool Is64Bit,
2760 int FPDiff, SDLoc dl) const {
2761 // Adjust the Return address stack slot.
2762 EVT VT = getPointerTy();
2763 OutRetAddr = getReturnAddressFrameIndex(DAG);
2765 // Load the "old" Return address.
2766 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2767 false, false, false, 0);
2768 return SDValue(OutRetAddr.getNode(), 1);
2771 /// Emit a store of the return address if tail call
2772 /// optimization is performed and it is required (FPDiff!=0).
2773 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2774 SDValue Chain, SDValue RetAddrFrIdx,
2775 EVT PtrVT, unsigned SlotSize,
2776 int FPDiff, SDLoc dl) {
2777 // Store the return address to the appropriate stack slot.
2778 if (!FPDiff) return Chain;
2779 // Calculate the new stack slot for the return address.
2780 int NewReturnAddrFI =
2781 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2783 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2784 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2785 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2791 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2792 SmallVectorImpl<SDValue> &InVals) const {
2793 SelectionDAG &DAG = CLI.DAG;
2795 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2796 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2797 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2798 SDValue Chain = CLI.Chain;
2799 SDValue Callee = CLI.Callee;
2800 CallingConv::ID CallConv = CLI.CallConv;
2801 bool &isTailCall = CLI.IsTailCall;
2802 bool isVarArg = CLI.IsVarArg;
2804 MachineFunction &MF = DAG.getMachineFunction();
2805 bool Is64Bit = Subtarget->is64Bit();
2806 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2807 StructReturnType SR = callIsStructReturn(Outs);
2808 bool IsSibcall = false;
2809 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2811 if (MF.getTarget().Options.DisableTailCalls)
2814 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2816 // Force this to be a tail call. The verifier rules are enough to ensure
2817 // that we can lower this successfully without moving the return address
2820 } else if (isTailCall) {
2821 // Check if it's really possible to do a tail call.
2822 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2823 isVarArg, SR != NotStructReturn,
2824 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2825 Outs, OutVals, Ins, DAG);
2827 // Sibcalls are automatically detected tailcalls which do not require
2829 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2836 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2837 "Var args not supported with calling convention fastcc, ghc or hipe");
2839 // Analyze operands of the call, assigning locations to each operand.
2840 SmallVector<CCValAssign, 16> ArgLocs;
2841 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2843 // Allocate shadow area for Win64
2845 CCInfo.AllocateStack(32, 8);
2847 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2849 // Get a count of how many bytes are to be pushed on the stack.
2850 unsigned NumBytes = CCInfo.getNextStackOffset();
2852 // This is a sibcall. The memory operands are available in caller's
2853 // own caller's stack.
2855 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2856 IsTailCallConvention(CallConv))
2857 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2860 if (isTailCall && !IsSibcall && !IsMustTail) {
2861 // Lower arguments at fp - stackoffset + fpdiff.
2862 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2864 FPDiff = NumBytesCallerPushed - NumBytes;
2866 // Set the delta of movement of the returnaddr stackslot.
2867 // But only set if delta is greater than previous delta.
2868 if (FPDiff < X86Info->getTCReturnAddrDelta())
2869 X86Info->setTCReturnAddrDelta(FPDiff);
2872 unsigned NumBytesToPush = NumBytes;
2873 unsigned NumBytesToPop = NumBytes;
2875 // If we have an inalloca argument, all stack space has already been allocated
2876 // for us and be right at the top of the stack. We don't support multiple
2877 // arguments passed in memory when using inalloca.
2878 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2880 if (!ArgLocs.back().isMemLoc())
2881 report_fatal_error("cannot use inalloca attribute on a register "
2883 if (ArgLocs.back().getLocMemOffset() != 0)
2884 report_fatal_error("any parameter with the inalloca attribute must be "
2885 "the only memory argument");
2889 Chain = DAG.getCALLSEQ_START(
2890 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2892 SDValue RetAddrFrIdx;
2893 // Load return address for tail calls.
2894 if (isTailCall && FPDiff)
2895 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2896 Is64Bit, FPDiff, dl);
2898 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2899 SmallVector<SDValue, 8> MemOpChains;
2902 // Walk the register/memloc assignments, inserting copies/loads. In the case
2903 // of tail call optimization arguments are handle later.
2904 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2905 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2906 // Skip inalloca arguments, they have already been written.
2907 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2908 if (Flags.isInAlloca())
2911 CCValAssign &VA = ArgLocs[i];
2912 EVT RegVT = VA.getLocVT();
2913 SDValue Arg = OutVals[i];
2914 bool isByVal = Flags.isByVal();
2916 // Promote the value if needed.
2917 switch (VA.getLocInfo()) {
2918 default: llvm_unreachable("Unknown loc info!");
2919 case CCValAssign::Full: break;
2920 case CCValAssign::SExt:
2921 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2923 case CCValAssign::ZExt:
2924 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2926 case CCValAssign::AExt:
2927 if (RegVT.is128BitVector()) {
2928 // Special case: passing MMX values in XMM registers.
2929 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2930 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2931 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2933 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2935 case CCValAssign::BCvt:
2936 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2938 case CCValAssign::Indirect: {
2939 // Store the argument.
2940 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2941 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2942 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2943 MachinePointerInfo::getFixedStack(FI),
2950 if (VA.isRegLoc()) {
2951 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2952 if (isVarArg && IsWin64) {
2953 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2954 // shadow reg if callee is a varargs function.
2955 unsigned ShadowReg = 0;
2956 switch (VA.getLocReg()) {
2957 case X86::XMM0: ShadowReg = X86::RCX; break;
2958 case X86::XMM1: ShadowReg = X86::RDX; break;
2959 case X86::XMM2: ShadowReg = X86::R8; break;
2960 case X86::XMM3: ShadowReg = X86::R9; break;
2963 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2965 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2966 assert(VA.isMemLoc());
2967 if (!StackPtr.getNode())
2968 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2970 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2971 dl, DAG, VA, Flags));
2975 if (!MemOpChains.empty())
2976 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2978 if (Subtarget->isPICStyleGOT()) {
2979 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2982 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2983 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2985 // If we are tail calling and generating PIC/GOT style code load the
2986 // address of the callee into ECX. The value in ecx is used as target of
2987 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2988 // for tail calls on PIC/GOT architectures. Normally we would just put the
2989 // address of GOT into ebx and then call target@PLT. But for tail calls
2990 // ebx would be restored (since ebx is callee saved) before jumping to the
2993 // Note: The actual moving to ECX is done further down.
2994 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2995 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2996 !G->getGlobal()->hasProtectedVisibility())
2997 Callee = LowerGlobalAddress(Callee, DAG);
2998 else if (isa<ExternalSymbolSDNode>(Callee))
2999 Callee = LowerExternalSymbol(Callee, DAG);
3003 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3004 // From AMD64 ABI document:
3005 // For calls that may call functions that use varargs or stdargs
3006 // (prototype-less calls or calls to functions containing ellipsis (...) in
3007 // the declaration) %al is used as hidden argument to specify the number
3008 // of SSE registers used. The contents of %al do not need to match exactly
3009 // the number of registers, but must be an ubound on the number of SSE
3010 // registers used and is in the range 0 - 8 inclusive.
3012 // Count the number of XMM registers allocated.
3013 static const MCPhysReg XMMArgRegs[] = {
3014 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3015 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3017 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
3018 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3019 && "SSE registers cannot be used when SSE is disabled");
3021 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3022 DAG.getConstant(NumXMMRegs, MVT::i8)));
3025 if (isVarArg && IsMustTail) {
3026 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3027 for (const auto &F : Forwards) {
3028 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3029 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3033 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3034 // don't need this because the eligibility check rejects calls that require
3035 // shuffling arguments passed in memory.
3036 if (!IsSibcall && isTailCall) {
3037 // Force all the incoming stack arguments to be loaded from the stack
3038 // before any new outgoing arguments are stored to the stack, because the
3039 // outgoing stack slots may alias the incoming argument stack slots, and
3040 // the alias isn't otherwise explicit. This is slightly more conservative
3041 // than necessary, because it means that each store effectively depends
3042 // on every argument instead of just those arguments it would clobber.
3043 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3045 SmallVector<SDValue, 8> MemOpChains2;
3048 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3049 CCValAssign &VA = ArgLocs[i];
3052 assert(VA.isMemLoc());
3053 SDValue Arg = OutVals[i];
3054 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3055 // Skip inalloca arguments. They don't require any work.
3056 if (Flags.isInAlloca())
3058 // Create frame index.
3059 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3060 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3061 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3062 FIN = DAG.getFrameIndex(FI, getPointerTy());
3064 if (Flags.isByVal()) {
3065 // Copy relative to framepointer.
3066 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3067 if (!StackPtr.getNode())
3068 StackPtr = DAG.getCopyFromReg(Chain, dl,
3069 RegInfo->getStackRegister(),
3071 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3073 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3077 // Store relative to framepointer.
3078 MemOpChains2.push_back(
3079 DAG.getStore(ArgChain, dl, Arg, FIN,
3080 MachinePointerInfo::getFixedStack(FI),
3085 if (!MemOpChains2.empty())
3086 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3088 // Store the return address to the appropriate stack slot.
3089 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3090 getPointerTy(), RegInfo->getSlotSize(),
3094 // Build a sequence of copy-to-reg nodes chained together with token chain
3095 // and flag operands which copy the outgoing args into registers.
3097 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3098 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3099 RegsToPass[i].second, InFlag);
3100 InFlag = Chain.getValue(1);
3103 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3104 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3105 // In the 64-bit large code model, we have to make all calls
3106 // through a register, since the call instruction's 32-bit
3107 // pc-relative offset may not be large enough to hold the whole
3109 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3110 // If the callee is a GlobalAddress node (quite common, every direct call
3111 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3113 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3115 // We should use extra load for direct calls to dllimported functions in
3117 const GlobalValue *GV = G->getGlobal();
3118 if (!GV->hasDLLImportStorageClass()) {
3119 unsigned char OpFlags = 0;
3120 bool ExtraLoad = false;
3121 unsigned WrapperKind = ISD::DELETED_NODE;
3123 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3124 // external symbols most go through the PLT in PIC mode. If the symbol
3125 // has hidden or protected visibility, or if it is static or local, then
3126 // we don't need to use the PLT - we can directly call it.
3127 if (Subtarget->isTargetELF() &&
3128 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3129 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3130 OpFlags = X86II::MO_PLT;
3131 } else if (Subtarget->isPICStyleStubAny() &&
3132 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3133 (!Subtarget->getTargetTriple().isMacOSX() ||
3134 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3135 // PC-relative references to external symbols should go through $stub,
3136 // unless we're building with the leopard linker or later, which
3137 // automatically synthesizes these stubs.
3138 OpFlags = X86II::MO_DARWIN_STUB;
3139 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3140 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3141 // If the function is marked as non-lazy, generate an indirect call
3142 // which loads from the GOT directly. This avoids runtime overhead
3143 // at the cost of eager binding (and one extra byte of encoding).
3144 OpFlags = X86II::MO_GOTPCREL;
3145 WrapperKind = X86ISD::WrapperRIP;
3149 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3150 G->getOffset(), OpFlags);
3152 // Add a wrapper if needed.
3153 if (WrapperKind != ISD::DELETED_NODE)
3154 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3155 // Add extra indirection if needed.
3157 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3158 MachinePointerInfo::getGOT(),
3159 false, false, false, 0);
3161 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3162 unsigned char OpFlags = 0;
3164 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3165 // external symbols should go through the PLT.
3166 if (Subtarget->isTargetELF() &&
3167 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3168 OpFlags = X86II::MO_PLT;
3169 } else if (Subtarget->isPICStyleStubAny() &&
3170 (!Subtarget->getTargetTriple().isMacOSX() ||
3171 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3172 // PC-relative references to external symbols should go through $stub,
3173 // unless we're building with the leopard linker or later, which
3174 // automatically synthesizes these stubs.
3175 OpFlags = X86II::MO_DARWIN_STUB;
3178 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3180 } else if (Subtarget->isTarget64BitILP32() &&
3181 Callee->getValueType(0) == MVT::i32) {
3182 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3183 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3186 // Returns a chain & a flag for retval copy to use.
3187 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3188 SmallVector<SDValue, 8> Ops;
3190 if (!IsSibcall && isTailCall) {
3191 Chain = DAG.getCALLSEQ_END(Chain,
3192 DAG.getIntPtrConstant(NumBytesToPop, true),
3193 DAG.getIntPtrConstant(0, true), InFlag, dl);
3194 InFlag = Chain.getValue(1);
3197 Ops.push_back(Chain);
3198 Ops.push_back(Callee);
3201 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3203 // Add argument registers to the end of the list so that they are known live
3205 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3206 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3207 RegsToPass[i].second.getValueType()));
3209 // Add a register mask operand representing the call-preserved registers.
3210 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3211 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3212 assert(Mask && "Missing call preserved mask for calling convention");
3213 Ops.push_back(DAG.getRegisterMask(Mask));
3215 if (InFlag.getNode())
3216 Ops.push_back(InFlag);
3220 //// If this is the first return lowered for this function, add the regs
3221 //// to the liveout set for the function.
3222 // This isn't right, although it's probably harmless on x86; liveouts
3223 // should be computed from returns not tail calls. Consider a void
3224 // function making a tail call to a function returning int.
3225 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3228 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3229 InFlag = Chain.getValue(1);
3231 // Create the CALLSEQ_END node.
3232 unsigned NumBytesForCalleeToPop;
3233 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3234 DAG.getTarget().Options.GuaranteedTailCallOpt))
3235 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3236 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3237 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3238 SR == StackStructReturn)
3239 // If this is a call to a struct-return function, the callee
3240 // pops the hidden struct pointer, so we have to push it back.
3241 // This is common for Darwin/X86, Linux & Mingw32 targets.
3242 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3243 NumBytesForCalleeToPop = 4;
3245 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3247 // Returns a flag for retval copy to use.
3249 Chain = DAG.getCALLSEQ_END(Chain,
3250 DAG.getIntPtrConstant(NumBytesToPop, true),
3251 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3254 InFlag = Chain.getValue(1);
3257 // Handle result values, copying them out of physregs into vregs that we
3259 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3260 Ins, dl, DAG, InVals);
3263 //===----------------------------------------------------------------------===//
3264 // Fast Calling Convention (tail call) implementation
3265 //===----------------------------------------------------------------------===//
3267 // Like std call, callee cleans arguments, convention except that ECX is
3268 // reserved for storing the tail called function address. Only 2 registers are
3269 // free for argument passing (inreg). Tail call optimization is performed
3271 // * tailcallopt is enabled
3272 // * caller/callee are fastcc
3273 // On X86_64 architecture with GOT-style position independent code only local
3274 // (within module) calls are supported at the moment.
3275 // To keep the stack aligned according to platform abi the function
3276 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3277 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3278 // If a tail called function callee has more arguments than the caller the
3279 // caller needs to make sure that there is room to move the RETADDR to. This is
3280 // achieved by reserving an area the size of the argument delta right after the
3281 // original RETADDR, but before the saved framepointer or the spilled registers
3282 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3294 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3295 /// for a 16 byte align requirement.
3297 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3298 SelectionDAG& DAG) const {
3299 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3300 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3301 unsigned StackAlignment = TFI.getStackAlignment();
3302 uint64_t AlignMask = StackAlignment - 1;
3303 int64_t Offset = StackSize;
3304 unsigned SlotSize = RegInfo->getSlotSize();
3305 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3306 // Number smaller than 12 so just add the difference.
3307 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3309 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3310 Offset = ((~AlignMask) & Offset) + StackAlignment +
3311 (StackAlignment-SlotSize);
3316 /// MatchingStackOffset - Return true if the given stack call argument is
3317 /// already available in the same position (relatively) of the caller's
3318 /// incoming argument stack.
3320 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3321 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3322 const X86InstrInfo *TII) {
3323 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3325 if (Arg.getOpcode() == ISD::CopyFromReg) {
3326 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3327 if (!TargetRegisterInfo::isVirtualRegister(VR))
3329 MachineInstr *Def = MRI->getVRegDef(VR);
3332 if (!Flags.isByVal()) {
3333 if (!TII->isLoadFromStackSlot(Def, FI))
3336 unsigned Opcode = Def->getOpcode();
3337 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3338 Opcode == X86::LEA64_32r) &&
3339 Def->getOperand(1).isFI()) {
3340 FI = Def->getOperand(1).getIndex();
3341 Bytes = Flags.getByValSize();
3345 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3346 if (Flags.isByVal())
3347 // ByVal argument is passed in as a pointer but it's now being
3348 // dereferenced. e.g.
3349 // define @foo(%struct.X* %A) {
3350 // tail call @bar(%struct.X* byval %A)
3353 SDValue Ptr = Ld->getBasePtr();
3354 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3357 FI = FINode->getIndex();
3358 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3359 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3360 FI = FINode->getIndex();
3361 Bytes = Flags.getByValSize();
3365 assert(FI != INT_MAX);
3366 if (!MFI->isFixedObjectIndex(FI))
3368 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3371 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3372 /// for tail call optimization. Targets which want to do tail call
3373 /// optimization should implement this function.
3375 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3376 CallingConv::ID CalleeCC,
3378 bool isCalleeStructRet,
3379 bool isCallerStructRet,
3381 const SmallVectorImpl<ISD::OutputArg> &Outs,
3382 const SmallVectorImpl<SDValue> &OutVals,
3383 const SmallVectorImpl<ISD::InputArg> &Ins,
3384 SelectionDAG &DAG) const {
3385 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3388 // If -tailcallopt is specified, make fastcc functions tail-callable.
3389 const MachineFunction &MF = DAG.getMachineFunction();
3390 const Function *CallerF = MF.getFunction();
3392 // If the function return type is x86_fp80 and the callee return type is not,
3393 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3394 // perform a tailcall optimization here.
3395 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3398 CallingConv::ID CallerCC = CallerF->getCallingConv();
3399 bool CCMatch = CallerCC == CalleeCC;
3400 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3401 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3403 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3404 if (IsTailCallConvention(CalleeCC) && CCMatch)
3409 // Look for obvious safe cases to perform tail call optimization that do not
3410 // require ABI changes. This is what gcc calls sibcall.
3412 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3413 // emit a special epilogue.
3414 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3415 if (RegInfo->needsStackRealignment(MF))
3418 // Also avoid sibcall optimization if either caller or callee uses struct
3419 // return semantics.
3420 if (isCalleeStructRet || isCallerStructRet)
3423 // An stdcall/thiscall caller is expected to clean up its arguments; the
3424 // callee isn't going to do that.
3425 // FIXME: this is more restrictive than needed. We could produce a tailcall
3426 // when the stack adjustment matches. For example, with a thiscall that takes
3427 // only one argument.
3428 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3429 CallerCC == CallingConv::X86_ThisCall))
3432 // Do not sibcall optimize vararg calls unless all arguments are passed via
3434 if (isVarArg && !Outs.empty()) {
3436 // Optimizing for varargs on Win64 is unlikely to be safe without
3437 // additional testing.
3438 if (IsCalleeWin64 || IsCallerWin64)
3441 SmallVector<CCValAssign, 16> ArgLocs;
3442 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3445 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3446 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3447 if (!ArgLocs[i].isRegLoc())
3451 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3452 // stack. Therefore, if it's not used by the call it is not safe to optimize
3453 // this into a sibcall.
3454 bool Unused = false;
3455 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3462 SmallVector<CCValAssign, 16> RVLocs;
3463 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3465 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3466 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3467 CCValAssign &VA = RVLocs[i];
3468 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3473 // If the calling conventions do not match, then we'd better make sure the
3474 // results are returned in the same way as what the caller expects.
3476 SmallVector<CCValAssign, 16> RVLocs1;
3477 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3479 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3481 SmallVector<CCValAssign, 16> RVLocs2;
3482 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3484 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3486 if (RVLocs1.size() != RVLocs2.size())
3488 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3489 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3491 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3493 if (RVLocs1[i].isRegLoc()) {
3494 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3497 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3503 // If the callee takes no arguments then go on to check the results of the
3505 if (!Outs.empty()) {
3506 // Check if stack adjustment is needed. For now, do not do this if any
3507 // argument is passed on the stack.
3508 SmallVector<CCValAssign, 16> ArgLocs;
3509 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3512 // Allocate shadow area for Win64
3514 CCInfo.AllocateStack(32, 8);
3516 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3517 if (CCInfo.getNextStackOffset()) {
3518 MachineFunction &MF = DAG.getMachineFunction();
3519 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3522 // Check if the arguments are already laid out in the right way as
3523 // the caller's fixed stack objects.
3524 MachineFrameInfo *MFI = MF.getFrameInfo();
3525 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3526 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3528 CCValAssign &VA = ArgLocs[i];
3529 SDValue Arg = OutVals[i];
3530 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3531 if (VA.getLocInfo() == CCValAssign::Indirect)
3533 if (!VA.isRegLoc()) {
3534 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3541 // If the tailcall address may be in a register, then make sure it's
3542 // possible to register allocate for it. In 32-bit, the call address can
3543 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3544 // callee-saved registers are restored. These happen to be the same
3545 // registers used to pass 'inreg' arguments so watch out for those.
3546 if (!Subtarget->is64Bit() &&
3547 ((!isa<GlobalAddressSDNode>(Callee) &&
3548 !isa<ExternalSymbolSDNode>(Callee)) ||
3549 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3550 unsigned NumInRegs = 0;
3551 // In PIC we need an extra register to formulate the address computation
3553 unsigned MaxInRegs =
3554 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3556 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3557 CCValAssign &VA = ArgLocs[i];
3560 unsigned Reg = VA.getLocReg();
3563 case X86::EAX: case X86::EDX: case X86::ECX:
3564 if (++NumInRegs == MaxInRegs)
3576 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3577 const TargetLibraryInfo *libInfo) const {
3578 return X86::createFastISel(funcInfo, libInfo);
3581 //===----------------------------------------------------------------------===//
3582 // Other Lowering Hooks
3583 //===----------------------------------------------------------------------===//
3585 static bool MayFoldLoad(SDValue Op) {
3586 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3589 static bool MayFoldIntoStore(SDValue Op) {
3590 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3593 static bool isTargetShuffle(unsigned Opcode) {
3595 default: return false;
3596 case X86ISD::BLENDI:
3597 case X86ISD::PSHUFB:
3598 case X86ISD::PSHUFD:
3599 case X86ISD::PSHUFHW:
3600 case X86ISD::PSHUFLW:
3602 case X86ISD::PALIGNR:
3603 case X86ISD::MOVLHPS:
3604 case X86ISD::MOVLHPD:
3605 case X86ISD::MOVHLPS:
3606 case X86ISD::MOVLPS:
3607 case X86ISD::MOVLPD:
3608 case X86ISD::MOVSHDUP:
3609 case X86ISD::MOVSLDUP:
3610 case X86ISD::MOVDDUP:
3613 case X86ISD::UNPCKL:
3614 case X86ISD::UNPCKH:
3615 case X86ISD::VPERMILPI:
3616 case X86ISD::VPERM2X128:
3617 case X86ISD::VPERMI:
3622 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3623 SDValue V1, SelectionDAG &DAG) {
3625 default: llvm_unreachable("Unknown x86 shuffle node");
3626 case X86ISD::MOVSHDUP:
3627 case X86ISD::MOVSLDUP:
3628 case X86ISD::MOVDDUP:
3629 return DAG.getNode(Opc, dl, VT, V1);
3633 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3634 SDValue V1, unsigned TargetMask,
3635 SelectionDAG &DAG) {
3637 default: llvm_unreachable("Unknown x86 shuffle node");
3638 case X86ISD::PSHUFD:
3639 case X86ISD::PSHUFHW:
3640 case X86ISD::PSHUFLW:
3641 case X86ISD::VPERMILPI:
3642 case X86ISD::VPERMI:
3643 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3647 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3648 SDValue V1, SDValue V2, unsigned TargetMask,
3649 SelectionDAG &DAG) {
3651 default: llvm_unreachable("Unknown x86 shuffle node");
3652 case X86ISD::PALIGNR:
3653 case X86ISD::VALIGN:
3655 case X86ISD::VPERM2X128:
3656 return DAG.getNode(Opc, dl, VT, V1, V2,
3657 DAG.getConstant(TargetMask, MVT::i8));
3661 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3662 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3664 default: llvm_unreachable("Unknown x86 shuffle node");
3665 case X86ISD::MOVLHPS:
3666 case X86ISD::MOVLHPD:
3667 case X86ISD::MOVHLPS:
3668 case X86ISD::MOVLPS:
3669 case X86ISD::MOVLPD:
3672 case X86ISD::UNPCKL:
3673 case X86ISD::UNPCKH:
3674 return DAG.getNode(Opc, dl, VT, V1, V2);
3678 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3679 MachineFunction &MF = DAG.getMachineFunction();
3680 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3681 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3682 int ReturnAddrIndex = FuncInfo->getRAIndex();
3684 if (ReturnAddrIndex == 0) {
3685 // Set up a frame object for the return address.
3686 unsigned SlotSize = RegInfo->getSlotSize();
3687 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3690 FuncInfo->setRAIndex(ReturnAddrIndex);
3693 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3696 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3697 bool hasSymbolicDisplacement) {
3698 // Offset should fit into 32 bit immediate field.
3699 if (!isInt<32>(Offset))
3702 // If we don't have a symbolic displacement - we don't have any extra
3704 if (!hasSymbolicDisplacement)
3707 // FIXME: Some tweaks might be needed for medium code model.
3708 if (M != CodeModel::Small && M != CodeModel::Kernel)
3711 // For small code model we assume that latest object is 16MB before end of 31
3712 // bits boundary. We may also accept pretty large negative constants knowing
3713 // that all objects are in the positive half of address space.
3714 if (M == CodeModel::Small && Offset < 16*1024*1024)
3717 // For kernel code model we know that all object resist in the negative half
3718 // of 32bits address space. We may not accept negative offsets, since they may
3719 // be just off and we may accept pretty large positive ones.
3720 if (M == CodeModel::Kernel && Offset >= 0)
3726 /// isCalleePop - Determines whether the callee is required to pop its
3727 /// own arguments. Callee pop is necessary to support tail calls.
3728 bool X86::isCalleePop(CallingConv::ID CallingConv,
3729 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3730 switch (CallingConv) {
3733 case CallingConv::X86_StdCall:
3734 case CallingConv::X86_FastCall:
3735 case CallingConv::X86_ThisCall:
3737 case CallingConv::Fast:
3738 case CallingConv::GHC:
3739 case CallingConv::HiPE:
3746 /// \brief Return true if the condition is an unsigned comparison operation.
3747 static bool isX86CCUnsigned(unsigned X86CC) {
3749 default: llvm_unreachable("Invalid integer condition!");
3750 case X86::COND_E: return true;
3751 case X86::COND_G: return false;
3752 case X86::COND_GE: return false;
3753 case X86::COND_L: return false;
3754 case X86::COND_LE: return false;
3755 case X86::COND_NE: return true;
3756 case X86::COND_B: return true;
3757 case X86::COND_A: return true;
3758 case X86::COND_BE: return true;
3759 case X86::COND_AE: return true;
3761 llvm_unreachable("covered switch fell through?!");
3764 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3765 /// specific condition code, returning the condition code and the LHS/RHS of the
3766 /// comparison to make.
3767 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3768 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3770 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3771 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3772 // X > -1 -> X == 0, jump !sign.
3773 RHS = DAG.getConstant(0, RHS.getValueType());
3774 return X86::COND_NS;
3776 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3777 // X < 0 -> X == 0, jump on sign.
3780 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3782 RHS = DAG.getConstant(0, RHS.getValueType());
3783 return X86::COND_LE;
3787 switch (SetCCOpcode) {
3788 default: llvm_unreachable("Invalid integer condition!");
3789 case ISD::SETEQ: return X86::COND_E;
3790 case ISD::SETGT: return X86::COND_G;
3791 case ISD::SETGE: return X86::COND_GE;
3792 case ISD::SETLT: return X86::COND_L;
3793 case ISD::SETLE: return X86::COND_LE;
3794 case ISD::SETNE: return X86::COND_NE;
3795 case ISD::SETULT: return X86::COND_B;
3796 case ISD::SETUGT: return X86::COND_A;
3797 case ISD::SETULE: return X86::COND_BE;
3798 case ISD::SETUGE: return X86::COND_AE;
3802 // First determine if it is required or is profitable to flip the operands.
3804 // If LHS is a foldable load, but RHS is not, flip the condition.
3805 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3806 !ISD::isNON_EXTLoad(RHS.getNode())) {
3807 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3808 std::swap(LHS, RHS);
3811 switch (SetCCOpcode) {
3817 std::swap(LHS, RHS);
3821 // On a floating point condition, the flags are set as follows:
3823 // 0 | 0 | 0 | X > Y
3824 // 0 | 0 | 1 | X < Y
3825 // 1 | 0 | 0 | X == Y
3826 // 1 | 1 | 1 | unordered
3827 switch (SetCCOpcode) {
3828 default: llvm_unreachable("Condcode should be pre-legalized away");
3830 case ISD::SETEQ: return X86::COND_E;
3831 case ISD::SETOLT: // flipped
3833 case ISD::SETGT: return X86::COND_A;
3834 case ISD::SETOLE: // flipped
3836 case ISD::SETGE: return X86::COND_AE;
3837 case ISD::SETUGT: // flipped
3839 case ISD::SETLT: return X86::COND_B;
3840 case ISD::SETUGE: // flipped
3842 case ISD::SETLE: return X86::COND_BE;
3844 case ISD::SETNE: return X86::COND_NE;
3845 case ISD::SETUO: return X86::COND_P;
3846 case ISD::SETO: return X86::COND_NP;
3848 case ISD::SETUNE: return X86::COND_INVALID;
3852 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3853 /// code. Current x86 isa includes the following FP cmov instructions:
3854 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3855 static bool hasFPCMov(unsigned X86CC) {
3871 /// isFPImmLegal - Returns true if the target can instruction select the
3872 /// specified FP immediate natively. If false, the legalizer will
3873 /// materialize the FP immediate as a load from a constant pool.
3874 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3875 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3876 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3882 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3883 ISD::LoadExtType ExtTy,
3885 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3886 // relocation target a movq or addq instruction: don't let the load shrink.
3887 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3888 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3889 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3890 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3894 /// \brief Returns true if it is beneficial to convert a load of a constant
3895 /// to just the constant itself.
3896 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3898 assert(Ty->isIntegerTy());
3900 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3901 if (BitSize == 0 || BitSize > 64)
3906 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3907 unsigned Index) const {
3908 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3911 return (Index == 0 || Index == ResVT.getVectorNumElements());
3914 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3915 // Speculate cttz only if we can directly use TZCNT.
3916 return Subtarget->hasBMI();
3919 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3920 // Speculate ctlz only if we can directly use LZCNT.
3921 return Subtarget->hasLZCNT();
3924 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3925 /// the specified range (L, H].
3926 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3927 return (Val < 0) || (Val >= Low && Val < Hi);
3930 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3931 /// specified value.
3932 static bool isUndefOrEqual(int Val, int CmpVal) {
3933 return (Val < 0 || Val == CmpVal);
3936 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3937 /// from position Pos and ending in Pos+Size, falls within the specified
3938 /// sequential range (Low, Low+Size]. or is undef.
3939 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3940 unsigned Pos, unsigned Size, int Low) {
3941 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3942 if (!isUndefOrEqual(Mask[i], Low))
3947 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3948 /// is suitable for input to PSHUFD. That is, it doesn't reference the other
3949 /// operand - by default will match for first operand.
3950 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT,
3951 bool TestSecondOperand = false) {
3952 if (VT != MVT::v4f32 && VT != MVT::v4i32 &&
3953 VT != MVT::v2f64 && VT != MVT::v2i64)
3956 unsigned NumElems = VT.getVectorNumElements();
3957 unsigned Lo = TestSecondOperand ? NumElems : 0;
3958 unsigned Hi = Lo + NumElems;
3960 for (unsigned i = 0; i < NumElems; ++i)
3961 if (!isUndefOrInRange(Mask[i], (int)Lo, (int)Hi))
3967 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3968 /// is suitable for input to PSHUFHW.
3969 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3970 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3973 // Lower quadword copied in order or undef.
3974 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3977 // Upper quadword shuffled.
3978 for (unsigned i = 4; i != 8; ++i)
3979 if (!isUndefOrInRange(Mask[i], 4, 8))
3982 if (VT == MVT::v16i16) {
3983 // Lower quadword copied in order or undef.
3984 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3987 // Upper quadword shuffled.
3988 for (unsigned i = 12; i != 16; ++i)
3989 if (!isUndefOrInRange(Mask[i], 12, 16))
3996 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3997 /// is suitable for input to PSHUFLW.
3998 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3999 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
4002 // Upper quadword copied in order.
4003 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
4006 // Lower quadword shuffled.
4007 for (unsigned i = 0; i != 4; ++i)
4008 if (!isUndefOrInRange(Mask[i], 0, 4))
4011 if (VT == MVT::v16i16) {
4012 // Upper quadword copied in order.
4013 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
4016 // Lower quadword shuffled.
4017 for (unsigned i = 8; i != 12; ++i)
4018 if (!isUndefOrInRange(Mask[i], 8, 12))
4025 /// \brief Return true if the mask specifies a shuffle of elements that is
4026 /// suitable for input to intralane (palignr) or interlane (valign) vector
4028 static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) {
4029 unsigned NumElts = VT.getVectorNumElements();
4030 unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128;
4031 unsigned NumLaneElts = NumElts/NumLanes;
4033 // Do not handle 64-bit element shuffles with palignr.
4034 if (NumLaneElts == 2)
4037 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
4039 for (i = 0; i != NumLaneElts; ++i) {
4044 // Lane is all undef, go to next lane
4045 if (i == NumLaneElts)
4048 int Start = Mask[i+l];
4050 // Make sure its in this lane in one of the sources
4051 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
4052 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
4055 // If not lane 0, then we must match lane 0
4056 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
4059 // Correct second source to be contiguous with first source
4060 if (Start >= (int)NumElts)
4061 Start -= NumElts - NumLaneElts;
4063 // Make sure we're shifting in the right direction.
4064 if (Start <= (int)(i+l))
4069 // Check the rest of the elements to see if they are consecutive.
4070 for (++i; i != NumLaneElts; ++i) {
4071 int Idx = Mask[i+l];
4073 // Make sure its in this lane
4074 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
4075 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
4078 // If not lane 0, then we must match lane 0
4079 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
4082 if (Idx >= (int)NumElts)
4083 Idx -= NumElts - NumLaneElts;
4085 if (!isUndefOrEqual(Idx, Start+i))
4094 /// \brief Return true if the node specifies a shuffle of elements that is
4095 /// suitable for input to PALIGNR.
4096 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
4097 const X86Subtarget *Subtarget) {
4098 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
4099 (VT.is256BitVector() && !Subtarget->hasInt256()) ||
4100 VT.is512BitVector())
4101 // FIXME: Add AVX512BW.
4104 return isAlignrMask(Mask, VT, false);
4107 /// \brief Return true if the node specifies a shuffle of elements that is
4108 /// suitable for input to VALIGN.
4109 static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT,
4110 const X86Subtarget *Subtarget) {
4111 // FIXME: Add AVX512VL.
4112 if (!VT.is512BitVector() || !Subtarget->hasAVX512())
4114 return isAlignrMask(Mask, VT, true);
4117 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4118 /// the two vector operands have swapped position.
4119 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
4120 unsigned NumElems) {
4121 for (unsigned i = 0; i != NumElems; ++i) {
4125 else if (idx < (int)NumElems)
4126 Mask[i] = idx + NumElems;
4128 Mask[i] = idx - NumElems;
4132 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
4133 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
4134 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
4135 /// reverse of what x86 shuffles want.
4136 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
4138 unsigned NumElems = VT.getVectorNumElements();
4139 unsigned NumLanes = VT.getSizeInBits()/128;
4140 unsigned NumLaneElems = NumElems/NumLanes;
4142 if (NumLaneElems != 2 && NumLaneElems != 4)
4145 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4146 bool symmetricMaskRequired =
4147 (VT.getSizeInBits() >= 256) && (EltSize == 32);
4149 // VSHUFPSY divides the resulting vector into 4 chunks.
4150 // The sources are also splitted into 4 chunks, and each destination
4151 // chunk must come from a different source chunk.
4153 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
4154 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
4156 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
4157 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
4159 // VSHUFPDY divides the resulting vector into 4 chunks.
4160 // The sources are also splitted into 4 chunks, and each destination
4161 // chunk must come from a different source chunk.
4163 // SRC1 => X3 X2 X1 X0
4164 // SRC2 => Y3 Y2 Y1 Y0
4166 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
4168 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
4169 unsigned HalfLaneElems = NumLaneElems/2;
4170 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
4171 for (unsigned i = 0; i != NumLaneElems; ++i) {
4172 int Idx = Mask[i+l];
4173 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
4174 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
4176 // For VSHUFPSY, the mask of the second half must be the same as the
4177 // first but with the appropriate offsets. This works in the same way as
4178 // VPERMILPS works with masks.
4179 if (!symmetricMaskRequired || Idx < 0)
4181 if (MaskVal[i] < 0) {
4182 MaskVal[i] = Idx - l;
4185 if ((signed)(Idx - l) != MaskVal[i])
4193 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
4194 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
4195 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
4196 if (!VT.is128BitVector())
4199 unsigned NumElems = VT.getVectorNumElements();
4204 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
4205 return isUndefOrEqual(Mask[0], 6) &&
4206 isUndefOrEqual(Mask[1], 7) &&
4207 isUndefOrEqual(Mask[2], 2) &&
4208 isUndefOrEqual(Mask[3], 3);
4211 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
4212 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
4214 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
4215 if (!VT.is128BitVector())
4218 unsigned NumElems = VT.getVectorNumElements();
4223 return isUndefOrEqual(Mask[0], 2) &&
4224 isUndefOrEqual(Mask[1], 3) &&
4225 isUndefOrEqual(Mask[2], 2) &&
4226 isUndefOrEqual(Mask[3], 3);
4229 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
4230 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
4231 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
4232 if (!VT.is128BitVector())
4235 unsigned NumElems = VT.getVectorNumElements();
4237 if (NumElems != 2 && NumElems != 4)
4240 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4241 if (!isUndefOrEqual(Mask[i], i + NumElems))
4244 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4245 if (!isUndefOrEqual(Mask[i], i))
4251 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
4252 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
4253 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
4254 if (!VT.is128BitVector())
4257 unsigned NumElems = VT.getVectorNumElements();
4259 if (NumElems != 2 && NumElems != 4)
4262 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4263 if (!isUndefOrEqual(Mask[i], i))
4266 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4267 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
4273 /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
4274 /// specifies a shuffle of elements that is suitable for input to INSERTPS.
4275 /// i. e: If all but one element come from the same vector.
4276 static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
4277 // TODO: Deal with AVX's VINSERTPS
4278 if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
4281 unsigned CorrectPosV1 = 0;
4282 unsigned CorrectPosV2 = 0;
4283 for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
4284 if (Mask[i] == -1) {
4292 else if (Mask[i] == i + 4)
4296 if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
4297 // We have 3 elements (undefs count as elements from any vector) from one
4298 // vector, and one from another.
4305 // Some special combinations that can be optimized.
4308 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
4309 SelectionDAG &DAG) {
4310 MVT VT = SVOp->getSimpleValueType(0);
4313 if (VT != MVT::v8i32 && VT != MVT::v8f32)
4316 ArrayRef<int> Mask = SVOp->getMask();
4318 // These are the special masks that may be optimized.
4319 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
4320 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
4321 bool MatchEvenMask = true;
4322 bool MatchOddMask = true;
4323 for (int i=0; i<8; ++i) {
4324 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
4325 MatchEvenMask = false;
4326 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
4327 MatchOddMask = false;
4330 if (!MatchEvenMask && !MatchOddMask)
4333 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
4335 SDValue Op0 = SVOp->getOperand(0);
4336 SDValue Op1 = SVOp->getOperand(1);
4338 if (MatchEvenMask) {
4339 // Shift the second operand right to 32 bits.
4340 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
4341 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
4343 // Shift the first operand left to 32 bits.
4344 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
4345 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
4347 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
4348 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
4351 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
4352 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
4353 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
4354 bool HasInt256, bool V2IsSplat = false) {
4356 assert(VT.getSizeInBits() >= 128 &&
4357 "Unsupported vector type for unpckl");
4359 unsigned NumElts = VT.getVectorNumElements();
4360 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4361 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4364 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4365 "Unsupported vector type for unpckh");
4367 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4368 unsigned NumLanes = VT.getSizeInBits()/128;
4369 unsigned NumLaneElts = NumElts/NumLanes;
4371 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4372 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4373 int BitI = Mask[l+i];
4374 int BitI1 = Mask[l+i+1];
4375 if (!isUndefOrEqual(BitI, j))
4378 if (!isUndefOrEqual(BitI1, NumElts))
4381 if (!isUndefOrEqual(BitI1, j + NumElts))
4390 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
4391 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
4392 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
4393 bool HasInt256, bool V2IsSplat = false) {
4394 assert(VT.getSizeInBits() >= 128 &&
4395 "Unsupported vector type for unpckh");
4397 unsigned NumElts = VT.getVectorNumElements();
4398 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4399 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4402 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4403 "Unsupported vector type for unpckh");
4405 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4406 unsigned NumLanes = VT.getSizeInBits()/128;
4407 unsigned NumLaneElts = NumElts/NumLanes;
4409 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4410 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4411 int BitI = Mask[l+i];
4412 int BitI1 = Mask[l+i+1];
4413 if (!isUndefOrEqual(BitI, j))
4416 if (isUndefOrEqual(BitI1, NumElts))
4419 if (!isUndefOrEqual(BitI1, j+NumElts))
4427 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
4428 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
4430 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4431 unsigned NumElts = VT.getVectorNumElements();
4432 bool Is256BitVec = VT.is256BitVector();
4434 if (VT.is512BitVector())
4436 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4437 "Unsupported vector type for unpckh");
4439 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
4440 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4443 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
4444 // FIXME: Need a better way to get rid of this, there's no latency difference
4445 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
4446 // the former later. We should also remove the "_undef" special mask.
4447 if (NumElts == 4 && Is256BitVec)
4450 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4451 // independently on 128-bit lanes.
4452 unsigned NumLanes = VT.getSizeInBits()/128;
4453 unsigned NumLaneElts = NumElts/NumLanes;
4455 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4456 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4457 int BitI = Mask[l+i];
4458 int BitI1 = Mask[l+i+1];
4460 if (!isUndefOrEqual(BitI, j))
4462 if (!isUndefOrEqual(BitI1, j))
4470 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4471 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4473 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4474 unsigned NumElts = VT.getVectorNumElements();
4476 if (VT.is512BitVector())
4479 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4480 "Unsupported vector type for unpckh");
4482 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4483 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4486 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4487 // independently on 128-bit lanes.
4488 unsigned NumLanes = VT.getSizeInBits()/128;
4489 unsigned NumLaneElts = NumElts/NumLanes;
4491 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4492 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4493 int BitI = Mask[l+i];
4494 int BitI1 = Mask[l+i+1];
4495 if (!isUndefOrEqual(BitI, j))
4497 if (!isUndefOrEqual(BitI1, j))
4504 // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
4505 // (src1[0], src0[1]), manipulation with 256-bit sub-vectors
4506 static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
4507 if (!VT.is512BitVector())
4510 unsigned NumElts = VT.getVectorNumElements();
4511 unsigned HalfSize = NumElts/2;
4512 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
4513 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
4518 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
4519 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
4527 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4528 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4529 /// MOVSD, and MOVD, i.e. setting the lowest element.
4530 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4531 if (VT.getVectorElementType().getSizeInBits() < 32)
4533 if (!VT.is128BitVector())
4536 unsigned NumElts = VT.getVectorNumElements();
4538 if (!isUndefOrEqual(Mask[0], NumElts))
4541 for (unsigned i = 1; i != NumElts; ++i)
4542 if (!isUndefOrEqual(Mask[i], i))
4548 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4549 /// as permutations between 128-bit chunks or halves. As an example: this
4551 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4552 /// The first half comes from the second half of V1 and the second half from the
4553 /// the second half of V2.
4554 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4555 if (!HasFp256 || !VT.is256BitVector())
4558 // The shuffle result is divided into half A and half B. In total the two
4559 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4560 // B must come from C, D, E or F.
4561 unsigned HalfSize = VT.getVectorNumElements()/2;
4562 bool MatchA = false, MatchB = false;
4564 // Check if A comes from one of C, D, E, F.
4565 for (unsigned Half = 0; Half != 4; ++Half) {
4566 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4572 // Check if B comes from one of C, D, E, F.
4573 for (unsigned Half = 0; Half != 4; ++Half) {
4574 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4580 return MatchA && MatchB;
4583 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4584 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4585 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4586 MVT VT = SVOp->getSimpleValueType(0);
4588 unsigned HalfSize = VT.getVectorNumElements()/2;
4590 unsigned FstHalf = 0, SndHalf = 0;
4591 for (unsigned i = 0; i < HalfSize; ++i) {
4592 if (SVOp->getMaskElt(i) > 0) {
4593 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4597 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4598 if (SVOp->getMaskElt(i) > 0) {
4599 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4604 return (FstHalf | (SndHalf << 4));
4607 // Symmetric in-lane mask. Each lane has 4 elements (for imm8)
4608 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4609 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4613 unsigned NumElts = VT.getVectorNumElements();
4615 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4616 for (unsigned i = 0; i != NumElts; ++i) {
4619 Imm8 |= Mask[i] << (i*2);
4624 unsigned LaneSize = 4;
4625 SmallVector<int, 4> MaskVal(LaneSize, -1);
4627 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4628 for (unsigned i = 0; i != LaneSize; ++i) {
4629 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4633 if (MaskVal[i] < 0) {
4634 MaskVal[i] = Mask[i+l] - l;
4635 Imm8 |= MaskVal[i] << (i*2);
4638 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4645 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4646 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4647 /// Note that VPERMIL mask matching is different depending whether theunderlying
4648 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4649 /// to the same elements of the low, but to the higher half of the source.
4650 /// In VPERMILPD the two lanes could be shuffled independently of each other
4651 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4652 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4653 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4654 if (VT.getSizeInBits() < 256 || EltSize < 32)
4656 bool symmetricMaskRequired = (EltSize == 32);
4657 unsigned NumElts = VT.getVectorNumElements();
4659 unsigned NumLanes = VT.getSizeInBits()/128;
4660 unsigned LaneSize = NumElts/NumLanes;
4661 // 2 or 4 elements in one lane
4663 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4664 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4665 for (unsigned i = 0; i != LaneSize; ++i) {
4666 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4668 if (symmetricMaskRequired) {
4669 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4670 ExpectedMaskVal[i] = Mask[i+l] - l;
4673 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4681 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4682 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4683 /// element of vector 2 and the other elements to come from vector 1 in order.
4684 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4685 bool V2IsSplat = false, bool V2IsUndef = false) {
4686 if (!VT.is128BitVector())
4689 unsigned NumOps = VT.getVectorNumElements();
4690 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4693 if (!isUndefOrEqual(Mask[0], 0))
4696 for (unsigned i = 1; i != NumOps; ++i)
4697 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4698 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4699 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4705 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4706 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4707 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4708 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4709 const X86Subtarget *Subtarget) {
4710 if (!Subtarget->hasSSE3())
4713 unsigned NumElems = VT.getVectorNumElements();
4715 if ((VT.is128BitVector() && NumElems != 4) ||
4716 (VT.is256BitVector() && NumElems != 8) ||
4717 (VT.is512BitVector() && NumElems != 16))
4720 // "i+1" is the value the indexed mask element must have
4721 for (unsigned i = 0; i != NumElems; i += 2)
4722 if (!isUndefOrEqual(Mask[i], i+1) ||
4723 !isUndefOrEqual(Mask[i+1], i+1))
4729 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4730 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4731 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4732 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4733 const X86Subtarget *Subtarget) {
4734 if (!Subtarget->hasSSE3())
4737 unsigned NumElems = VT.getVectorNumElements();
4739 if ((VT.is128BitVector() && NumElems != 4) ||
4740 (VT.is256BitVector() && NumElems != 8) ||
4741 (VT.is512BitVector() && NumElems != 16))
4744 // "i" is the value the indexed mask element must have
4745 for (unsigned i = 0; i != NumElems; i += 2)
4746 if (!isUndefOrEqual(Mask[i], i) ||
4747 !isUndefOrEqual(Mask[i+1], i))
4753 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4754 /// specifies a shuffle of elements that is suitable for input to 256-bit
4755 /// version of MOVDDUP.
4756 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4757 if (!HasFp256 || !VT.is256BitVector())
4760 unsigned NumElts = VT.getVectorNumElements();
4764 for (unsigned i = 0; i != NumElts/2; ++i)
4765 if (!isUndefOrEqual(Mask[i], 0))
4767 for (unsigned i = NumElts/2; i != NumElts; ++i)
4768 if (!isUndefOrEqual(Mask[i], NumElts/2))
4773 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4774 /// specifies a shuffle of elements that is suitable for input to 128-bit
4775 /// version of MOVDDUP.
4776 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4777 if (!VT.is128BitVector())
4780 unsigned e = VT.getVectorNumElements() / 2;
4781 for (unsigned i = 0; i != e; ++i)
4782 if (!isUndefOrEqual(Mask[i], i))
4784 for (unsigned i = 0; i != e; ++i)
4785 if (!isUndefOrEqual(Mask[e+i], i))
4790 /// isVEXTRACTIndex - Return true if the specified
4791 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4792 /// suitable for instruction that extract 128 or 256 bit vectors
4793 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4794 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4795 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4798 // The index should be aligned on a vecWidth-bit boundary.
4800 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4802 MVT VT = N->getSimpleValueType(0);
4803 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4804 bool Result = (Index * ElSize) % vecWidth == 0;
4809 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4810 /// operand specifies a subvector insert that is suitable for input to
4811 /// insertion of 128 or 256-bit subvectors
4812 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4813 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4814 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4816 // The index should be aligned on a vecWidth-bit boundary.
4818 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4820 MVT VT = N->getSimpleValueType(0);
4821 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4822 bool Result = (Index * ElSize) % vecWidth == 0;
4827 bool X86::isVINSERT128Index(SDNode *N) {
4828 return isVINSERTIndex(N, 128);
4831 bool X86::isVINSERT256Index(SDNode *N) {
4832 return isVINSERTIndex(N, 256);
4835 bool X86::isVEXTRACT128Index(SDNode *N) {
4836 return isVEXTRACTIndex(N, 128);
4839 bool X86::isVEXTRACT256Index(SDNode *N) {
4840 return isVEXTRACTIndex(N, 256);
4843 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4844 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4845 /// Handles 128-bit and 256-bit.
4846 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4847 MVT VT = N->getSimpleValueType(0);
4849 assert((VT.getSizeInBits() >= 128) &&
4850 "Unsupported vector type for PSHUF/SHUFP");
4852 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4853 // independently on 128-bit lanes.
4854 unsigned NumElts = VT.getVectorNumElements();
4855 unsigned NumLanes = VT.getSizeInBits()/128;
4856 unsigned NumLaneElts = NumElts/NumLanes;
4858 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4859 "Only supports 2, 4 or 8 elements per lane");
4861 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4863 for (unsigned i = 0; i != NumElts; ++i) {
4864 int Elt = N->getMaskElt(i);
4865 if (Elt < 0) continue;
4866 Elt &= NumLaneElts - 1;
4867 unsigned ShAmt = (i << Shift) % 8;
4868 Mask |= Elt << ShAmt;
4874 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4875 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4876 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4877 MVT VT = N->getSimpleValueType(0);
4879 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4880 "Unsupported vector type for PSHUFHW");
4882 unsigned NumElts = VT.getVectorNumElements();
4885 for (unsigned l = 0; l != NumElts; l += 8) {
4886 // 8 nodes per lane, but we only care about the last 4.
4887 for (unsigned i = 0; i < 4; ++i) {
4888 int Elt = N->getMaskElt(l+i+4);
4889 if (Elt < 0) continue;
4890 Elt &= 0x3; // only 2-bits.
4891 Mask |= Elt << (i * 2);
4898 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4899 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4900 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4901 MVT VT = N->getSimpleValueType(0);
4903 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4904 "Unsupported vector type for PSHUFHW");
4906 unsigned NumElts = VT.getVectorNumElements();
4909 for (unsigned l = 0; l != NumElts; l += 8) {
4910 // 8 nodes per lane, but we only care about the first 4.
4911 for (unsigned i = 0; i < 4; ++i) {
4912 int Elt = N->getMaskElt(l+i);
4913 if (Elt < 0) continue;
4914 Elt &= 0x3; // only 2-bits
4915 Mask |= Elt << (i * 2);
4922 /// \brief Return the appropriate immediate to shuffle the specified
4923 /// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with
4924 /// VALIGN (if Interlane is true) instructions.
4925 static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp,
4927 MVT VT = SVOp->getSimpleValueType(0);
4928 unsigned EltSize = InterLane ? 1 :
4929 VT.getVectorElementType().getSizeInBits() >> 3;
4931 unsigned NumElts = VT.getVectorNumElements();
4932 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4933 unsigned NumLaneElts = NumElts/NumLanes;
4937 for (i = 0; i != NumElts; ++i) {
4938 Val = SVOp->getMaskElt(i);
4942 if (Val >= (int)NumElts)
4943 Val -= NumElts - NumLaneElts;
4945 assert(Val - i > 0 && "PALIGNR imm should be positive");
4946 return (Val - i) * EltSize;
4949 /// \brief Return the appropriate immediate to shuffle the specified
4950 /// VECTOR_SHUFFLE mask with the PALIGNR instruction.
4951 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4952 return getShuffleAlignrImmediate(SVOp, false);
4955 /// \brief Return the appropriate immediate to shuffle the specified
4956 /// VECTOR_SHUFFLE mask with the VALIGN instruction.
4957 static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) {
4958 return getShuffleAlignrImmediate(SVOp, true);
4962 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4963 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4964 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4965 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4968 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4970 MVT VecVT = N->getOperand(0).getSimpleValueType();
4971 MVT ElVT = VecVT.getVectorElementType();
4973 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4974 return Index / NumElemsPerChunk;
4977 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4978 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4979 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4980 llvm_unreachable("Illegal insert subvector for VINSERT");
4983 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4985 MVT VecVT = N->getSimpleValueType(0);
4986 MVT ElVT = VecVT.getVectorElementType();
4988 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4989 return Index / NumElemsPerChunk;
4992 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4993 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4994 /// and VINSERTI128 instructions.
4995 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4996 return getExtractVEXTRACTImmediate(N, 128);
4999 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
5000 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
5001 /// and VINSERTI64x4 instructions.
5002 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
5003 return getExtractVEXTRACTImmediate(N, 256);
5006 /// getInsertVINSERT128Immediate - Return the appropriate immediate
5007 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
5008 /// and VINSERTI128 instructions.
5009 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
5010 return getInsertVINSERTImmediate(N, 128);
5013 /// getInsertVINSERT256Immediate - Return the appropriate immediate
5014 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
5015 /// and VINSERTI64x4 instructions.
5016 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
5017 return getInsertVINSERTImmediate(N, 256);
5020 /// isZero - Returns true if Elt is a constant integer zero
5021 static bool isZero(SDValue V) {
5022 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
5023 return C && C->isNullValue();
5026 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
5028 bool X86::isZeroNode(SDValue Elt) {
5031 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
5032 return CFP->getValueAPF().isPosZero();
5036 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
5037 /// match movhlps. The lower half elements should come from upper half of
5038 /// V1 (and in order), and the upper half elements should come from the upper
5039 /// half of V2 (and in order).
5040 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
5041 if (!VT.is128BitVector())
5043 if (VT.getVectorNumElements() != 4)
5045 for (unsigned i = 0, e = 2; i != e; ++i)
5046 if (!isUndefOrEqual(Mask[i], i+2))
5048 for (unsigned i = 2; i != 4; ++i)
5049 if (!isUndefOrEqual(Mask[i], i+4))
5054 /// isScalarLoadToVector - Returns true if the node is a scalar load that
5055 /// is promoted to a vector. It also returns the LoadSDNode by reference if
5057 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
5058 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
5060 N = N->getOperand(0).getNode();
5061 if (!ISD::isNON_EXTLoad(N))
5064 *LD = cast<LoadSDNode>(N);
5068 // Test whether the given value is a vector value which will be legalized
5070 static bool WillBeConstantPoolLoad(SDNode *N) {
5071 if (N->getOpcode() != ISD::BUILD_VECTOR)
5074 // Check for any non-constant elements.
5075 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5076 switch (N->getOperand(i).getNode()->getOpcode()) {
5078 case ISD::ConstantFP:
5085 // Vectors of all-zeros and all-ones are materialized with special
5086 // instructions rather than being loaded.
5087 return !ISD::isBuildVectorAllZeros(N) &&
5088 !ISD::isBuildVectorAllOnes(N);
5091 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
5092 /// match movlp{s|d}. The lower half elements should come from lower half of
5093 /// V1 (and in order), and the upper half elements should come from the upper
5094 /// half of V2 (and in order). And since V1 will become the source of the
5095 /// MOVLP, it must be either a vector load or a scalar load to vector.
5096 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
5097 ArrayRef<int> Mask, MVT VT) {
5098 if (!VT.is128BitVector())
5101 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
5103 // Is V2 is a vector load, don't do this transformation. We will try to use
5104 // load folding shufps op.
5105 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
5108 unsigned NumElems = VT.getVectorNumElements();
5110 if (NumElems != 2 && NumElems != 4)
5112 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
5113 if (!isUndefOrEqual(Mask[i], i))
5115 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
5116 if (!isUndefOrEqual(Mask[i], i+NumElems))
5121 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
5122 /// to an zero vector.
5123 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
5124 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
5125 SDValue V1 = N->getOperand(0);
5126 SDValue V2 = N->getOperand(1);
5127 unsigned NumElems = N->getValueType(0).getVectorNumElements();
5128 for (unsigned i = 0; i != NumElems; ++i) {
5129 int Idx = N->getMaskElt(i);
5130 if (Idx >= (int)NumElems) {
5131 unsigned Opc = V2.getOpcode();
5132 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
5134 if (Opc != ISD::BUILD_VECTOR ||
5135 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
5137 } else if (Idx >= 0) {
5138 unsigned Opc = V1.getOpcode();
5139 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
5141 if (Opc != ISD::BUILD_VECTOR ||
5142 !X86::isZeroNode(V1.getOperand(Idx)))
5149 /// getZeroVector - Returns a vector of specified type with all zero elements.
5151 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
5152 SelectionDAG &DAG, SDLoc dl) {
5153 assert(VT.isVector() && "Expected a vector type");
5155 // Always build SSE zero vectors as <4 x i32> bitcasted
5156 // to their dest type. This ensures they get CSE'd.
5158 if (VT.is128BitVector()) { // SSE
5159 if (Subtarget->hasSSE2()) { // SSE2
5160 SDValue Cst = DAG.getConstant(0, MVT::i32);
5161 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5163 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5164 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
5166 } else if (VT.is256BitVector()) { // AVX
5167 if (Subtarget->hasInt256()) { // AVX2
5168 SDValue Cst = DAG.getConstant(0, MVT::i32);
5169 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5170 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5172 // 256-bit logic and arithmetic instructions in AVX are all
5173 // floating-point, no support for integer ops. Emit fp zeroed vectors.
5174 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5175 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5176 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
5178 } else if (VT.is512BitVector()) { // AVX-512
5179 SDValue Cst = DAG.getConstant(0, MVT::i32);
5180 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5181 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5182 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
5183 } else if (VT.getScalarType() == MVT::i1) {
5184 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
5185 SDValue Cst = DAG.getConstant(0, MVT::i1);
5186 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5187 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5189 llvm_unreachable("Unexpected vector type");
5191 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5194 /// getOnesVector - Returns a vector of specified type with all bits set.
5195 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
5196 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
5197 /// Then bitcast to their original type, ensuring they get CSE'd.
5198 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
5200 assert(VT.isVector() && "Expected a vector type");
5202 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
5204 if (VT.is256BitVector()) {
5205 if (HasInt256) { // AVX2
5206 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5207 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5209 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5210 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
5212 } else if (VT.is128BitVector()) {
5213 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5215 llvm_unreachable("Unexpected vector type");
5217 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5220 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
5221 /// that point to V2 points to its first element.
5222 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
5223 for (unsigned i = 0; i != NumElems; ++i) {
5224 if (Mask[i] > (int)NumElems) {
5230 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
5231 /// operation of specified width.
5232 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
5234 unsigned NumElems = VT.getVectorNumElements();
5235 SmallVector<int, 8> Mask;
5236 Mask.push_back(NumElems);
5237 for (unsigned i = 1; i != NumElems; ++i)
5239 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5242 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
5243 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5245 unsigned NumElems = VT.getVectorNumElements();
5246 SmallVector<int, 8> Mask;
5247 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
5249 Mask.push_back(i + NumElems);
5251 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5254 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
5255 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5257 unsigned NumElems = VT.getVectorNumElements();
5258 SmallVector<int, 8> Mask;
5259 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
5260 Mask.push_back(i + Half);
5261 Mask.push_back(i + NumElems + Half);
5263 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5266 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
5267 // a generic shuffle instruction because the target has no such instructions.
5268 // Generate shuffles which repeat i16 and i8 several times until they can be
5269 // represented by v4f32 and then be manipulated by target suported shuffles.
5270 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
5271 MVT VT = V.getSimpleValueType();
5272 int NumElems = VT.getVectorNumElements();
5275 while (NumElems > 4) {
5276 if (EltNo < NumElems/2) {
5277 V = getUnpackl(DAG, dl, VT, V, V);
5279 V = getUnpackh(DAG, dl, VT, V, V);
5280 EltNo -= NumElems/2;
5287 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
5288 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
5289 MVT VT = V.getSimpleValueType();
5292 if (VT.is128BitVector()) {
5293 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
5294 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
5295 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
5297 } else if (VT.is256BitVector()) {
5298 // To use VPERMILPS to splat scalars, the second half of indicies must
5299 // refer to the higher part, which is a duplication of the lower one,
5300 // because VPERMILPS can only handle in-lane permutations.
5301 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
5302 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
5304 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
5305 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
5308 llvm_unreachable("Vector size not supported");
5310 return DAG.getNode(ISD::BITCAST, dl, VT, V);
5313 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
5314 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
5315 MVT SrcVT = SV->getSimpleValueType(0);
5316 SDValue V1 = SV->getOperand(0);
5319 int EltNo = SV->getSplatIndex();
5320 int NumElems = SrcVT.getVectorNumElements();
5321 bool Is256BitVec = SrcVT.is256BitVector();
5323 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
5324 "Unknown how to promote splat for type");
5326 // Extract the 128-bit part containing the splat element and update
5327 // the splat element index when it refers to the higher register.
5329 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
5330 if (EltNo >= NumElems/2)
5331 EltNo -= NumElems/2;
5334 // All i16 and i8 vector types can't be used directly by a generic shuffle
5335 // instruction because the target has no such instruction. Generate shuffles
5336 // which repeat i16 and i8 several times until they fit in i32, and then can
5337 // be manipulated by target suported shuffles.
5338 MVT EltVT = SrcVT.getVectorElementType();
5339 if (EltVT == MVT::i8 || EltVT == MVT::i16)
5340 V1 = PromoteSplati8i16(V1, DAG, EltNo);
5342 // Recreate the 256-bit vector and place the same 128-bit vector
5343 // into the low and high part. This is necessary because we want
5344 // to use VPERM* to shuffle the vectors
5346 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
5349 return getLegalSplat(DAG, V1, EltNo);
5352 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
5353 /// vector of zero or undef vector. This produces a shuffle where the low
5354 /// element of V2 is swizzled into the zero/undef vector, landing at element
5355 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
5356 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
5358 const X86Subtarget *Subtarget,
5359 SelectionDAG &DAG) {
5360 MVT VT = V2.getSimpleValueType();
5362 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
5363 unsigned NumElems = VT.getVectorNumElements();
5364 SmallVector<int, 16> MaskVec;
5365 for (unsigned i = 0; i != NumElems; ++i)
5366 // If this is the insertion idx, put the low elt of V2 here.
5367 MaskVec.push_back(i == Idx ? NumElems : i);
5368 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
5371 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
5372 /// target specific opcode. Returns true if the Mask could be calculated. Sets
5373 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
5374 /// shuffles which use a single input multiple times, and in those cases it will
5375 /// adjust the mask to only have indices within that single input.
5376 static bool getTargetShuffleMask(SDNode *N, MVT VT,
5377 SmallVectorImpl<int> &Mask, bool &IsUnary) {
5378 unsigned NumElems = VT.getVectorNumElements();
5382 bool IsFakeUnary = false;
5383 switch(N->getOpcode()) {
5384 case X86ISD::BLENDI:
5385 ImmN = N->getOperand(N->getNumOperands()-1);
5386 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5389 ImmN = N->getOperand(N->getNumOperands()-1);
5390 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5391 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5393 case X86ISD::UNPCKH:
5394 DecodeUNPCKHMask(VT, Mask);
5395 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5397 case X86ISD::UNPCKL:
5398 DecodeUNPCKLMask(VT, Mask);
5399 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5401 case X86ISD::MOVHLPS:
5402 DecodeMOVHLPSMask(NumElems, Mask);
5403 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5405 case X86ISD::MOVLHPS:
5406 DecodeMOVLHPSMask(NumElems, Mask);
5407 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5409 case X86ISD::PALIGNR:
5410 ImmN = N->getOperand(N->getNumOperands()-1);
5411 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5413 case X86ISD::PSHUFD:
5414 case X86ISD::VPERMILPI:
5415 ImmN = N->getOperand(N->getNumOperands()-1);
5416 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5419 case X86ISD::PSHUFHW:
5420 ImmN = N->getOperand(N->getNumOperands()-1);
5421 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5424 case X86ISD::PSHUFLW:
5425 ImmN = N->getOperand(N->getNumOperands()-1);
5426 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5429 case X86ISD::PSHUFB: {
5431 SDValue MaskNode = N->getOperand(1);
5432 while (MaskNode->getOpcode() == ISD::BITCAST)
5433 MaskNode = MaskNode->getOperand(0);
5435 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
5436 // If we have a build-vector, then things are easy.
5437 EVT VT = MaskNode.getValueType();
5438 assert(VT.isVector() &&
5439 "Can't produce a non-vector with a build_vector!");
5440 if (!VT.isInteger())
5443 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
5445 SmallVector<uint64_t, 32> RawMask;
5446 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
5447 SDValue Op = MaskNode->getOperand(i);
5448 if (Op->getOpcode() == ISD::UNDEF) {
5449 RawMask.push_back((uint64_t)SM_SentinelUndef);
5452 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
5455 APInt MaskElement = CN->getAPIntValue();
5457 // We now have to decode the element which could be any integer size and
5458 // extract each byte of it.
5459 for (int j = 0; j < NumBytesPerElement; ++j) {
5460 // Note that this is x86 and so always little endian: the low byte is
5461 // the first byte of the mask.
5462 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
5463 MaskElement = MaskElement.lshr(8);
5466 DecodePSHUFBMask(RawMask, Mask);
5470 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
5474 SDValue Ptr = MaskLoad->getBasePtr();
5475 if (Ptr->getOpcode() == X86ISD::Wrapper)
5476 Ptr = Ptr->getOperand(0);
5478 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
5479 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5482 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5483 DecodePSHUFBMask(C, Mask);
5491 case X86ISD::VPERMI:
5492 ImmN = N->getOperand(N->getNumOperands()-1);
5493 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5498 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
5500 case X86ISD::VPERM2X128:
5501 ImmN = N->getOperand(N->getNumOperands()-1);
5502 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5503 if (Mask.empty()) return false;
5505 case X86ISD::MOVSLDUP:
5506 DecodeMOVSLDUPMask(VT, Mask);
5509 case X86ISD::MOVSHDUP:
5510 DecodeMOVSHDUPMask(VT, Mask);
5513 case X86ISD::MOVDDUP:
5514 DecodeMOVDDUPMask(VT, Mask);
5517 case X86ISD::MOVLHPD:
5518 case X86ISD::MOVLPD:
5519 case X86ISD::MOVLPS:
5520 // Not yet implemented
5522 default: llvm_unreachable("unknown target shuffle node");
5525 // If we have a fake unary shuffle, the shuffle mask is spread across two
5526 // inputs that are actually the same node. Re-map the mask to always point
5527 // into the first input.
5530 if (M >= (int)Mask.size())
5536 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5537 /// element of the result of the vector shuffle.
5538 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5541 return SDValue(); // Limit search depth.
5543 SDValue V = SDValue(N, 0);
5544 EVT VT = V.getValueType();
5545 unsigned Opcode = V.getOpcode();
5547 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5548 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5549 int Elt = SV->getMaskElt(Index);
5552 return DAG.getUNDEF(VT.getVectorElementType());
5554 unsigned NumElems = VT.getVectorNumElements();
5555 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5556 : SV->getOperand(1);
5557 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5560 // Recurse into target specific vector shuffles to find scalars.
5561 if (isTargetShuffle(Opcode)) {
5562 MVT ShufVT = V.getSimpleValueType();
5563 unsigned NumElems = ShufVT.getVectorNumElements();
5564 SmallVector<int, 16> ShuffleMask;
5567 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5570 int Elt = ShuffleMask[Index];
5572 return DAG.getUNDEF(ShufVT.getVectorElementType());
5574 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5576 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5580 // Actual nodes that may contain scalar elements
5581 if (Opcode == ISD::BITCAST) {
5582 V = V.getOperand(0);
5583 EVT SrcVT = V.getValueType();
5584 unsigned NumElems = VT.getVectorNumElements();
5586 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5590 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5591 return (Index == 0) ? V.getOperand(0)
5592 : DAG.getUNDEF(VT.getVectorElementType());
5594 if (V.getOpcode() == ISD::BUILD_VECTOR)
5595 return V.getOperand(Index);
5600 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5601 /// shuffle operation which come from a consecutively from a zero. The
5602 /// search can start in two different directions, from left or right.
5603 /// We count undefs as zeros until PreferredNum is reached.
5604 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5605 unsigned NumElems, bool ZerosFromLeft,
5607 unsigned PreferredNum = -1U) {
5608 unsigned NumZeros = 0;
5609 for (unsigned i = 0; i != NumElems; ++i) {
5610 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5611 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5615 if (X86::isZeroNode(Elt))
5617 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5618 NumZeros = std::min(NumZeros + 1, PreferredNum);
5626 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5627 /// correspond consecutively to elements from one of the vector operands,
5628 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5630 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5631 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5632 unsigned NumElems, unsigned &OpNum) {
5633 bool SeenV1 = false;
5634 bool SeenV2 = false;
5636 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5637 int Idx = SVOp->getMaskElt(i);
5638 // Ignore undef indicies
5642 if (Idx < (int)NumElems)
5647 // Only accept consecutive elements from the same vector
5648 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5652 OpNum = SeenV1 ? 0 : 1;
5656 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5657 /// logical left shift of a vector.
5658 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5659 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5661 SVOp->getSimpleValueType(0).getVectorNumElements();
5662 unsigned NumZeros = getNumOfConsecutiveZeros(
5663 SVOp, NumElems, false /* check zeros from right */, DAG,
5664 SVOp->getMaskElt(0));
5670 // Considering the elements in the mask that are not consecutive zeros,
5671 // check if they consecutively come from only one of the source vectors.
5673 // V1 = {X, A, B, C} 0
5675 // vector_shuffle V1, V2 <1, 2, 3, X>
5677 if (!isShuffleMaskConsecutive(SVOp,
5678 0, // Mask Start Index
5679 NumElems-NumZeros, // Mask End Index(exclusive)
5680 NumZeros, // Where to start looking in the src vector
5681 NumElems, // Number of elements in vector
5682 OpSrc)) // Which source operand ?
5687 ShVal = SVOp->getOperand(OpSrc);
5691 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5692 /// logical left shift of a vector.
5693 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5694 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5696 SVOp->getSimpleValueType(0).getVectorNumElements();
5697 unsigned NumZeros = getNumOfConsecutiveZeros(
5698 SVOp, NumElems, true /* check zeros from left */, DAG,
5699 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5705 // Considering the elements in the mask that are not consecutive zeros,
5706 // check if they consecutively come from only one of the source vectors.
5708 // 0 { A, B, X, X } = V2
5710 // vector_shuffle V1, V2 <X, X, 4, 5>
5712 if (!isShuffleMaskConsecutive(SVOp,
5713 NumZeros, // Mask Start Index
5714 NumElems, // Mask End Index(exclusive)
5715 0, // Where to start looking in the src vector
5716 NumElems, // Number of elements in vector
5717 OpSrc)) // Which source operand ?
5722 ShVal = SVOp->getOperand(OpSrc);
5726 /// isVectorShift - Returns true if the shuffle can be implemented as a
5727 /// logical left or right shift of a vector.
5728 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5729 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5730 // Although the logic below support any bitwidth size, there are no
5731 // shift instructions which handle more than 128-bit vectors.
5732 if (!SVOp->getSimpleValueType(0).is128BitVector())
5735 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5736 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5742 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5744 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5745 unsigned NumNonZero, unsigned NumZero,
5747 const X86Subtarget* Subtarget,
5748 const TargetLowering &TLI) {
5755 for (unsigned i = 0; i < 16; ++i) {
5756 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5757 if (ThisIsNonZero && First) {
5759 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5761 V = DAG.getUNDEF(MVT::v8i16);
5766 SDValue ThisElt, LastElt;
5767 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5768 if (LastIsNonZero) {
5769 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5770 MVT::i16, Op.getOperand(i-1));
5772 if (ThisIsNonZero) {
5773 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5774 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5775 ThisElt, DAG.getConstant(8, MVT::i8));
5777 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5781 if (ThisElt.getNode())
5782 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5783 DAG.getIntPtrConstant(i/2));
5787 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5790 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5792 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5793 unsigned NumNonZero, unsigned NumZero,
5795 const X86Subtarget* Subtarget,
5796 const TargetLowering &TLI) {
5803 for (unsigned i = 0; i < 8; ++i) {
5804 bool isNonZero = (NonZeros & (1 << i)) != 0;
5808 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5810 V = DAG.getUNDEF(MVT::v8i16);
5813 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5814 MVT::v8i16, V, Op.getOperand(i),
5815 DAG.getIntPtrConstant(i));
5822 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
5823 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5824 const X86Subtarget *Subtarget,
5825 const TargetLowering &TLI) {
5826 // Find all zeroable elements.
5827 std::bitset<4> Zeroable;
5828 for (int i=0; i < 4; ++i) {
5829 SDValue Elt = Op->getOperand(i);
5830 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5832 assert(Zeroable.size() - Zeroable.count() > 1 &&
5833 "We expect at least two non-zero elements!");
5835 // We only know how to deal with build_vector nodes where elements are either
5836 // zeroable or extract_vector_elt with constant index.
5837 SDValue FirstNonZero;
5838 unsigned FirstNonZeroIdx;
5839 for (unsigned i=0; i < 4; ++i) {
5842 SDValue Elt = Op->getOperand(i);
5843 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5844 !isa<ConstantSDNode>(Elt.getOperand(1)))
5846 // Make sure that this node is extracting from a 128-bit vector.
5847 MVT VT = Elt.getOperand(0).getSimpleValueType();
5848 if (!VT.is128BitVector())
5850 if (!FirstNonZero.getNode()) {
5852 FirstNonZeroIdx = i;
5856 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5857 SDValue V1 = FirstNonZero.getOperand(0);
5858 MVT VT = V1.getSimpleValueType();
5860 // See if this build_vector can be lowered as a blend with zero.
5862 unsigned EltMaskIdx, EltIdx;
5864 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5865 if (Zeroable[EltIdx]) {
5866 // The zero vector will be on the right hand side.
5867 Mask[EltIdx] = EltIdx+4;
5871 Elt = Op->getOperand(EltIdx);
5872 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5873 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5874 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5876 Mask[EltIdx] = EltIdx;
5880 // Let the shuffle legalizer deal with blend operations.
5881 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5882 if (V1.getSimpleValueType() != VT)
5883 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5884 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5887 // See if we can lower this build_vector to a INSERTPS.
5888 if (!Subtarget->hasSSE41())
5891 SDValue V2 = Elt.getOperand(0);
5892 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5895 bool CanFold = true;
5896 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5900 SDValue Current = Op->getOperand(i);
5901 SDValue SrcVector = Current->getOperand(0);
5904 CanFold = SrcVector == V1 &&
5905 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5911 assert(V1.getNode() && "Expected at least two non-zero elements!");
5912 if (V1.getSimpleValueType() != MVT::v4f32)
5913 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5914 if (V2.getSimpleValueType() != MVT::v4f32)
5915 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5917 // Ok, we can emit an INSERTPS instruction.
5918 unsigned ZMask = Zeroable.to_ulong();
5920 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5921 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5922 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
5923 DAG.getIntPtrConstant(InsertPSMask));
5924 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
5927 /// Return a vector logical shift node.
5928 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5929 unsigned NumBits, SelectionDAG &DAG,
5930 const TargetLowering &TLI, SDLoc dl) {
5931 assert(VT.is128BitVector() && "Unknown type for VShift");
5932 MVT ShVT = MVT::v2i64;
5933 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5934 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5935 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
5936 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5937 SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
5938 return DAG.getNode(ISD::BITCAST, dl, VT,
5939 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5943 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5945 // Check if the scalar load can be widened into a vector load. And if
5946 // the address is "base + cst" see if the cst can be "absorbed" into
5947 // the shuffle mask.
5948 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5949 SDValue Ptr = LD->getBasePtr();
5950 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5952 EVT PVT = LD->getValueType(0);
5953 if (PVT != MVT::i32 && PVT != MVT::f32)
5958 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5959 FI = FINode->getIndex();
5961 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5962 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5963 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5964 Offset = Ptr.getConstantOperandVal(1);
5965 Ptr = Ptr.getOperand(0);
5970 // FIXME: 256-bit vector instructions don't require a strict alignment,
5971 // improve this code to support it better.
5972 unsigned RequiredAlign = VT.getSizeInBits()/8;
5973 SDValue Chain = LD->getChain();
5974 // Make sure the stack object alignment is at least 16 or 32.
5975 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5976 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5977 if (MFI->isFixedObjectIndex(FI)) {
5978 // Can't change the alignment. FIXME: It's possible to compute
5979 // the exact stack offset and reference FI + adjust offset instead.
5980 // If someone *really* cares about this. That's the way to implement it.
5983 MFI->setObjectAlignment(FI, RequiredAlign);
5987 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5988 // Ptr + (Offset & ~15).
5991 if ((Offset % RequiredAlign) & 3)
5993 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5995 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5996 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5998 int EltNo = (Offset - StartOffset) >> 2;
5999 unsigned NumElems = VT.getVectorNumElements();
6001 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
6002 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
6003 LD->getPointerInfo().getWithOffset(StartOffset),
6004 false, false, false, 0);
6006 SmallVector<int, 8> Mask(NumElems, EltNo);
6008 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
6014 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
6015 /// elements can be replaced by a single large load which has the same value as
6016 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
6018 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
6020 /// FIXME: we'd also like to handle the case where the last elements are zero
6021 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
6022 /// There's even a handy isZeroNode for that purpose.
6023 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
6024 SDLoc &DL, SelectionDAG &DAG,
6025 bool isAfterLegalize) {
6026 unsigned NumElems = Elts.size();
6028 LoadSDNode *LDBase = nullptr;
6029 unsigned LastLoadedElt = -1U;
6031 // For each element in the initializer, see if we've found a load or an undef.
6032 // If we don't find an initial load element, or later load elements are
6033 // non-consecutive, bail out.
6034 for (unsigned i = 0; i < NumElems; ++i) {
6035 SDValue Elt = Elts[i];
6036 // Look through a bitcast.
6037 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
6038 Elt = Elt.getOperand(0);
6039 if (!Elt.getNode() ||
6040 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6043 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
6045 LDBase = cast<LoadSDNode>(Elt.getNode());
6049 if (Elt.getOpcode() == ISD::UNDEF)
6052 LoadSDNode *LD = cast<LoadSDNode>(Elt);
6053 EVT LdVT = Elt.getValueType();
6054 // Each loaded element must be the correct fractional portion of the
6055 // requested vector load.
6056 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
6058 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
6063 // If we have found an entire vector of loads and undefs, then return a large
6064 // load of the entire vector width starting at the base pointer. If we found
6065 // consecutive loads for the low half, generate a vzext_load node.
6066 if (LastLoadedElt == NumElems - 1) {
6067 assert(LDBase && "Did not find base load for merging consecutive loads");
6068 EVT EltVT = LDBase->getValueType(0);
6069 // Ensure that the input vector size for the merged loads matches the
6070 // cumulative size of the input elements.
6071 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
6074 if (isAfterLegalize &&
6075 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
6078 SDValue NewLd = SDValue();
6080 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
6081 LDBase->getPointerInfo(), LDBase->isVolatile(),
6082 LDBase->isNonTemporal(), LDBase->isInvariant(),
6083 LDBase->getAlignment());
6085 if (LDBase->hasAnyUseOfValue(1)) {
6086 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6088 SDValue(NewLd.getNode(), 1));
6089 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6090 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6091 SDValue(NewLd.getNode(), 1));
6097 //TODO: The code below fires only for for loading the low v2i32 / v2f32
6098 //of a v4i32 / v4f32. It's probably worth generalizing.
6099 EVT EltVT = VT.getVectorElementType();
6100 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
6101 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
6102 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
6103 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
6105 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
6106 LDBase->getPointerInfo(),
6107 LDBase->getAlignment(),
6108 false/*isVolatile*/, true/*ReadMem*/,
6111 // Make sure the newly-created LOAD is in the same position as LDBase in
6112 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
6113 // update uses of LDBase's output chain to use the TokenFactor.
6114 if (LDBase->hasAnyUseOfValue(1)) {
6115 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6116 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
6117 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6118 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6119 SDValue(ResNode.getNode(), 1));
6122 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
6127 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
6128 /// to generate a splat value for the following cases:
6129 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
6130 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
6131 /// a scalar load, or a constant.
6132 /// The VBROADCAST node is returned when a pattern is found,
6133 /// or SDValue() otherwise.
6134 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
6135 SelectionDAG &DAG) {
6136 // VBROADCAST requires AVX.
6137 // TODO: Splats could be generated for non-AVX CPUs using SSE
6138 // instructions, but there's less potential gain for only 128-bit vectors.
6139 if (!Subtarget->hasAVX())
6142 MVT VT = Op.getSimpleValueType();
6145 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
6146 "Unsupported vector type for broadcast.");
6151 switch (Op.getOpcode()) {
6153 // Unknown pattern found.
6156 case ISD::BUILD_VECTOR: {
6157 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
6158 BitVector UndefElements;
6159 SDValue Splat = BVOp->getSplatValue(&UndefElements);
6161 // We need a splat of a single value to use broadcast, and it doesn't
6162 // make any sense if the value is only in one element of the vector.
6163 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
6167 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6168 Ld.getOpcode() == ISD::ConstantFP);
6170 // Make sure that all of the users of a non-constant load are from the
6171 // BUILD_VECTOR node.
6172 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
6177 case ISD::VECTOR_SHUFFLE: {
6178 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6180 // Shuffles must have a splat mask where the first element is
6182 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
6185 SDValue Sc = Op.getOperand(0);
6186 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
6187 Sc.getOpcode() != ISD::BUILD_VECTOR) {
6189 if (!Subtarget->hasInt256())
6192 // Use the register form of the broadcast instruction available on AVX2.
6193 if (VT.getSizeInBits() >= 256)
6194 Sc = Extract128BitVector(Sc, 0, DAG, dl);
6195 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
6198 Ld = Sc.getOperand(0);
6199 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6200 Ld.getOpcode() == ISD::ConstantFP);
6202 // The scalar_to_vector node and the suspected
6203 // load node must have exactly one user.
6204 // Constants may have multiple users.
6206 // AVX-512 has register version of the broadcast
6207 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
6208 Ld.getValueType().getSizeInBits() >= 32;
6209 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
6216 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
6217 bool IsGE256 = (VT.getSizeInBits() >= 256);
6219 // When optimizing for size, generate up to 5 extra bytes for a broadcast
6220 // instruction to save 8 or more bytes of constant pool data.
6221 // TODO: If multiple splats are generated to load the same constant,
6222 // it may be detrimental to overall size. There needs to be a way to detect
6223 // that condition to know if this is truly a size win.
6224 const Function *F = DAG.getMachineFunction().getFunction();
6225 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
6227 // Handle broadcasting a single constant scalar from the constant pool
6229 // On Sandybridge (no AVX2), it is still better to load a constant vector
6230 // from the constant pool and not to broadcast it from a scalar.
6231 // But override that restriction when optimizing for size.
6232 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
6233 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
6234 EVT CVT = Ld.getValueType();
6235 assert(!CVT.isVector() && "Must not broadcast a vector type");
6237 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
6238 // For size optimization, also splat v2f64 and v2i64, and for size opt
6239 // with AVX2, also splat i8 and i16.
6240 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
6241 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6242 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
6243 const Constant *C = nullptr;
6244 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
6245 C = CI->getConstantIntValue();
6246 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
6247 C = CF->getConstantFPValue();
6249 assert(C && "Invalid constant type");
6251 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6252 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
6253 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
6254 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
6255 MachinePointerInfo::getConstantPool(),
6256 false, false, false, Alignment);
6258 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6262 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
6264 // Handle AVX2 in-register broadcasts.
6265 if (!IsLoad && Subtarget->hasInt256() &&
6266 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
6267 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6269 // The scalar source must be a normal load.
6273 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6274 (Subtarget->hasVLX() && ScalarSize == 64))
6275 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6277 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
6278 // double since there is no vbroadcastsd xmm
6279 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
6280 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
6281 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6284 // Unsupported broadcast.
6288 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
6289 /// underlying vector and index.
6291 /// Modifies \p ExtractedFromVec to the real vector and returns the real
6293 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
6295 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
6296 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
6299 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
6301 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
6303 // (extract_vector_elt (vector_shuffle<2,u,u,u>
6304 // (extract_subvector (v8f32 %vreg0), Constant<4>),
6307 // In this case the vector is the extract_subvector expression and the index
6308 // is 2, as specified by the shuffle.
6309 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
6310 SDValue ShuffleVec = SVOp->getOperand(0);
6311 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
6312 assert(ShuffleVecVT.getVectorElementType() ==
6313 ExtractedFromVec.getSimpleValueType().getVectorElementType());
6315 int ShuffleIdx = SVOp->getMaskElt(Idx);
6316 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
6317 ExtractedFromVec = ShuffleVec;
6323 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
6324 MVT VT = Op.getSimpleValueType();
6326 // Skip if insert_vec_elt is not supported.
6327 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6328 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
6332 unsigned NumElems = Op.getNumOperands();
6336 SmallVector<unsigned, 4> InsertIndices;
6337 SmallVector<int, 8> Mask(NumElems, -1);
6339 for (unsigned i = 0; i != NumElems; ++i) {
6340 unsigned Opc = Op.getOperand(i).getOpcode();
6342 if (Opc == ISD::UNDEF)
6345 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
6346 // Quit if more than 1 elements need inserting.
6347 if (InsertIndices.size() > 1)
6350 InsertIndices.push_back(i);
6354 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
6355 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
6356 // Quit if non-constant index.
6357 if (!isa<ConstantSDNode>(ExtIdx))
6359 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
6361 // Quit if extracted from vector of different type.
6362 if (ExtractedFromVec.getValueType() != VT)
6365 if (!VecIn1.getNode())
6366 VecIn1 = ExtractedFromVec;
6367 else if (VecIn1 != ExtractedFromVec) {
6368 if (!VecIn2.getNode())
6369 VecIn2 = ExtractedFromVec;
6370 else if (VecIn2 != ExtractedFromVec)
6371 // Quit if more than 2 vectors to shuffle
6375 if (ExtractedFromVec == VecIn1)
6377 else if (ExtractedFromVec == VecIn2)
6378 Mask[i] = Idx + NumElems;
6381 if (!VecIn1.getNode())
6384 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
6385 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
6386 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
6387 unsigned Idx = InsertIndices[i];
6388 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
6389 DAG.getIntPtrConstant(Idx));
6395 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
6397 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
6399 MVT VT = Op.getSimpleValueType();
6400 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
6401 "Unexpected type in LowerBUILD_VECTORvXi1!");
6404 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6405 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
6406 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6407 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6410 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
6411 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
6412 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6413 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6416 bool AllContants = true;
6417 uint64_t Immediate = 0;
6418 int NonConstIdx = -1;
6419 bool IsSplat = true;
6420 unsigned NumNonConsts = 0;
6421 unsigned NumConsts = 0;
6422 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
6423 SDValue In = Op.getOperand(idx);
6424 if (In.getOpcode() == ISD::UNDEF)
6426 if (!isa<ConstantSDNode>(In)) {
6427 AllContants = false;
6432 if (cast<ConstantSDNode>(In)->getZExtValue())
6433 Immediate |= (1ULL << idx);
6435 if (In != Op.getOperand(0))
6440 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
6441 DAG.getConstant(Immediate, MVT::i16));
6442 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
6443 DAG.getIntPtrConstant(0));
6446 if (NumNonConsts == 1 && NonConstIdx != 0) {
6449 SDValue VecAsImm = DAG.getConstant(Immediate,
6450 MVT::getIntegerVT(VT.getSizeInBits()));
6451 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
6454 DstVec = DAG.getUNDEF(VT);
6455 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
6456 Op.getOperand(NonConstIdx),
6457 DAG.getIntPtrConstant(NonConstIdx));
6459 if (!IsSplat && (NonConstIdx != 0))
6460 llvm_unreachable("Unsupported BUILD_VECTOR operation");
6461 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
6464 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6465 DAG.getConstant(-1, SelectVT),
6466 DAG.getConstant(0, SelectVT));
6468 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6469 DAG.getConstant((Immediate | 1), SelectVT),
6470 DAG.getConstant(Immediate, SelectVT));
6471 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
6474 /// \brief Return true if \p N implements a horizontal binop and return the
6475 /// operands for the horizontal binop into V0 and V1.
6477 /// This is a helper function of PerformBUILD_VECTORCombine.
6478 /// This function checks that the build_vector \p N in input implements a
6479 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
6480 /// operation to match.
6481 /// For example, if \p Opcode is equal to ISD::ADD, then this function
6482 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
6483 /// is equal to ISD::SUB, then this function checks if this is a horizontal
6486 /// This function only analyzes elements of \p N whose indices are
6487 /// in range [BaseIdx, LastIdx).
6488 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
6490 unsigned BaseIdx, unsigned LastIdx,
6491 SDValue &V0, SDValue &V1) {
6492 EVT VT = N->getValueType(0);
6494 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
6495 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
6496 "Invalid Vector in input!");
6498 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
6499 bool CanFold = true;
6500 unsigned ExpectedVExtractIdx = BaseIdx;
6501 unsigned NumElts = LastIdx - BaseIdx;
6502 V0 = DAG.getUNDEF(VT);
6503 V1 = DAG.getUNDEF(VT);
6505 // Check if N implements a horizontal binop.
6506 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
6507 SDValue Op = N->getOperand(i + BaseIdx);
6510 if (Op->getOpcode() == ISD::UNDEF) {
6511 // Update the expected vector extract index.
6512 if (i * 2 == NumElts)
6513 ExpectedVExtractIdx = BaseIdx;
6514 ExpectedVExtractIdx += 2;
6518 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
6523 SDValue Op0 = Op.getOperand(0);
6524 SDValue Op1 = Op.getOperand(1);
6526 // Try to match the following pattern:
6527 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
6528 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6529 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6530 Op0.getOperand(0) == Op1.getOperand(0) &&
6531 isa<ConstantSDNode>(Op0.getOperand(1)) &&
6532 isa<ConstantSDNode>(Op1.getOperand(1)));
6536 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6537 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
6539 if (i * 2 < NumElts) {
6540 if (V0.getOpcode() == ISD::UNDEF)
6541 V0 = Op0.getOperand(0);
6543 if (V1.getOpcode() == ISD::UNDEF)
6544 V1 = Op0.getOperand(0);
6545 if (i * 2 == NumElts)
6546 ExpectedVExtractIdx = BaseIdx;
6549 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
6550 if (I0 == ExpectedVExtractIdx)
6551 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
6552 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
6553 // Try to match the following dag sequence:
6554 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
6555 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
6559 ExpectedVExtractIdx += 2;
6565 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
6566 /// a concat_vector.
6568 /// This is a helper function of PerformBUILD_VECTORCombine.
6569 /// This function expects two 256-bit vectors called V0 and V1.
6570 /// At first, each vector is split into two separate 128-bit vectors.
6571 /// Then, the resulting 128-bit vectors are used to implement two
6572 /// horizontal binary operations.
6574 /// The kind of horizontal binary operation is defined by \p X86Opcode.
6576 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
6577 /// the two new horizontal binop.
6578 /// When Mode is set, the first horizontal binop dag node would take as input
6579 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
6580 /// horizontal binop dag node would take as input the lower 128-bit of V1
6581 /// and the upper 128-bit of V1.
6583 /// HADD V0_LO, V0_HI
6584 /// HADD V1_LO, V1_HI
6586 /// Otherwise, the first horizontal binop dag node takes as input the lower
6587 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
6588 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
6590 /// HADD V0_LO, V1_LO
6591 /// HADD V0_HI, V1_HI
6593 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
6594 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6595 /// the upper 128-bits of the result.
6596 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6597 SDLoc DL, SelectionDAG &DAG,
6598 unsigned X86Opcode, bool Mode,
6599 bool isUndefLO, bool isUndefHI) {
6600 EVT VT = V0.getValueType();
6601 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6602 "Invalid nodes in input!");
6604 unsigned NumElts = VT.getVectorNumElements();
6605 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6606 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6607 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6608 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6609 EVT NewVT = V0_LO.getValueType();
6611 SDValue LO = DAG.getUNDEF(NewVT);
6612 SDValue HI = DAG.getUNDEF(NewVT);
6615 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6616 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6617 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6618 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6619 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6621 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6622 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6623 V1_LO->getOpcode() != ISD::UNDEF))
6624 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6626 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6627 V1_HI->getOpcode() != ISD::UNDEF))
6628 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6631 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6634 /// \brief Try to fold a build_vector that performs an 'addsub' into the
6635 /// sequence of 'vadd + vsub + blendi'.
6636 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
6637 const X86Subtarget *Subtarget) {
6639 EVT VT = BV->getValueType(0);
6640 unsigned NumElts = VT.getVectorNumElements();
6641 SDValue InVec0 = DAG.getUNDEF(VT);
6642 SDValue InVec1 = DAG.getUNDEF(VT);
6644 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6645 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6647 // Odd-numbered elements in the input build vector are obtained from
6648 // adding two integer/float elements.
6649 // Even-numbered elements in the input build vector are obtained from
6650 // subtracting two integer/float elements.
6651 unsigned ExpectedOpcode = ISD::FSUB;
6652 unsigned NextExpectedOpcode = ISD::FADD;
6653 bool AddFound = false;
6654 bool SubFound = false;
6656 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6657 SDValue Op = BV->getOperand(i);
6659 // Skip 'undef' values.
6660 unsigned Opcode = Op.getOpcode();
6661 if (Opcode == ISD::UNDEF) {
6662 std::swap(ExpectedOpcode, NextExpectedOpcode);
6666 // Early exit if we found an unexpected opcode.
6667 if (Opcode != ExpectedOpcode)
6670 SDValue Op0 = Op.getOperand(0);
6671 SDValue Op1 = Op.getOperand(1);
6673 // Try to match the following pattern:
6674 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6675 // Early exit if we cannot match that sequence.
6676 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6677 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6678 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6679 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6680 Op0.getOperand(1) != Op1.getOperand(1))
6683 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6687 // We found a valid add/sub node. Update the information accordingly.
6693 // Update InVec0 and InVec1.
6694 if (InVec0.getOpcode() == ISD::UNDEF)
6695 InVec0 = Op0.getOperand(0);
6696 if (InVec1.getOpcode() == ISD::UNDEF)
6697 InVec1 = Op1.getOperand(0);
6699 // Make sure that operands in input to each add/sub node always
6700 // come from a same pair of vectors.
6701 if (InVec0 != Op0.getOperand(0)) {
6702 if (ExpectedOpcode == ISD::FSUB)
6705 // FADD is commutable. Try to commute the operands
6706 // and then test again.
6707 std::swap(Op0, Op1);
6708 if (InVec0 != Op0.getOperand(0))
6712 if (InVec1 != Op1.getOperand(0))
6715 // Update the pair of expected opcodes.
6716 std::swap(ExpectedOpcode, NextExpectedOpcode);
6719 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6720 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6721 InVec1.getOpcode() != ISD::UNDEF)
6722 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6727 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
6728 const X86Subtarget *Subtarget) {
6730 EVT VT = N->getValueType(0);
6731 unsigned NumElts = VT.getVectorNumElements();
6732 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
6733 SDValue InVec0, InVec1;
6735 // Try to match an ADDSUB.
6736 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
6737 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
6738 SDValue Value = matchAddSub(BV, DAG, Subtarget);
6739 if (Value.getNode())
6743 // Try to match horizontal ADD/SUB.
6744 unsigned NumUndefsLO = 0;
6745 unsigned NumUndefsHI = 0;
6746 unsigned Half = NumElts/2;
6748 // Count the number of UNDEF operands in the build_vector in input.
6749 for (unsigned i = 0, e = Half; i != e; ++i)
6750 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6753 for (unsigned i = Half, e = NumElts; i != e; ++i)
6754 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6757 // Early exit if this is either a build_vector of all UNDEFs or all the
6758 // operands but one are UNDEF.
6759 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6762 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6763 // Try to match an SSE3 float HADD/HSUB.
6764 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6765 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6767 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6768 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6769 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6770 // Try to match an SSSE3 integer HADD/HSUB.
6771 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6772 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6774 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6775 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6778 if (!Subtarget->hasAVX())
6781 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6782 // Try to match an AVX horizontal add/sub of packed single/double
6783 // precision floating point values from 256-bit vectors.
6784 SDValue InVec2, InVec3;
6785 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6786 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6787 ((InVec0.getOpcode() == ISD::UNDEF ||
6788 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6789 ((InVec1.getOpcode() == ISD::UNDEF ||
6790 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6791 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6793 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6794 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6795 ((InVec0.getOpcode() == ISD::UNDEF ||
6796 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6797 ((InVec1.getOpcode() == ISD::UNDEF ||
6798 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6799 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6800 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6801 // Try to match an AVX2 horizontal add/sub of signed integers.
6802 SDValue InVec2, InVec3;
6804 bool CanFold = true;
6806 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6807 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6808 ((InVec0.getOpcode() == ISD::UNDEF ||
6809 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6810 ((InVec1.getOpcode() == ISD::UNDEF ||
6811 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6812 X86Opcode = X86ISD::HADD;
6813 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6814 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6815 ((InVec0.getOpcode() == ISD::UNDEF ||
6816 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6817 ((InVec1.getOpcode() == ISD::UNDEF ||
6818 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6819 X86Opcode = X86ISD::HSUB;
6824 // Fold this build_vector into a single horizontal add/sub.
6825 // Do this only if the target has AVX2.
6826 if (Subtarget->hasAVX2())
6827 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6829 // Do not try to expand this build_vector into a pair of horizontal
6830 // add/sub if we can emit a pair of scalar add/sub.
6831 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6834 // Convert this build_vector into a pair of horizontal binop followed by
6836 bool isUndefLO = NumUndefsLO == Half;
6837 bool isUndefHI = NumUndefsHI == Half;
6838 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6839 isUndefLO, isUndefHI);
6843 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6844 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6846 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6847 X86Opcode = X86ISD::HADD;
6848 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6849 X86Opcode = X86ISD::HSUB;
6850 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6851 X86Opcode = X86ISD::FHADD;
6852 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6853 X86Opcode = X86ISD::FHSUB;
6857 // Don't try to expand this build_vector into a pair of horizontal add/sub
6858 // if we can simply emit a pair of scalar add/sub.
6859 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6862 // Convert this build_vector into two horizontal add/sub followed by
6864 bool isUndefLO = NumUndefsLO == Half;
6865 bool isUndefHI = NumUndefsHI == Half;
6866 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6867 isUndefLO, isUndefHI);
6874 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6877 MVT VT = Op.getSimpleValueType();
6878 MVT ExtVT = VT.getVectorElementType();
6879 unsigned NumElems = Op.getNumOperands();
6881 // Generate vectors for predicate vectors.
6882 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6883 return LowerBUILD_VECTORvXi1(Op, DAG);
6885 // Vectors containing all zeros can be matched by pxor and xorps later
6886 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6887 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6888 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6889 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6892 return getZeroVector(VT, Subtarget, DAG, dl);
6895 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6896 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6897 // vpcmpeqd on 256-bit vectors.
6898 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6899 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6902 if (!VT.is512BitVector())
6903 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
6906 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
6907 if (Broadcast.getNode())
6910 unsigned EVTBits = ExtVT.getSizeInBits();
6912 unsigned NumZero = 0;
6913 unsigned NumNonZero = 0;
6914 unsigned NonZeros = 0;
6915 bool IsAllConstants = true;
6916 SmallSet<SDValue, 8> Values;
6917 for (unsigned i = 0; i < NumElems; ++i) {
6918 SDValue Elt = Op.getOperand(i);
6919 if (Elt.getOpcode() == ISD::UNDEF)
6922 if (Elt.getOpcode() != ISD::Constant &&
6923 Elt.getOpcode() != ISD::ConstantFP)
6924 IsAllConstants = false;
6925 if (X86::isZeroNode(Elt))
6928 NonZeros |= (1 << i);
6933 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6934 if (NumNonZero == 0)
6935 return DAG.getUNDEF(VT);
6937 // Special case for single non-zero, non-undef, element.
6938 if (NumNonZero == 1) {
6939 unsigned Idx = countTrailingZeros(NonZeros);
6940 SDValue Item = Op.getOperand(Idx);
6942 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6943 // the value are obviously zero, truncate the value to i32 and do the
6944 // insertion that way. Only do this if the value is non-constant or if the
6945 // value is a constant being inserted into element 0. It is cheaper to do
6946 // a constant pool load than it is to do a movd + shuffle.
6947 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6948 (!IsAllConstants || Idx == 0)) {
6949 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6951 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6952 EVT VecVT = MVT::v4i32;
6953 unsigned VecElts = 4;
6955 // Truncate the value (which may itself be a constant) to i32, and
6956 // convert it to a vector with movd (S2V+shuffle to zero extend).
6957 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6958 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6960 // If using the new shuffle lowering, just directly insert this.
6961 if (ExperimentalVectorShuffleLowering)
6963 ISD::BITCAST, dl, VT,
6964 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
6966 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6968 // Now we have our 32-bit value zero extended in the low element of
6969 // a vector. If Idx != 0, swizzle it into place.
6971 SmallVector<int, 4> Mask;
6972 Mask.push_back(Idx);
6973 for (unsigned i = 1; i != VecElts; ++i)
6975 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
6978 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6982 // If we have a constant or non-constant insertion into the low element of
6983 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6984 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6985 // depending on what the source datatype is.
6988 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6990 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6991 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6992 if (VT.is256BitVector() || VT.is512BitVector()) {
6993 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6994 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6995 Item, DAG.getIntPtrConstant(0));
6997 assert(VT.is128BitVector() && "Expected an SSE value type!");
6998 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6999 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
7000 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
7003 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
7004 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
7005 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
7006 if (VT.is256BitVector()) {
7007 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
7008 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
7010 assert(VT.is128BitVector() && "Expected an SSE value type!");
7011 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
7013 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
7017 // Is it a vector logical left shift?
7018 if (NumElems == 2 && Idx == 1 &&
7019 X86::isZeroNode(Op.getOperand(0)) &&
7020 !X86::isZeroNode(Op.getOperand(1))) {
7021 unsigned NumBits = VT.getSizeInBits();
7022 return getVShift(true, VT,
7023 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7024 VT, Op.getOperand(1)),
7025 NumBits/2, DAG, *this, dl);
7028 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
7031 // Otherwise, if this is a vector with i32 or f32 elements, and the element
7032 // is a non-constant being inserted into an element other than the low one,
7033 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
7034 // movd/movss) to move this into the low element, then shuffle it into
7036 if (EVTBits == 32) {
7037 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
7039 // If using the new shuffle lowering, just directly insert this.
7040 if (ExperimentalVectorShuffleLowering)
7041 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
7043 // Turn it into a shuffle of zero and zero-extended scalar to vector.
7044 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
7045 SmallVector<int, 8> MaskVec;
7046 for (unsigned i = 0; i != NumElems; ++i)
7047 MaskVec.push_back(i == Idx ? 0 : 1);
7048 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
7052 // Splat is obviously ok. Let legalizer expand it to a shuffle.
7053 if (Values.size() == 1) {
7054 if (EVTBits == 32) {
7055 // Instead of a shuffle like this:
7056 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
7057 // Check if it's possible to issue this instead.
7058 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
7059 unsigned Idx = countTrailingZeros(NonZeros);
7060 SDValue Item = Op.getOperand(Idx);
7061 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
7062 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
7067 // A vector full of immediates; various special cases are already
7068 // handled, so this is best done with a single constant-pool load.
7072 // For AVX-length vectors, see if we can use a vector load to get all of the
7073 // elements, otherwise build the individual 128-bit pieces and use
7074 // shuffles to put them in place.
7075 if (VT.is256BitVector() || VT.is512BitVector()) {
7076 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
7078 // Check for a build vector of consecutive loads.
7079 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
7082 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
7084 // Build both the lower and upper subvector.
7085 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7086 makeArrayRef(&V[0], NumElems/2));
7087 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7088 makeArrayRef(&V[NumElems / 2], NumElems/2));
7090 // Recreate the wider vector with the lower and upper part.
7091 if (VT.is256BitVector())
7092 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7093 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7096 // Let legalizer expand 2-wide build_vectors.
7097 if (EVTBits == 64) {
7098 if (NumNonZero == 1) {
7099 // One half is zero or undef.
7100 unsigned Idx = countTrailingZeros(NonZeros);
7101 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
7102 Op.getOperand(Idx));
7103 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
7108 // If element VT is < 32 bits, convert it to inserts into a zero vector.
7109 if (EVTBits == 8 && NumElems == 16) {
7110 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
7112 if (V.getNode()) return V;
7115 if (EVTBits == 16 && NumElems == 8) {
7116 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
7118 if (V.getNode()) return V;
7121 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
7122 if (EVTBits == 32 && NumElems == 4) {
7123 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
7128 // If element VT is == 32 bits, turn it into a number of shuffles.
7129 SmallVector<SDValue, 8> V(NumElems);
7130 if (NumElems == 4 && NumZero > 0) {
7131 for (unsigned i = 0; i < 4; ++i) {
7132 bool isZero = !(NonZeros & (1 << i));
7134 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
7136 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7139 for (unsigned i = 0; i < 2; ++i) {
7140 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
7143 V[i] = V[i*2]; // Must be a zero vector.
7146 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
7149 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
7152 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
7157 bool Reverse1 = (NonZeros & 0x3) == 2;
7158 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
7162 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
7163 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
7165 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
7168 if (Values.size() > 1 && VT.is128BitVector()) {
7169 // Check for a build vector of consecutive loads.
7170 for (unsigned i = 0; i < NumElems; ++i)
7171 V[i] = Op.getOperand(i);
7173 // Check for elements which are consecutive loads.
7174 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
7178 // Check for a build vector from mostly shuffle plus few inserting.
7179 SDValue Sh = buildFromShuffleMostly(Op, DAG);
7183 // For SSE 4.1, use insertps to put the high elements into the low element.
7184 if (Subtarget->hasSSE41()) {
7186 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
7187 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
7189 Result = DAG.getUNDEF(VT);
7191 for (unsigned i = 1; i < NumElems; ++i) {
7192 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
7193 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
7194 Op.getOperand(i), DAG.getIntPtrConstant(i));
7199 // Otherwise, expand into a number of unpckl*, start by extending each of
7200 // our (non-undef) elements to the full vector width with the element in the
7201 // bottom slot of the vector (which generates no code for SSE).
7202 for (unsigned i = 0; i < NumElems; ++i) {
7203 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
7204 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7206 V[i] = DAG.getUNDEF(VT);
7209 // Next, we iteratively mix elements, e.g. for v4f32:
7210 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
7211 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
7212 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
7213 unsigned EltStride = NumElems >> 1;
7214 while (EltStride != 0) {
7215 for (unsigned i = 0; i < EltStride; ++i) {
7216 // If V[i+EltStride] is undef and this is the first round of mixing,
7217 // then it is safe to just drop this shuffle: V[i] is already in the
7218 // right place, the one element (since it's the first round) being
7219 // inserted as undef can be dropped. This isn't safe for successive
7220 // rounds because they will permute elements within both vectors.
7221 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
7222 EltStride == NumElems/2)
7225 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
7234 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
7235 // to create 256-bit vectors from two other 128-bit ones.
7236 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7238 MVT ResVT = Op.getSimpleValueType();
7240 assert((ResVT.is256BitVector() ||
7241 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
7243 SDValue V1 = Op.getOperand(0);
7244 SDValue V2 = Op.getOperand(1);
7245 unsigned NumElems = ResVT.getVectorNumElements();
7246 if(ResVT.is256BitVector())
7247 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7249 if (Op.getNumOperands() == 4) {
7250 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
7251 ResVT.getVectorNumElements()/2);
7252 SDValue V3 = Op.getOperand(2);
7253 SDValue V4 = Op.getOperand(3);
7254 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
7255 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
7257 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7260 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7261 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
7262 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
7263 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
7264 Op.getNumOperands() == 4)));
7266 // AVX can use the vinsertf128 instruction to create 256-bit vectors
7267 // from two other 128-bit ones.
7269 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
7270 return LowerAVXCONCAT_VECTORS(Op, DAG);
7274 //===----------------------------------------------------------------------===//
7275 // Vector shuffle lowering
7277 // This is an experimental code path for lowering vector shuffles on x86. It is
7278 // designed to handle arbitrary vector shuffles and blends, gracefully
7279 // degrading performance as necessary. It works hard to recognize idiomatic
7280 // shuffles and lower them to optimal instruction patterns without leaving
7281 // a framework that allows reasonably efficient handling of all vector shuffle
7283 //===----------------------------------------------------------------------===//
7285 /// \brief Tiny helper function to identify a no-op mask.
7287 /// This is a somewhat boring predicate function. It checks whether the mask
7288 /// array input, which is assumed to be a single-input shuffle mask of the kind
7289 /// used by the X86 shuffle instructions (not a fully general
7290 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
7291 /// in-place shuffle are 'no-op's.
7292 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
7293 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7294 if (Mask[i] != -1 && Mask[i] != i)
7299 /// \brief Helper function to classify a mask as a single-input mask.
7301 /// This isn't a generic single-input test because in the vector shuffle
7302 /// lowering we canonicalize single inputs to be the first input operand. This
7303 /// means we can more quickly test for a single input by only checking whether
7304 /// an input from the second operand exists. We also assume that the size of
7305 /// mask corresponds to the size of the input vectors which isn't true in the
7306 /// fully general case.
7307 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
7309 if (M >= (int)Mask.size())
7314 /// \brief Test whether there are elements crossing 128-bit lanes in this
7317 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
7318 /// and we routinely test for these.
7319 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
7320 int LaneSize = 128 / VT.getScalarSizeInBits();
7321 int Size = Mask.size();
7322 for (int i = 0; i < Size; ++i)
7323 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
7328 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
7330 /// This checks a shuffle mask to see if it is performing the same
7331 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
7332 /// that it is also not lane-crossing. It may however involve a blend from the
7333 /// same lane of a second vector.
7335 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
7336 /// non-trivial to compute in the face of undef lanes. The representation is
7337 /// *not* suitable for use with existing 128-bit shuffles as it will contain
7338 /// entries from both V1 and V2 inputs to the wider mask.
7340 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
7341 SmallVectorImpl<int> &RepeatedMask) {
7342 int LaneSize = 128 / VT.getScalarSizeInBits();
7343 RepeatedMask.resize(LaneSize, -1);
7344 int Size = Mask.size();
7345 for (int i = 0; i < Size; ++i) {
7348 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
7349 // This entry crosses lanes, so there is no way to model this shuffle.
7352 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
7353 if (RepeatedMask[i % LaneSize] == -1)
7354 // This is the first non-undef entry in this slot of a 128-bit lane.
7355 RepeatedMask[i % LaneSize] =
7356 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
7357 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
7358 // Found a mismatch with the repeated mask.
7364 /// \brief Base case helper for testing a single mask element.
7365 static bool isShuffleEquivalentImpl(SDValue V1, SDValue V2,
7366 BuildVectorSDNode *BV1,
7367 BuildVectorSDNode *BV2, ArrayRef<int> Mask,
7369 int Size = Mask.size();
7370 if (Mask[i] != -1 && Mask[i] != Arg) {
7371 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
7372 auto *ArgsBV = Arg < Size ? BV1 : BV2;
7373 if (!MaskBV || !ArgsBV ||
7374 MaskBV->getOperand(Mask[i] % Size) != ArgsBV->getOperand(Arg % Size))
7380 /// \brief Recursive helper to peel off and test each mask element.
7381 template <typename... Ts>
7382 static bool isShuffleEquivalentImpl(SDValue V1, SDValue V2,
7383 BuildVectorSDNode *BV1,
7384 BuildVectorSDNode *BV2, ArrayRef<int> Mask,
7385 int i, int Arg, Ts... Args) {
7386 if (!isShuffleEquivalentImpl(V1, V2, BV1, BV2, Mask, i, Arg))
7389 return isShuffleEquivalentImpl(V1, V2, BV1, BV2, Mask, i + 1, Args...);
7392 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
7395 /// This is a fast way to test a shuffle mask against a fixed pattern:
7397 /// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... }
7399 /// It returns true if the mask is exactly as wide as the argument list, and
7400 /// each element of the mask is either -1 (signifying undef) or the value given
7401 /// in the argument.
7402 template <typename... Ts>
7403 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
7405 if (Mask.size() != sizeof...(Args))
7408 // If the values are build vectors, we can look through them to find
7409 // equivalent inputs that make the shuffles equivalent.
7410 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
7411 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
7413 // Recursively peel off arguments and test them against the mask.
7414 return isShuffleEquivalentImpl(V1, V2, BV1, BV2, Mask, 0, Args...);
7417 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
7419 /// This helper function produces an 8-bit shuffle immediate corresponding to
7420 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
7421 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
7424 /// NB: We rely heavily on "undef" masks preserving the input lane.
7425 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
7426 SelectionDAG &DAG) {
7427 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
7428 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
7429 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
7430 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
7431 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
7434 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
7435 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
7436 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
7437 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
7438 return DAG.getConstant(Imm, MVT::i8);
7441 /// \brief Try to emit a blend instruction for a shuffle using bit math.
7443 /// This is used as a fallback approach when first class blend instructions are
7444 /// unavailable. Currently it is only suitable for integer vectors, but could
7445 /// be generalized for floating point vectors if desirable.
7446 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
7447 SDValue V2, ArrayRef<int> Mask,
7448 SelectionDAG &DAG) {
7449 assert(VT.isInteger() && "Only supports integer vector types!");
7450 MVT EltVT = VT.getScalarType();
7451 int NumEltBits = EltVT.getSizeInBits();
7452 SDValue Zero = DAG.getConstant(0, EltVT);
7453 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
7454 SmallVector<SDValue, 16> MaskOps;
7455 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7456 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
7457 return SDValue(); // Shuffled input!
7458 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
7461 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
7462 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
7463 // We have to cast V2 around.
7464 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
7465 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
7466 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
7467 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
7468 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
7469 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
7472 /// \brief Try to emit a blend instruction for a shuffle.
7474 /// This doesn't do any checks for the availability of instructions for blending
7475 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7476 /// be matched in the backend with the type given. What it does check for is
7477 /// that the shuffle mask is in fact a blend.
7478 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7479 SDValue V2, ArrayRef<int> Mask,
7480 const X86Subtarget *Subtarget,
7481 SelectionDAG &DAG) {
7482 unsigned BlendMask = 0;
7483 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7484 if (Mask[i] >= Size) {
7485 if (Mask[i] != i + Size)
7486 return SDValue(); // Shuffled V2 input!
7487 BlendMask |= 1u << i;
7490 if (Mask[i] >= 0 && Mask[i] != i)
7491 return SDValue(); // Shuffled V1 input!
7493 switch (VT.SimpleTy) {
7498 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7499 DAG.getConstant(BlendMask, MVT::i8));
7503 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7507 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7508 // that instruction.
7509 if (Subtarget->hasAVX2()) {
7510 // Scale the blend by the number of 32-bit dwords per element.
7511 int Scale = VT.getScalarSizeInBits() / 32;
7513 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7514 if (Mask[i] >= Size)
7515 for (int j = 0; j < Scale; ++j)
7516 BlendMask |= 1u << (i * Scale + j);
7518 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7519 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
7520 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
7521 return DAG.getNode(ISD::BITCAST, DL, VT,
7522 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7523 DAG.getConstant(BlendMask, MVT::i8)));
7527 // For integer shuffles we need to expand the mask and cast the inputs to
7528 // v8i16s prior to blending.
7529 int Scale = 8 / VT.getVectorNumElements();
7531 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7532 if (Mask[i] >= Size)
7533 for (int j = 0; j < Scale; ++j)
7534 BlendMask |= 1u << (i * Scale + j);
7536 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
7537 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
7538 return DAG.getNode(ISD::BITCAST, DL, VT,
7539 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7540 DAG.getConstant(BlendMask, MVT::i8)));
7544 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7545 SmallVector<int, 8> RepeatedMask;
7546 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7547 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7548 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7550 for (int i = 0; i < 8; ++i)
7551 if (RepeatedMask[i] >= 16)
7552 BlendMask |= 1u << i;
7553 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7554 DAG.getConstant(BlendMask, MVT::i8));
7560 // Scale the blend by the number of bytes per element.
7561 int Scale = VT.getScalarSizeInBits() / 8;
7563 // This form of blend is always done on bytes. Compute the byte vector
7565 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7567 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7568 // mix of LLVM's code generator and the x86 backend. We tell the code
7569 // generator that boolean values in the elements of an x86 vector register
7570 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7571 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7572 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7573 // of the element (the remaining are ignored) and 0 in that high bit would
7574 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7575 // the LLVM model for boolean values in vector elements gets the relevant
7576 // bit set, it is set backwards and over constrained relative to x86's
7578 SmallVector<SDValue, 32> VSELECTMask;
7579 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7580 for (int j = 0; j < Scale; ++j)
7581 VSELECTMask.push_back(
7582 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7583 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
7585 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
7586 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
7588 ISD::BITCAST, DL, VT,
7589 DAG.getNode(ISD::VSELECT, DL, BlendVT,
7590 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
7595 llvm_unreachable("Not a supported integer vector type!");
7599 /// \brief Try to lower as a blend of elements from two inputs followed by
7600 /// a single-input permutation.
7602 /// This matches the pattern where we can blend elements from two inputs and
7603 /// then reduce the shuffle to a single-input permutation.
7604 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7607 SelectionDAG &DAG) {
7608 // We build up the blend mask while checking whether a blend is a viable way
7609 // to reduce the shuffle.
7610 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7611 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7613 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7617 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7619 if (BlendMask[Mask[i] % Size] == -1)
7620 BlendMask[Mask[i] % Size] = Mask[i];
7621 else if (BlendMask[Mask[i] % Size] != Mask[i])
7622 return SDValue(); // Can't blend in the needed input!
7624 PermuteMask[i] = Mask[i] % Size;
7627 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7628 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7631 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7632 /// blends and permutes.
7634 /// This matches the extremely common pattern for handling combined
7635 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7636 /// operations. It will try to pick the best arrangement of shuffles and
7638 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7642 SelectionDAG &DAG) {
7643 // Shuffle the input elements into the desired positions in V1 and V2 and
7644 // blend them together.
7645 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7646 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7647 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7648 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7649 if (Mask[i] >= 0 && Mask[i] < Size) {
7650 V1Mask[i] = Mask[i];
7652 } else if (Mask[i] >= Size) {
7653 V2Mask[i] = Mask[i] - Size;
7654 BlendMask[i] = i + Size;
7657 // Try to lower with the simpler initial blend strategy unless one of the
7658 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7659 // shuffle may be able to fold with a load or other benefit. However, when
7660 // we'll have to do 2x as many shuffles in order to achieve this, blending
7661 // first is a better strategy.
7662 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7663 if (SDValue BlendPerm =
7664 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7667 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7668 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7669 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7672 /// \brief Try to lower a vector shuffle as a byte rotation.
7674 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7675 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7676 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7677 /// try to generically lower a vector shuffle through such an pattern. It
7678 /// does not check for the profitability of lowering either as PALIGNR or
7679 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7680 /// This matches shuffle vectors that look like:
7682 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7684 /// Essentially it concatenates V1 and V2, shifts right by some number of
7685 /// elements, and takes the low elements as the result. Note that while this is
7686 /// specified as a *right shift* because x86 is little-endian, it is a *left
7687 /// rotate* of the vector lanes.
7688 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7691 const X86Subtarget *Subtarget,
7692 SelectionDAG &DAG) {
7693 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7695 int NumElts = Mask.size();
7696 int NumLanes = VT.getSizeInBits() / 128;
7697 int NumLaneElts = NumElts / NumLanes;
7699 // We need to detect various ways of spelling a rotation:
7700 // [11, 12, 13, 14, 15, 0, 1, 2]
7701 // [-1, 12, 13, 14, -1, -1, 1, -1]
7702 // [-1, -1, -1, -1, -1, -1, 1, 2]
7703 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7704 // [-1, 4, 5, 6, -1, -1, 9, -1]
7705 // [-1, 4, 5, 6, -1, -1, -1, -1]
7708 for (int l = 0; l < NumElts; l += NumLaneElts) {
7709 for (int i = 0; i < NumLaneElts; ++i) {
7710 if (Mask[l + i] == -1)
7712 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7714 // Get the mod-Size index and lane correct it.
7715 int LaneIdx = (Mask[l + i] % NumElts) - l;
7716 // Make sure it was in this lane.
7717 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7720 // Determine where a rotated vector would have started.
7721 int StartIdx = i - LaneIdx;
7723 // The identity rotation isn't interesting, stop.
7726 // If we found the tail of a vector the rotation must be the missing
7727 // front. If we found the head of a vector, it must be how much of the
7729 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7732 Rotation = CandidateRotation;
7733 else if (Rotation != CandidateRotation)
7734 // The rotations don't match, so we can't match this mask.
7737 // Compute which value this mask is pointing at.
7738 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7740 // Compute which of the two target values this index should be assigned
7741 // to. This reflects whether the high elements are remaining or the low
7742 // elements are remaining.
7743 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7745 // Either set up this value if we've not encountered it before, or check
7746 // that it remains consistent.
7749 else if (TargetV != MaskV)
7750 // This may be a rotation, but it pulls from the inputs in some
7751 // unsupported interleaving.
7756 // Check that we successfully analyzed the mask, and normalize the results.
7757 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7758 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7764 // The actual rotate instruction rotates bytes, so we need to scale the
7765 // rotation based on how many bytes are in the vector lane.
7766 int Scale = 16 / NumLaneElts;
7768 // SSSE3 targets can use the palignr instruction.
7769 if (Subtarget->hasSSSE3()) {
7770 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7771 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7772 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
7773 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
7775 return DAG.getNode(ISD::BITCAST, DL, VT,
7776 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
7777 DAG.getConstant(Rotation * Scale, MVT::i8)));
7780 assert(VT.getSizeInBits() == 128 &&
7781 "Rotate-based lowering only supports 128-bit lowering!");
7782 assert(Mask.size() <= 16 &&
7783 "Can shuffle at most 16 bytes in a 128-bit vector!");
7785 // Default SSE2 implementation
7786 int LoByteShift = 16 - Rotation * Scale;
7787 int HiByteShift = Rotation * Scale;
7789 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7790 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
7791 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
7793 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7794 DAG.getConstant(LoByteShift, MVT::i8));
7795 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7796 DAG.getConstant(HiByteShift, MVT::i8));
7797 return DAG.getNode(ISD::BITCAST, DL, VT,
7798 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7801 /// \brief Compute whether each element of a shuffle is zeroable.
7803 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
7804 /// Either it is an undef element in the shuffle mask, the element of the input
7805 /// referenced is undef, or the element of the input referenced is known to be
7806 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
7807 /// as many lanes with this technique as possible to simplify the remaining
7809 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
7810 SDValue V1, SDValue V2) {
7811 SmallBitVector Zeroable(Mask.size(), false);
7813 while (V1.getOpcode() == ISD::BITCAST)
7814 V1 = V1->getOperand(0);
7815 while (V2.getOpcode() == ISD::BITCAST)
7816 V2 = V2->getOperand(0);
7818 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7819 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7821 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7823 // Handle the easy cases.
7824 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
7829 // If this is an index into a build_vector node (which has the same number
7830 // of elements), dig out the input value and use it.
7831 SDValue V = M < Size ? V1 : V2;
7832 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
7835 SDValue Input = V.getOperand(M % Size);
7836 // The UNDEF opcode check really should be dead code here, but not quite
7837 // worth asserting on (it isn't invalid, just unexpected).
7838 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
7845 /// \brief Try to emit a bitmask instruction for a shuffle.
7847 /// This handles cases where we can model a blend exactly as a bitmask due to
7848 /// one of the inputs being zeroable.
7849 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
7850 SDValue V2, ArrayRef<int> Mask,
7851 SelectionDAG &DAG) {
7852 MVT EltVT = VT.getScalarType();
7853 int NumEltBits = EltVT.getSizeInBits();
7854 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
7855 SDValue Zero = DAG.getConstant(0, IntEltVT);
7856 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
7857 if (EltVT.isFloatingPoint()) {
7858 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
7859 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
7861 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
7862 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7864 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7867 if (Mask[i] % Size != i)
7868 return SDValue(); // Not a blend.
7870 V = Mask[i] < Size ? V1 : V2;
7871 else if (V != (Mask[i] < Size ? V1 : V2))
7872 return SDValue(); // Can only let one input through the mask.
7874 VMaskOps[i] = AllOnes;
7877 return SDValue(); // No non-zeroable elements!
7879 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
7880 V = DAG.getNode(VT.isFloatingPoint()
7881 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
7886 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7888 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7889 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7890 /// matches elements from one of the input vectors shuffled to the left or
7891 /// right with zeroable elements 'shifted in'. It handles both the strictly
7892 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7895 /// PSHL : (little-endian) left bit shift.
7896 /// [ zz, 0, zz, 2 ]
7897 /// [ -1, 4, zz, -1 ]
7898 /// PSRL : (little-endian) right bit shift.
7900 /// [ -1, -1, 7, zz]
7901 /// PSLLDQ : (little-endian) left byte shift
7902 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7903 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7904 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7905 /// PSRLDQ : (little-endian) right byte shift
7906 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7907 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7908 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7909 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7910 SDValue V2, ArrayRef<int> Mask,
7911 SelectionDAG &DAG) {
7912 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7914 int Size = Mask.size();
7915 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7917 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7918 for (int i = 0; i < Size; i += Scale)
7919 for (int j = 0; j < Shift; ++j)
7920 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7926 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7927 for (int i = 0; i != Size; i += Scale) {
7928 unsigned Pos = Left ? i + Shift : i;
7929 unsigned Low = Left ? i : i + Shift;
7930 unsigned Len = Scale - Shift;
7931 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7932 Low + (V == V1 ? 0 : Size)))
7936 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7937 bool ByteShift = ShiftEltBits > 64;
7938 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7939 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7940 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7942 // Normalize the scale for byte shifts to still produce an i64 element
7944 Scale = ByteShift ? Scale / 2 : Scale;
7946 // We need to round trip through the appropriate type for the shift.
7947 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7948 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7949 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7950 "Illegal integer vector type");
7951 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
7953 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
7954 return DAG.getNode(ISD::BITCAST, DL, VT, V);
7957 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7958 // keep doubling the size of the integer elements up to that. We can
7959 // then shift the elements of the integer vector by whole multiples of
7960 // their width within the elements of the larger integer vector. Test each
7961 // multiple to see if we can find a match with the moved element indices
7962 // and that the shifted in elements are all zeroable.
7963 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7964 for (int Shift = 1; Shift != Scale; ++Shift)
7965 for (bool Left : {true, false})
7966 if (CheckZeros(Shift, Scale, Left))
7967 for (SDValue V : {V1, V2})
7968 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7975 /// \brief Lower a vector shuffle as a zero or any extension.
7977 /// Given a specific number of elements, element bit width, and extension
7978 /// stride, produce either a zero or any extension based on the available
7979 /// features of the subtarget.
7980 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7981 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7982 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7983 assert(Scale > 1 && "Need a scale to extend.");
7984 int NumElements = VT.getVectorNumElements();
7985 int EltBits = VT.getScalarSizeInBits();
7986 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7987 "Only 8, 16, and 32 bit elements can be extended.");
7988 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7990 // Found a valid zext mask! Try various lowering strategies based on the
7991 // input type and available ISA extensions.
7992 if (Subtarget->hasSSE41()) {
7993 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7994 NumElements / Scale);
7995 return DAG.getNode(ISD::BITCAST, DL, VT,
7996 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7999 // For any extends we can cheat for larger element sizes and use shuffle
8000 // instructions that can fold with a load and/or copy.
8001 if (AnyExt && EltBits == 32) {
8002 int PSHUFDMask[4] = {0, -1, 1, -1};
8004 ISD::BITCAST, DL, VT,
8005 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8006 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
8007 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8009 if (AnyExt && EltBits == 16 && Scale > 2) {
8010 int PSHUFDMask[4] = {0, -1, 0, -1};
8011 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8012 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
8013 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
8014 int PSHUFHWMask[4] = {1, -1, -1, -1};
8016 ISD::BITCAST, DL, VT,
8017 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
8018 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
8019 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
8022 // If this would require more than 2 unpack instructions to expand, use
8023 // pshufb when available. We can only use more than 2 unpack instructions
8024 // when zero extending i8 elements which also makes it easier to use pshufb.
8025 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
8026 assert(NumElements == 16 && "Unexpected byte vector width!");
8027 SDValue PSHUFBMask[16];
8028 for (int i = 0; i < 16; ++i)
8030 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
8031 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
8032 return DAG.getNode(ISD::BITCAST, DL, VT,
8033 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
8034 DAG.getNode(ISD::BUILD_VECTOR, DL,
8035 MVT::v16i8, PSHUFBMask)));
8038 // Otherwise emit a sequence of unpacks.
8040 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
8041 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
8042 : getZeroVector(InputVT, Subtarget, DAG, DL);
8043 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
8044 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
8048 } while (Scale > 1);
8049 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
8052 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
8054 /// This routine will try to do everything in its power to cleverly lower
8055 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
8056 /// check for the profitability of this lowering, it tries to aggressively
8057 /// match this pattern. It will use all of the micro-architectural details it
8058 /// can to emit an efficient lowering. It handles both blends with all-zero
8059 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
8060 /// masking out later).
8062 /// The reason we have dedicated lowering for zext-style shuffles is that they
8063 /// are both incredibly common and often quite performance sensitive.
8064 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
8065 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
8066 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8067 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8069 int Bits = VT.getSizeInBits();
8070 int NumElements = VT.getVectorNumElements();
8071 assert(VT.getScalarSizeInBits() <= 32 &&
8072 "Exceeds 32-bit integer zero extension limit");
8073 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
8075 // Define a helper function to check a particular ext-scale and lower to it if
8077 auto Lower = [&](int Scale) -> SDValue {
8080 for (int i = 0; i < NumElements; ++i) {
8082 continue; // Valid anywhere but doesn't tell us anything.
8083 if (i % Scale != 0) {
8084 // Each of the extended elements need to be zeroable.
8088 // We no longer are in the anyext case.
8093 // Each of the base elements needs to be consecutive indices into the
8094 // same input vector.
8095 SDValue V = Mask[i] < NumElements ? V1 : V2;
8098 else if (InputV != V)
8099 return SDValue(); // Flip-flopping inputs.
8101 if (Mask[i] % NumElements != i / Scale)
8102 return SDValue(); // Non-consecutive strided elements.
8105 // If we fail to find an input, we have a zero-shuffle which should always
8106 // have already been handled.
8107 // FIXME: Maybe handle this here in case during blending we end up with one?
8111 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
8112 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
8115 // The widest scale possible for extending is to a 64-bit integer.
8116 assert(Bits % 64 == 0 &&
8117 "The number of bits in a vector must be divisible by 64 on x86!");
8118 int NumExtElements = Bits / 64;
8120 // Each iteration, try extending the elements half as much, but into twice as
8122 for (; NumExtElements < NumElements; NumExtElements *= 2) {
8123 assert(NumElements % NumExtElements == 0 &&
8124 "The input vector size must be divisible by the extended size.");
8125 if (SDValue V = Lower(NumElements / NumExtElements))
8129 // General extends failed, but 128-bit vectors may be able to use MOVQ.
8133 // Returns one of the source operands if the shuffle can be reduced to a
8134 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
8135 auto CanZExtLowHalf = [&]() {
8136 for (int i = NumElements / 2; i != NumElements; ++i)
8139 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
8141 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
8146 if (SDValue V = CanZExtLowHalf()) {
8147 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
8148 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
8149 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8152 // No viable ext lowering found.
8156 /// \brief Try to get a scalar value for a specific element of a vector.
8158 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
8159 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
8160 SelectionDAG &DAG) {
8161 MVT VT = V.getSimpleValueType();
8162 MVT EltVT = VT.getVectorElementType();
8163 while (V.getOpcode() == ISD::BITCAST)
8164 V = V.getOperand(0);
8165 // If the bitcasts shift the element size, we can't extract an equivalent
8167 MVT NewVT = V.getSimpleValueType();
8168 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
8171 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8172 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
8173 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
8178 /// \brief Helper to test for a load that can be folded with x86 shuffles.
8180 /// This is particularly important because the set of instructions varies
8181 /// significantly based on whether the operand is a load or not.
8182 static bool isShuffleFoldableLoad(SDValue V) {
8183 while (V.getOpcode() == ISD::BITCAST)
8184 V = V.getOperand(0);
8186 return ISD::isNON_EXTLoad(V.getNode());
8189 /// \brief Try to lower insertion of a single element into a zero vector.
8191 /// This is a common pattern that we have especially efficient patterns to lower
8192 /// across all subtarget feature sets.
8193 static SDValue lowerVectorShuffleAsElementInsertion(
8194 MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
8195 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8196 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8198 MVT EltVT = VT.getVectorElementType();
8200 int V2Index = std::find_if(Mask.begin(), Mask.end(),
8201 [&Mask](int M) { return M >= (int)Mask.size(); }) -
8203 bool IsV1Zeroable = true;
8204 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8205 if (i != V2Index && !Zeroable[i]) {
8206 IsV1Zeroable = false;
8210 // Check for a single input from a SCALAR_TO_VECTOR node.
8211 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
8212 // all the smarts here sunk into that routine. However, the current
8213 // lowering of BUILD_VECTOR makes that nearly impossible until the old
8214 // vector shuffle lowering is dead.
8215 if (SDValue V2S = getScalarValueForVectorElement(
8216 V2, Mask[V2Index] - Mask.size(), DAG)) {
8217 // We need to zext the scalar if it is smaller than an i32.
8218 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
8219 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
8220 // Using zext to expand a narrow element won't work for non-zero
8225 // Zero-extend directly to i32.
8227 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
8229 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
8230 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
8231 EltVT == MVT::i16) {
8232 // Either not inserting from the low element of the input or the input
8233 // element size is too small to use VZEXT_MOVL to clear the high bits.
8237 if (!IsV1Zeroable) {
8238 // If V1 can't be treated as a zero vector we have fewer options to lower
8239 // this. We can't support integer vectors or non-zero targets cheaply, and
8240 // the V1 elements can't be permuted in any way.
8241 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8242 if (!VT.isFloatingPoint() || V2Index != 0)
8244 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8245 V1Mask[V2Index] = -1;
8246 if (!isNoopShuffleMask(V1Mask))
8248 // This is essentially a special case blend operation, but if we have
8249 // general purpose blend operations, they are always faster. Bail and let
8250 // the rest of the lowering handle these as blends.
8251 if (Subtarget->hasSSE41())
8254 // Otherwise, use MOVSD or MOVSS.
8255 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8256 "Only two types of floating point element types to handle!");
8257 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8261 // This lowering only works for the low element with floating point vectors.
8262 if (VT.isFloatingPoint() && V2Index != 0)
8265 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8267 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8270 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8271 // the desired position. Otherwise it is more efficient to do a vector
8272 // shift left. We know that we can do a vector shift left because all
8273 // the inputs are zero.
8274 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8275 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8276 V2Shuffle[V2Index] = 0;
8277 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8279 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
8281 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8283 V2Index * EltVT.getSizeInBits()/8,
8284 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
8285 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8291 /// \brief Try to lower broadcast of a single element.
8293 /// For convenience, this code also bundles all of the subtarget feature set
8294 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8295 /// a convenient way to factor it out.
8296 static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V,
8298 const X86Subtarget *Subtarget,
8299 SelectionDAG &DAG) {
8300 if (!Subtarget->hasAVX())
8302 if (VT.isInteger() && !Subtarget->hasAVX2())
8305 // Check that the mask is a broadcast.
8306 int BroadcastIdx = -1;
8308 if (M >= 0 && BroadcastIdx == -1)
8310 else if (M >= 0 && M != BroadcastIdx)
8313 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8314 "a sorted mask where the broadcast "
8317 // Go up the chain of (vector) values to try and find a scalar load that
8318 // we can combine with the broadcast.
8320 switch (V.getOpcode()) {
8321 case ISD::CONCAT_VECTORS: {
8322 int OperandSize = Mask.size() / V.getNumOperands();
8323 V = V.getOperand(BroadcastIdx / OperandSize);
8324 BroadcastIdx %= OperandSize;
8328 case ISD::INSERT_SUBVECTOR: {
8329 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8330 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8334 int BeginIdx = (int)ConstantIdx->getZExtValue();
8336 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
8337 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8338 BroadcastIdx -= BeginIdx;
8349 // Check if this is a broadcast of a scalar. We special case lowering
8350 // for scalars so that we can more effectively fold with loads.
8351 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8352 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8353 V = V.getOperand(BroadcastIdx);
8355 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
8357 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8359 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8360 // We can't broadcast from a vector register w/o AVX2, and we can only
8361 // broadcast from the zero-element of a vector register.
8365 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8368 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8369 // INSERTPS when the V1 elements are already in the correct locations
8370 // because otherwise we can just always use two SHUFPS instructions which
8371 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8372 // perform INSERTPS if a single V1 element is out of place and all V2
8373 // elements are zeroable.
8374 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8376 SelectionDAG &DAG) {
8377 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8378 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8379 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8380 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8382 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8385 int V1DstIndex = -1;
8386 int V2DstIndex = -1;
8387 bool V1UsedInPlace = false;
8389 for (int i = 0; i < 4; ++i) {
8390 // Synthesize a zero mask from the zeroable elements (includes undefs).
8396 // Flag if we use any V1 inputs in place.
8398 V1UsedInPlace = true;
8402 // We can only insert a single non-zeroable element.
8403 if (V1DstIndex != -1 || V2DstIndex != -1)
8407 // V1 input out of place for insertion.
8410 // V2 input for insertion.
8415 // Don't bother if we have no (non-zeroable) element for insertion.
8416 if (V1DstIndex == -1 && V2DstIndex == -1)
8419 // Determine element insertion src/dst indices. The src index is from the
8420 // start of the inserted vector, not the start of the concatenated vector.
8421 unsigned V2SrcIndex = 0;
8422 if (V1DstIndex != -1) {
8423 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8424 // and don't use the original V2 at all.
8425 V2SrcIndex = Mask[V1DstIndex];
8426 V2DstIndex = V1DstIndex;
8429 V2SrcIndex = Mask[V2DstIndex] - 4;
8432 // If no V1 inputs are used in place, then the result is created only from
8433 // the zero mask and the V2 insertion - so remove V1 dependency.
8435 V1 = DAG.getUNDEF(MVT::v4f32);
8437 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8438 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8440 // Insert the V2 element into the desired position.
8442 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8443 DAG.getConstant(InsertPSMask, MVT::i8));
8446 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8447 /// UNPCK instruction.
8449 /// This specifically targets cases where we end up with alternating between
8450 /// the two inputs, and so can permute them into something that feeds a single
8451 /// UNPCK instruction. Note that this routine only targets integer vectors
8452 /// because for floating point vectors we have a generalized SHUFPS lowering
8453 /// strategy that handles everything that doesn't *exactly* match an unpack,
8454 /// making this clever lowering unnecessary.
8455 static SDValue lowerVectorShuffleAsUnpack(MVT VT, SDLoc DL, SDValue V1,
8456 SDValue V2, ArrayRef<int> Mask,
8457 SelectionDAG &DAG) {
8458 assert(!VT.isFloatingPoint() &&
8459 "This routine only supports integer vectors.");
8460 assert(!isSingleInputShuffleMask(Mask) &&
8461 "This routine should only be used when blending two inputs.");
8462 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8464 int Size = Mask.size();
8466 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8467 return M >= 0 && M % Size < Size / 2;
8469 int NumHiInputs = std::count_if(
8470 Mask.begin(), Mask.end(), [Size](int M) { return M % Size > Size / 2; });
8472 bool UnpackLo = NumLoInputs >= NumHiInputs;
8474 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8475 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8476 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8478 for (int i = 0; i < Size; ++i) {
8482 // Each element of the unpack contains Scale elements from this mask.
8483 int UnpackIdx = i / Scale;
8485 // We only handle the case where V1 feeds the first slots of the unpack.
8486 // We rely on canonicalization to ensure this is the case.
8487 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8490 // Setup the mask for this input. The indexing is tricky as we have to
8491 // handle the unpack stride.
8492 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8493 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8497 // Shuffle the inputs into place.
8498 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8499 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8501 // Cast the inputs to the type we will use to unpack them.
8502 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
8503 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
8505 // Unpack the inputs and cast the result back to the desired type.
8506 return DAG.getNode(ISD::BITCAST, DL, VT,
8507 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
8508 DL, UnpackVT, V1, V2));
8511 // We try each unpack from the largest to the smallest to try and find one
8512 // that fits this mask.
8513 int OrigNumElements = VT.getVectorNumElements();
8514 int OrigScalarSize = VT.getScalarSizeInBits();
8515 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8516 int Scale = ScalarSize / OrigScalarSize;
8517 int NumElements = OrigNumElements / Scale;
8518 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8519 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8526 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8528 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8529 /// support for floating point shuffles but not integer shuffles. These
8530 /// instructions will incur a domain crossing penalty on some chips though so
8531 /// it is better to avoid lowering through this for integer vectors where
8533 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8534 const X86Subtarget *Subtarget,
8535 SelectionDAG &DAG) {
8537 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8538 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8539 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8540 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8541 ArrayRef<int> Mask = SVOp->getMask();
8542 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8544 if (isSingleInputShuffleMask(Mask)) {
8545 // Use low duplicate instructions for masks that match their pattern.
8546 if (Subtarget->hasSSE3())
8547 if (isShuffleEquivalent(V1, V2, Mask, 0, 0))
8548 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8550 // Straight shuffle of a single input vector. Simulate this by using the
8551 // single input as both of the "inputs" to this instruction..
8552 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8554 if (Subtarget->hasAVX()) {
8555 // If we have AVX, we can use VPERMILPS which will allow folding a load
8556 // into the shuffle.
8557 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8558 DAG.getConstant(SHUFPDMask, MVT::i8));
8561 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
8562 DAG.getConstant(SHUFPDMask, MVT::i8));
8564 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8565 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8567 // If we have a single input, insert that into V1 if we can do so cheaply.
8568 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8569 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8570 MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
8572 // Try inverting the insertion since for v2 masks it is easy to do and we
8573 // can't reliably sort the mask one way or the other.
8574 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8575 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8576 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8577 MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG))
8581 // Try to use one of the special instruction patterns to handle two common
8582 // blend patterns if a zero-blend above didn't work.
8583 if (isShuffleEquivalent(V1, V2, Mask, 0, 3) || isShuffleEquivalent(V1, V2, Mask, 1, 3))
8584 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8585 // We can either use a special instruction to load over the low double or
8586 // to move just the low double.
8588 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8590 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8592 if (Subtarget->hasSSE41())
8593 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8597 // Use dedicated unpack instructions for masks that match their pattern.
8598 if (isShuffleEquivalent(V1, V2, Mask, 0, 2))
8599 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8600 if (isShuffleEquivalent(V1, V2, Mask, 1, 3))
8601 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8603 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8604 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
8605 DAG.getConstant(SHUFPDMask, MVT::i8));
8608 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8610 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8611 /// the integer unit to minimize domain crossing penalties. However, for blends
8612 /// it falls back to the floating point shuffle operation with appropriate bit
8614 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8615 const X86Subtarget *Subtarget,
8616 SelectionDAG &DAG) {
8618 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8619 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8620 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8621 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8622 ArrayRef<int> Mask = SVOp->getMask();
8623 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8625 if (isSingleInputShuffleMask(Mask)) {
8626 // Check for being able to broadcast a single element.
8627 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1,
8628 Mask, Subtarget, DAG))
8631 // Straight shuffle of a single input vector. For everything from SSE2
8632 // onward this has a single fast instruction with no scary immediates.
8633 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8634 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
8635 int WidenedMask[4] = {
8636 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8637 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8639 ISD::BITCAST, DL, MVT::v2i64,
8640 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
8641 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
8643 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8644 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8645 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8646 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8648 // If we have a blend of two PACKUS operations an the blend aligns with the
8649 // low and half halves, we can just merge the PACKUS operations. This is
8650 // particularly important as it lets us merge shuffles that this routine itself
8652 auto GetPackNode = [](SDValue V) {
8653 while (V.getOpcode() == ISD::BITCAST)
8654 V = V.getOperand(0);
8656 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8658 if (SDValue V1Pack = GetPackNode(V1))
8659 if (SDValue V2Pack = GetPackNode(V2))
8660 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
8661 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8662 Mask[0] == 0 ? V1Pack.getOperand(0)
8663 : V1Pack.getOperand(1),
8664 Mask[1] == 2 ? V2Pack.getOperand(0)
8665 : V2Pack.getOperand(1)));
8667 // Try to use shift instructions.
8669 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8672 // When loading a scalar and then shuffling it into a vector we can often do
8673 // the insertion cheaply.
8674 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8675 MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
8677 // Try inverting the insertion since for v2 masks it is easy to do and we
8678 // can't reliably sort the mask one way or the other.
8679 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8680 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8681 MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG))
8684 // We have different paths for blend lowering, but they all must use the
8685 // *exact* same predicate.
8686 bool IsBlendSupported = Subtarget->hasSSE41();
8687 if (IsBlendSupported)
8688 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8692 // Use dedicated unpack instructions for masks that match their pattern.
8693 if (isShuffleEquivalent(V1, V2, Mask, 0, 2))
8694 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8695 if (isShuffleEquivalent(V1, V2, Mask, 1, 3))
8696 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8698 // Try to use byte rotation instructions.
8699 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8700 if (Subtarget->hasSSSE3())
8701 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8702 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8705 // If we have direct support for blends, we should lower by decomposing into
8706 // a permute. That will be faster than the domain cross.
8707 if (IsBlendSupported)
8708 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8711 // We implement this with SHUFPD which is pretty lame because it will likely
8712 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8713 // However, all the alternatives are still more cycles and newer chips don't
8714 // have this problem. It would be really nice if x86 had better shuffles here.
8715 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
8716 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
8717 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
8718 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8721 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8723 /// This is used to disable more specialized lowerings when the shufps lowering
8724 /// will happen to be efficient.
8725 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8726 // This routine only handles 128-bit shufps.
8727 assert(Mask.size() == 4 && "Unsupported mask size!");
8729 // To lower with a single SHUFPS we need to have the low half and high half
8730 // each requiring a single input.
8731 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8733 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8739 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8741 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8742 /// It makes no assumptions about whether this is the *best* lowering, it simply
8744 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8745 ArrayRef<int> Mask, SDValue V1,
8746 SDValue V2, SelectionDAG &DAG) {
8747 SDValue LowV = V1, HighV = V2;
8748 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8751 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8753 if (NumV2Elements == 1) {
8755 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8758 // Compute the index adjacent to V2Index and in the same half by toggling
8760 int V2AdjIndex = V2Index ^ 1;
8762 if (Mask[V2AdjIndex] == -1) {
8763 // Handles all the cases where we have a single V2 element and an undef.
8764 // This will only ever happen in the high lanes because we commute the
8765 // vector otherwise.
8767 std::swap(LowV, HighV);
8768 NewMask[V2Index] -= 4;
8770 // Handle the case where the V2 element ends up adjacent to a V1 element.
8771 // To make this work, blend them together as the first step.
8772 int V1Index = V2AdjIndex;
8773 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8774 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8775 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8777 // Now proceed to reconstruct the final blend as we have the necessary
8778 // high or low half formed.
8785 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8786 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8788 } else if (NumV2Elements == 2) {
8789 if (Mask[0] < 4 && Mask[1] < 4) {
8790 // Handle the easy case where we have V1 in the low lanes and V2 in the
8794 } else if (Mask[2] < 4 && Mask[3] < 4) {
8795 // We also handle the reversed case because this utility may get called
8796 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8797 // arrange things in the right direction.
8803 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8804 // trying to place elements directly, just blend them and set up the final
8805 // shuffle to place them.
8807 // The first two blend mask elements are for V1, the second two are for
8809 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8810 Mask[2] < 4 ? Mask[2] : Mask[3],
8811 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8812 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8813 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8814 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8816 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8819 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8820 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8821 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8822 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8825 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8826 getV4X86ShuffleImm8ForMask(NewMask, DAG));
8829 /// \brief Lower 4-lane 32-bit floating point shuffles.
8831 /// Uses instructions exclusively from the floating point unit to minimize
8832 /// domain crossing penalties, as these are sufficient to implement all v4f32
8834 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8835 const X86Subtarget *Subtarget,
8836 SelectionDAG &DAG) {
8838 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8839 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8840 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8841 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8842 ArrayRef<int> Mask = SVOp->getMask();
8843 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8846 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8848 if (NumV2Elements == 0) {
8849 // Check for being able to broadcast a single element.
8850 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1,
8851 Mask, Subtarget, DAG))
8854 // Use even/odd duplicate instructions for masks that match their pattern.
8855 if (Subtarget->hasSSE3()) {
8856 if (isShuffleEquivalent(V1, V2, Mask, 0, 0, 2, 2))
8857 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8858 if (isShuffleEquivalent(V1, V2, Mask, 1, 1, 3, 3))
8859 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8862 if (Subtarget->hasAVX()) {
8863 // If we have AVX, we can use VPERMILPS which will allow folding a load
8864 // into the shuffle.
8865 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8866 getV4X86ShuffleImm8ForMask(Mask, DAG));
8869 // Otherwise, use a straight shuffle of a single input vector. We pass the
8870 // input vector to both operands to simulate this with a SHUFPS.
8871 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8872 getV4X86ShuffleImm8ForMask(Mask, DAG));
8875 // There are special ways we can lower some single-element blends. However, we
8876 // have custom ways we can lower more complex single-element blends below that
8877 // we defer to if both this and BLENDPS fail to match, so restrict this to
8878 // when the V2 input is targeting element 0 of the mask -- that is the fast
8880 if (NumV2Elements == 1 && Mask[0] >= 4)
8881 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
8882 Mask, Subtarget, DAG))
8885 if (Subtarget->hasSSE41()) {
8886 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8890 // Use INSERTPS if we can complete the shuffle efficiently.
8891 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8894 if (!isSingleSHUFPSMask(Mask))
8895 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8896 DL, MVT::v4f32, V1, V2, Mask, DAG))
8900 // Use dedicated unpack instructions for masks that match their pattern.
8901 if (isShuffleEquivalent(V1, V2, Mask, 0, 4, 1, 5))
8902 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8903 if (isShuffleEquivalent(V1, V2, Mask, 2, 6, 3, 7))
8904 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8905 if (isShuffleEquivalent(V1, V2, Mask, 4, 0, 5, 1))
8906 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8907 if (isShuffleEquivalent(V1, V2, Mask, 6, 2, 7, 3))
8908 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8910 // Otherwise fall back to a SHUFPS lowering strategy.
8911 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8914 /// \brief Lower 4-lane i32 vector shuffles.
8916 /// We try to handle these with integer-domain shuffles where we can, but for
8917 /// blends we use the floating point domain blend instructions.
8918 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8919 const X86Subtarget *Subtarget,
8920 SelectionDAG &DAG) {
8922 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8923 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8924 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8925 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8926 ArrayRef<int> Mask = SVOp->getMask();
8927 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8929 // Whenever we can lower this as a zext, that instruction is strictly faster
8930 // than any alternative. It also allows us to fold memory operands into the
8931 // shuffle in many cases.
8932 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8933 Mask, Subtarget, DAG))
8937 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8939 if (NumV2Elements == 0) {
8940 // Check for being able to broadcast a single element.
8941 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1,
8942 Mask, Subtarget, DAG))
8945 // Straight shuffle of a single input vector. For everything from SSE2
8946 // onward this has a single fast instruction with no scary immediates.
8947 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8948 // but we aren't actually going to use the UNPCK instruction because doing
8949 // so prevents folding a load into this instruction or making a copy.
8950 const int UnpackLoMask[] = {0, 0, 1, 1};
8951 const int UnpackHiMask[] = {2, 2, 3, 3};
8952 if (isShuffleEquivalent(V1, V2, Mask, 0, 0, 1, 1))
8953 Mask = UnpackLoMask;
8954 else if (isShuffleEquivalent(V1, V2, Mask, 2, 2, 3, 3))
8955 Mask = UnpackHiMask;
8957 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8958 getV4X86ShuffleImm8ForMask(Mask, DAG));
8961 // Try to use shift instructions.
8963 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8966 // There are special ways we can lower some single-element blends.
8967 if (NumV2Elements == 1)
8968 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
8969 Mask, Subtarget, DAG))
8972 // We have different paths for blend lowering, but they all must use the
8973 // *exact* same predicate.
8974 bool IsBlendSupported = Subtarget->hasSSE41();
8975 if (IsBlendSupported)
8976 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8980 if (SDValue Masked =
8981 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8984 // Use dedicated unpack instructions for masks that match their pattern.
8985 if (isShuffleEquivalent(V1, V2, Mask, 0, 4, 1, 5))
8986 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8987 if (isShuffleEquivalent(V1, V2, Mask, 2, 6, 3, 7))
8988 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8989 if (isShuffleEquivalent(V1, V2, Mask, 4, 0, 5, 1))
8990 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8991 if (isShuffleEquivalent(V1, V2, Mask, 6, 2, 7, 3))
8992 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8994 // Try to use byte rotation instructions.
8995 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8996 if (Subtarget->hasSSSE3())
8997 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8998 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
9001 // If we have direct support for blends, we should lower by decomposing into
9002 // a permute. That will be faster than the domain cross.
9003 if (IsBlendSupported)
9004 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
9007 // Try to lower by permuting the inputs into an unpack instruction.
9008 if (SDValue Unpack =
9009 lowerVectorShuffleAsUnpack(MVT::v4i32, DL, V1, V2, Mask, DAG))
9012 // We implement this with SHUFPS because it can blend from two vectors.
9013 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
9014 // up the inputs, bypassing domain shift penalties that we would encur if we
9015 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
9017 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
9018 DAG.getVectorShuffle(
9020 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
9021 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
9024 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
9025 /// shuffle lowering, and the most complex part.
9027 /// The lowering strategy is to try to form pairs of input lanes which are
9028 /// targeted at the same half of the final vector, and then use a dword shuffle
9029 /// to place them onto the right half, and finally unpack the paired lanes into
9030 /// their final position.
9032 /// The exact breakdown of how to form these dword pairs and align them on the
9033 /// correct sides is really tricky. See the comments within the function for
9034 /// more of the details.
9035 static SDValue lowerV8I16SingleInputVectorShuffle(
9036 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
9037 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9038 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9039 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
9040 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
9042 SmallVector<int, 4> LoInputs;
9043 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
9044 [](int M) { return M >= 0; });
9045 std::sort(LoInputs.begin(), LoInputs.end());
9046 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
9047 SmallVector<int, 4> HiInputs;
9048 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
9049 [](int M) { return M >= 0; });
9050 std::sort(HiInputs.begin(), HiInputs.end());
9051 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
9053 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
9054 int NumHToL = LoInputs.size() - NumLToL;
9056 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
9057 int NumHToH = HiInputs.size() - NumLToH;
9058 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
9059 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
9060 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
9061 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
9063 // Check for being able to broadcast a single element.
9064 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V,
9065 Mask, Subtarget, DAG))
9068 // Try to use shift instructions.
9070 lowerVectorShuffleAsShift(DL, MVT::v8i16, V, V, Mask, DAG))
9073 // Use dedicated unpack instructions for masks that match their pattern.
9074 if (isShuffleEquivalent(V, V, Mask, 0, 0, 1, 1, 2, 2, 3, 3))
9075 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
9076 if (isShuffleEquivalent(V, V, Mask, 4, 4, 5, 5, 6, 6, 7, 7))
9077 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
9079 // Try to use byte rotation instructions.
9080 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9081 DL, MVT::v8i16, V, V, Mask, Subtarget, DAG))
9084 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
9085 // such inputs we can swap two of the dwords across the half mark and end up
9086 // with <=2 inputs to each half in each half. Once there, we can fall through
9087 // to the generic code below. For example:
9089 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
9090 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
9092 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
9093 // and an existing 2-into-2 on the other half. In this case we may have to
9094 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
9095 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
9096 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
9097 // because any other situation (including a 3-into-1 or 1-into-3 in the other
9098 // half than the one we target for fixing) will be fixed when we re-enter this
9099 // path. We will also combine away any sequence of PSHUFD instructions that
9100 // result into a single instruction. Here is an example of the tricky case:
9102 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
9103 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
9105 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
9107 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
9108 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
9110 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
9111 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
9113 // The result is fine to be handled by the generic logic.
9114 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
9115 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
9116 int AOffset, int BOffset) {
9117 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
9118 "Must call this with A having 3 or 1 inputs from the A half.");
9119 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
9120 "Must call this with B having 1 or 3 inputs from the B half.");
9121 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
9122 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
9124 // Compute the index of dword with only one word among the three inputs in
9125 // a half by taking the sum of the half with three inputs and subtracting
9126 // the sum of the actual three inputs. The difference is the remaining
9129 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
9130 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
9131 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
9132 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
9133 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
9134 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
9135 int TripleNonInputIdx =
9136 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
9137 TripleDWord = TripleNonInputIdx / 2;
9139 // We use xor with one to compute the adjacent DWord to whichever one the
9141 OneInputDWord = (OneInput / 2) ^ 1;
9143 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
9144 // and BToA inputs. If there is also such a problem with the BToB and AToB
9145 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
9146 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
9147 // is essential that we don't *create* a 3<-1 as then we might oscillate.
9148 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
9149 // Compute how many inputs will be flipped by swapping these DWords. We
9151 // to balance this to ensure we don't form a 3-1 shuffle in the other
9153 int NumFlippedAToBInputs =
9154 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
9155 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
9156 int NumFlippedBToBInputs =
9157 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
9158 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
9159 if ((NumFlippedAToBInputs == 1 &&
9160 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
9161 (NumFlippedBToBInputs == 1 &&
9162 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
9163 // We choose whether to fix the A half or B half based on whether that
9164 // half has zero flipped inputs. At zero, we may not be able to fix it
9165 // with that half. We also bias towards fixing the B half because that
9166 // will more commonly be the high half, and we have to bias one way.
9167 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
9168 ArrayRef<int> Inputs) {
9169 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
9170 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
9171 PinnedIdx ^ 1) != Inputs.end();
9172 // Determine whether the free index is in the flipped dword or the
9173 // unflipped dword based on where the pinned index is. We use this bit
9174 // in an xor to conditionally select the adjacent dword.
9175 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
9176 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9177 FixFreeIdx) != Inputs.end();
9178 if (IsFixIdxInput == IsFixFreeIdxInput)
9180 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9181 FixFreeIdx) != Inputs.end();
9182 assert(IsFixIdxInput != IsFixFreeIdxInput &&
9183 "We need to be changing the number of flipped inputs!");
9184 int PSHUFHalfMask[] = {0, 1, 2, 3};
9185 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
9186 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
9188 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
9191 if (M != -1 && M == FixIdx)
9193 else if (M != -1 && M == FixFreeIdx)
9196 if (NumFlippedBToBInputs != 0) {
9198 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9199 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
9201 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
9203 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9204 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9209 int PSHUFDMask[] = {0, 1, 2, 3};
9210 PSHUFDMask[ADWord] = BDWord;
9211 PSHUFDMask[BDWord] = ADWord;
9212 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9213 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
9214 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
9215 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9217 // Adjust the mask to match the new locations of A and B.
9219 if (M != -1 && M/2 == ADWord)
9220 M = 2 * BDWord + M % 2;
9221 else if (M != -1 && M/2 == BDWord)
9222 M = 2 * ADWord + M % 2;
9224 // Recurse back into this routine to re-compute state now that this isn't
9225 // a 3 and 1 problem.
9226 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
9229 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9230 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9231 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9232 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9234 // At this point there are at most two inputs to the low and high halves from
9235 // each half. That means the inputs can always be grouped into dwords and
9236 // those dwords can then be moved to the correct half with a dword shuffle.
9237 // We use at most one low and one high word shuffle to collect these paired
9238 // inputs into dwords, and finally a dword shuffle to place them.
9239 int PSHUFLMask[4] = {-1, -1, -1, -1};
9240 int PSHUFHMask[4] = {-1, -1, -1, -1};
9241 int PSHUFDMask[4] = {-1, -1, -1, -1};
9243 // First fix the masks for all the inputs that are staying in their
9244 // original halves. This will then dictate the targets of the cross-half
9246 auto fixInPlaceInputs =
9247 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9248 MutableArrayRef<int> SourceHalfMask,
9249 MutableArrayRef<int> HalfMask, int HalfOffset) {
9250 if (InPlaceInputs.empty())
9252 if (InPlaceInputs.size() == 1) {
9253 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9254 InPlaceInputs[0] - HalfOffset;
9255 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9258 if (IncomingInputs.empty()) {
9259 // Just fix all of the in place inputs.
9260 for (int Input : InPlaceInputs) {
9261 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9262 PSHUFDMask[Input / 2] = Input / 2;
9267 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9268 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9269 InPlaceInputs[0] - HalfOffset;
9270 // Put the second input next to the first so that they are packed into
9271 // a dword. We find the adjacent index by toggling the low bit.
9272 int AdjIndex = InPlaceInputs[0] ^ 1;
9273 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9274 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9275 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9277 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9278 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9280 // Now gather the cross-half inputs and place them into a free dword of
9281 // their target half.
9282 // FIXME: This operation could almost certainly be simplified dramatically to
9283 // look more like the 3-1 fixing operation.
9284 auto moveInputsToRightHalf = [&PSHUFDMask](
9285 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9286 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9287 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9289 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9290 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9292 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9294 int LowWord = Word & ~1;
9295 int HighWord = Word | 1;
9296 return isWordClobbered(SourceHalfMask, LowWord) ||
9297 isWordClobbered(SourceHalfMask, HighWord);
9300 if (IncomingInputs.empty())
9303 if (ExistingInputs.empty()) {
9304 // Map any dwords with inputs from them into the right half.
9305 for (int Input : IncomingInputs) {
9306 // If the source half mask maps over the inputs, turn those into
9307 // swaps and use the swapped lane.
9308 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9309 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9310 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9311 Input - SourceOffset;
9312 // We have to swap the uses in our half mask in one sweep.
9313 for (int &M : HalfMask)
9314 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9316 else if (M == Input)
9317 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9319 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9320 Input - SourceOffset &&
9321 "Previous placement doesn't match!");
9323 // Note that this correctly re-maps both when we do a swap and when
9324 // we observe the other side of the swap above. We rely on that to
9325 // avoid swapping the members of the input list directly.
9326 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9329 // Map the input's dword into the correct half.
9330 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9331 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9333 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9335 "Previous placement doesn't match!");
9338 // And just directly shift any other-half mask elements to be same-half
9339 // as we will have mirrored the dword containing the element into the
9340 // same position within that half.
9341 for (int &M : HalfMask)
9342 if (M >= SourceOffset && M < SourceOffset + 4) {
9343 M = M - SourceOffset + DestOffset;
9344 assert(M >= 0 && "This should never wrap below zero!");
9349 // Ensure we have the input in a viable dword of its current half. This
9350 // is particularly tricky because the original position may be clobbered
9351 // by inputs being moved and *staying* in that half.
9352 if (IncomingInputs.size() == 1) {
9353 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9354 int InputFixed = std::find(std::begin(SourceHalfMask),
9355 std::end(SourceHalfMask), -1) -
9356 std::begin(SourceHalfMask) + SourceOffset;
9357 SourceHalfMask[InputFixed - SourceOffset] =
9358 IncomingInputs[0] - SourceOffset;
9359 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9361 IncomingInputs[0] = InputFixed;
9363 } else if (IncomingInputs.size() == 2) {
9364 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9365 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9366 // We have two non-adjacent or clobbered inputs we need to extract from
9367 // the source half. To do this, we need to map them into some adjacent
9368 // dword slot in the source mask.
9369 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9370 IncomingInputs[1] - SourceOffset};
9372 // If there is a free slot in the source half mask adjacent to one of
9373 // the inputs, place the other input in it. We use (Index XOR 1) to
9374 // compute an adjacent index.
9375 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9376 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9377 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9378 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9379 InputsFixed[1] = InputsFixed[0] ^ 1;
9380 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9381 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9382 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9383 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9384 InputsFixed[0] = InputsFixed[1] ^ 1;
9385 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9386 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9387 // The two inputs are in the same DWord but it is clobbered and the
9388 // adjacent DWord isn't used at all. Move both inputs to the free
9390 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9391 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9392 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9393 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9395 // The only way we hit this point is if there is no clobbering
9396 // (because there are no off-half inputs to this half) and there is no
9397 // free slot adjacent to one of the inputs. In this case, we have to
9398 // swap an input with a non-input.
9399 for (int i = 0; i < 4; ++i)
9400 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9401 "We can't handle any clobbers here!");
9402 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9403 "Cannot have adjacent inputs here!");
9405 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9406 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9408 // We also have to update the final source mask in this case because
9409 // it may need to undo the above swap.
9410 for (int &M : FinalSourceHalfMask)
9411 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9412 M = InputsFixed[1] + SourceOffset;
9413 else if (M == InputsFixed[1] + SourceOffset)
9414 M = (InputsFixed[0] ^ 1) + SourceOffset;
9416 InputsFixed[1] = InputsFixed[0] ^ 1;
9419 // Point everything at the fixed inputs.
9420 for (int &M : HalfMask)
9421 if (M == IncomingInputs[0])
9422 M = InputsFixed[0] + SourceOffset;
9423 else if (M == IncomingInputs[1])
9424 M = InputsFixed[1] + SourceOffset;
9426 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9427 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9430 llvm_unreachable("Unhandled input size!");
9433 // Now hoist the DWord down to the right half.
9434 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9435 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9436 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9437 for (int &M : HalfMask)
9438 for (int Input : IncomingInputs)
9440 M = FreeDWord * 2 + Input % 2;
9442 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9443 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9444 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9445 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9447 // Now enact all the shuffles we've computed to move the inputs into their
9449 if (!isNoopShuffleMask(PSHUFLMask))
9450 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9451 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
9452 if (!isNoopShuffleMask(PSHUFHMask))
9453 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9454 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
9455 if (!isNoopShuffleMask(PSHUFDMask))
9456 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9457 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
9458 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
9459 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9461 // At this point, each half should contain all its inputs, and we can then
9462 // just shuffle them into their final position.
9463 assert(std::count_if(LoMask.begin(), LoMask.end(),
9464 [](int M) { return M >= 4; }) == 0 &&
9465 "Failed to lift all the high half inputs to the low mask!");
9466 assert(std::count_if(HiMask.begin(), HiMask.end(),
9467 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9468 "Failed to lift all the low half inputs to the high mask!");
9470 // Do a half shuffle for the low mask.
9471 if (!isNoopShuffleMask(LoMask))
9472 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9473 getV4X86ShuffleImm8ForMask(LoMask, DAG));
9475 // Do a half shuffle with the high mask after shifting its values down.
9476 for (int &M : HiMask)
9479 if (!isNoopShuffleMask(HiMask))
9480 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9481 getV4X86ShuffleImm8ForMask(HiMask, DAG));
9486 /// \brief Detect whether the mask pattern should be lowered through
9489 /// This essentially tests whether viewing the mask as an interleaving of two
9490 /// sub-sequences reduces the cross-input traffic of a blend operation. If so,
9491 /// lowering it through interleaving is a significantly better strategy.
9492 static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
9493 int NumEvenInputs[2] = {0, 0};
9494 int NumOddInputs[2] = {0, 0};
9495 int NumLoInputs[2] = {0, 0};
9496 int NumHiInputs[2] = {0, 0};
9497 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
9501 int InputIdx = Mask[i] >= Size;
9504 ++NumLoInputs[InputIdx];
9506 ++NumHiInputs[InputIdx];
9509 ++NumEvenInputs[InputIdx];
9511 ++NumOddInputs[InputIdx];
9514 // The minimum number of cross-input results for both the interleaved and
9515 // split cases. If interleaving results in fewer cross-input results, return
9517 int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
9518 NumEvenInputs[0] + NumOddInputs[1]);
9519 int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
9520 NumLoInputs[0] + NumHiInputs[1]);
9521 return InterleavedCrosses < SplitCrosses;
9524 /// \brief Blend two v8i16 vectors using a naive unpack strategy.
9526 /// This strategy only works when the inputs from each vector fit into a single
9527 /// half of that vector, and generally there are not so many inputs as to leave
9528 /// the in-place shuffles required highly constrained (and thus expensive). It
9529 /// shifts all the inputs into a single side of both input vectors and then
9530 /// uses an unpack to interleave these inputs in a single vector. At that
9531 /// point, we will fall back on the generic single input shuffle lowering.
9532 static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
9534 MutableArrayRef<int> Mask,
9535 const X86Subtarget *Subtarget,
9536 SelectionDAG &DAG) {
9537 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9538 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9539 SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
9540 for (int i = 0; i < 8; ++i)
9541 if (Mask[i] >= 0 && Mask[i] < 4)
9542 LoV1Inputs.push_back(i);
9543 else if (Mask[i] >= 4 && Mask[i] < 8)
9544 HiV1Inputs.push_back(i);
9545 else if (Mask[i] >= 8 && Mask[i] < 12)
9546 LoV2Inputs.push_back(i);
9547 else if (Mask[i] >= 12)
9548 HiV2Inputs.push_back(i);
9550 int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
9551 int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
9554 assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
9555 assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
9556 assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
9558 bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
9559 HiV1Inputs.size() + HiV2Inputs.size();
9561 auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
9562 ArrayRef<int> HiInputs, bool MoveToLo,
9564 ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
9565 ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
9566 if (BadInputs.empty())
9569 int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9570 int MoveOffset = MoveToLo ? 0 : 4;
9572 if (GoodInputs.empty()) {
9573 for (int BadInput : BadInputs) {
9574 MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
9575 Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
9578 if (GoodInputs.size() == 2) {
9579 // If the low inputs are spread across two dwords, pack them into
9581 MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
9582 MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
9583 Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
9584 Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
9586 // Otherwise pin the good inputs.
9587 for (int GoodInput : GoodInputs)
9588 MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
9591 if (BadInputs.size() == 2) {
9592 // If we have two bad inputs then there may be either one or two good
9593 // inputs fixed in place. Find a fixed input, and then find the *other*
9594 // two adjacent indices by using modular arithmetic.
9596 std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
9597 [](int M) { return M >= 0; }) -
9598 std::begin(MoveMask);
9600 ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset;
9601 assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
9602 assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
9603 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9604 MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
9605 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9606 Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
9608 assert(BadInputs.size() == 1 && "All sizes handled");
9609 int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset,
9610 std::end(MoveMask), -1) -
9611 std::begin(MoveMask);
9612 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9613 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9617 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
9620 V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
9622 V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
9625 // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
9626 // cross-half traffic in the final shuffle.
9628 // Munge the mask to be a single-input mask after the unpack merges the
9632 M = 2 * (M % 4) + (M / 8);
9634 return DAG.getVectorShuffle(
9635 MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
9636 DL, MVT::v8i16, V1, V2),
9637 DAG.getUNDEF(MVT::v8i16), Mask);
9640 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9641 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9642 SDValue V2, ArrayRef<int> Mask,
9643 SelectionDAG &DAG, bool &V1InUse,
9645 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9651 int Size = Mask.size();
9652 int Scale = 16 / Size;
9653 for (int i = 0; i < 16; ++i) {
9654 if (Mask[i / Scale] == -1) {
9655 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9657 const int ZeroMask = 0x80;
9658 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9660 int V2Idx = Mask[i / Scale] < Size
9662 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9663 if (Zeroable[i / Scale])
9664 V1Idx = V2Idx = ZeroMask;
9665 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
9666 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
9667 V1InUse |= (ZeroMask != V1Idx);
9668 V2InUse |= (ZeroMask != V2Idx);
9673 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9674 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
9675 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9677 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9678 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
9679 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9681 // If we need shuffled inputs from both, blend the two.
9683 if (V1InUse && V2InUse)
9684 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9686 V = V1InUse ? V1 : V2;
9688 // Cast the result back to the correct type.
9689 return DAG.getNode(ISD::BITCAST, DL, VT, V);
9692 /// \brief Generic lowering of 8-lane i16 shuffles.
9694 /// This handles both single-input shuffles and combined shuffle/blends with
9695 /// two inputs. The single input shuffles are immediately delegated to
9696 /// a dedicated lowering routine.
9698 /// The blends are lowered in one of three fundamental ways. If there are few
9699 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9700 /// of the input is significantly cheaper when lowered as an interleaving of
9701 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9702 /// halves of the inputs separately (making them have relatively few inputs)
9703 /// and then concatenate them.
9704 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9705 const X86Subtarget *Subtarget,
9706 SelectionDAG &DAG) {
9708 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9709 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9710 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9711 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9712 ArrayRef<int> OrigMask = SVOp->getMask();
9713 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9714 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9715 MutableArrayRef<int> Mask(MaskStorage);
9717 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9719 // Whenever we can lower this as a zext, that instruction is strictly faster
9720 // than any alternative.
9721 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9722 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9725 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9726 auto isV2 = [](int M) { return M >= 8; };
9728 int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
9729 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9731 if (NumV2Inputs == 0)
9732 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
9734 assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
9735 "to be V1-input shuffles.");
9737 // Try to use shift instructions.
9739 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9742 // There are special ways we can lower some single-element blends.
9743 if (NumV2Inputs == 1)
9744 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
9745 Mask, Subtarget, DAG))
9748 // We have different paths for blend lowering, but they all must use the
9749 // *exact* same predicate.
9750 bool IsBlendSupported = Subtarget->hasSSE41();
9751 if (IsBlendSupported)
9752 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9756 if (SDValue Masked =
9757 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9760 // Use dedicated unpack instructions for masks that match their pattern.
9761 if (isShuffleEquivalent(V1, V2, Mask, 0, 8, 1, 9, 2, 10, 3, 11))
9762 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9763 if (isShuffleEquivalent(V1, V2, Mask, 4, 12, 5, 13, 6, 14, 7, 15))
9764 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9766 // Try to use byte rotation instructions.
9767 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9768 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9771 if (SDValue BitBlend =
9772 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9775 if (NumV1Inputs + NumV2Inputs <= 4)
9776 return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
9778 // Check whether an interleaving lowering is likely to be more efficient.
9779 // This isn't perfect but it is a strong heuristic that tends to work well on
9780 // the kinds of shuffles that show up in practice.
9782 // FIXME: Handle 1x, 2x, and 4x interleaving.
9783 if (shouldLowerAsInterleaving(Mask)) {
9784 // FIXME: Figure out whether we should pack these into the low or high
9787 int EMask[8], OMask[8];
9788 for (int i = 0; i < 4; ++i) {
9789 EMask[i] = Mask[2*i];
9790 OMask[i] = Mask[2*i + 1];
9795 SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
9796 SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
9798 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
9801 // Try to lower by permuting the inputs into an unpack instruction unless we
9802 // have direct support for blending.
9803 if (!IsBlendSupported) {
9804 if (SDValue Unpack =
9805 lowerVectorShuffleAsUnpack(MVT::v8i16, DL, V1, V2, Mask, DAG))
9808 // If we can use PSHUFB, that will be better as it can both shuffle and set
9809 // up an efficient blend.
9810 if (Subtarget->hasSSSE3()) {
9811 bool V1InUse, V2InUse;
9812 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9817 // We can always bit-blend if we have to so the fallback strategy is to
9818 // decompose into single-input permutes and blends.
9819 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9823 /// \brief Check whether a compaction lowering can be done by dropping even
9824 /// elements and compute how many times even elements must be dropped.
9826 /// This handles shuffles which take every Nth element where N is a power of
9827 /// two. Example shuffle masks:
9829 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9830 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9831 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9832 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9833 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9834 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9836 /// Any of these lanes can of course be undef.
9838 /// This routine only supports N <= 3.
9839 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9842 /// \returns N above, or the number of times even elements must be dropped if
9843 /// there is such a number. Otherwise returns zero.
9844 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9845 // Figure out whether we're looping over two inputs or just one.
9846 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9848 // The modulus for the shuffle vector entries is based on whether this is
9849 // a single input or not.
9850 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9851 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9852 "We should only be called with masks with a power-of-2 size!");
9854 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9856 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9857 // and 2^3 simultaneously. This is because we may have ambiguity with
9858 // partially undef inputs.
9859 bool ViableForN[3] = {true, true, true};
9861 for (int i = 0, e = Mask.size(); i < e; ++i) {
9862 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9867 bool IsAnyViable = false;
9868 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9869 if (ViableForN[j]) {
9872 // The shuffle mask must be equal to (i * 2^N) % M.
9873 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9876 ViableForN[j] = false;
9878 // Early exit if we exhaust the possible powers of two.
9883 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9887 // Return 0 as there is no viable power of two.
9891 /// \brief Generic lowering of v16i8 shuffles.
9893 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9894 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9895 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9896 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9898 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9899 const X86Subtarget *Subtarget,
9900 SelectionDAG &DAG) {
9902 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9903 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9904 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9905 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9906 ArrayRef<int> Mask = SVOp->getMask();
9907 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9909 // Try to use shift instructions.
9911 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9914 // Try to use byte rotation instructions.
9915 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9916 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9919 // Try to use a zext lowering.
9920 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9921 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9925 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9927 // For single-input shuffles, there are some nicer lowering tricks we can use.
9928 if (NumV2Elements == 0) {
9929 // Check for being able to broadcast a single element.
9930 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1,
9931 Mask, Subtarget, DAG))
9934 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9935 // Notably, this handles splat and partial-splat shuffles more efficiently.
9936 // However, it only makes sense if the pre-duplication shuffle simplifies
9937 // things significantly. Currently, this means we need to be able to
9938 // express the pre-duplication shuffle as an i16 shuffle.
9940 // FIXME: We should check for other patterns which can be widened into an
9941 // i16 shuffle as well.
9942 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9943 for (int i = 0; i < 16; i += 2)
9944 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9949 auto tryToWidenViaDuplication = [&]() -> SDValue {
9950 if (!canWidenViaDuplication(Mask))
9952 SmallVector<int, 4> LoInputs;
9953 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9954 [](int M) { return M >= 0 && M < 8; });
9955 std::sort(LoInputs.begin(), LoInputs.end());
9956 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9958 SmallVector<int, 4> HiInputs;
9959 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9960 [](int M) { return M >= 8; });
9961 std::sort(HiInputs.begin(), HiInputs.end());
9962 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9965 bool TargetLo = LoInputs.size() >= HiInputs.size();
9966 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9967 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9969 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9970 SmallDenseMap<int, int, 8> LaneMap;
9971 for (int I : InPlaceInputs) {
9972 PreDupI16Shuffle[I/2] = I/2;
9975 int j = TargetLo ? 0 : 4, je = j + 4;
9976 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9977 // Check if j is already a shuffle of this input. This happens when
9978 // there are two adjacent bytes after we move the low one.
9979 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9980 // If we haven't yet mapped the input, search for a slot into which
9982 while (j < je && PreDupI16Shuffle[j] != -1)
9986 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9989 // Map this input with the i16 shuffle.
9990 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9993 // Update the lane map based on the mapping we ended up with.
9994 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9997 ISD::BITCAST, DL, MVT::v16i8,
9998 DAG.getVectorShuffle(MVT::v8i16, DL,
9999 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
10000 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
10002 // Unpack the bytes to form the i16s that will be shuffled into place.
10003 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
10004 MVT::v16i8, V1, V1);
10006 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
10007 for (int i = 0; i < 16; ++i)
10008 if (Mask[i] != -1) {
10009 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
10010 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
10011 if (PostDupI16Shuffle[i / 2] == -1)
10012 PostDupI16Shuffle[i / 2] = MappedMask;
10014 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
10015 "Conflicting entrties in the original shuffle!");
10017 return DAG.getNode(
10018 ISD::BITCAST, DL, MVT::v16i8,
10019 DAG.getVectorShuffle(MVT::v8i16, DL,
10020 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
10021 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
10023 if (SDValue V = tryToWidenViaDuplication())
10027 // Use dedicated unpack instructions for masks that match their pattern.
10028 if (isShuffleEquivalent(V1, V2, Mask,
10029 0, 16, 1, 17, 2, 18, 3, 19,
10030 4, 20, 5, 21, 6, 22, 7, 23))
10031 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
10032 if (isShuffleEquivalent(V1, V2, Mask,
10033 8, 24, 9, 25, 10, 26, 11, 27,
10034 12, 28, 13, 29, 14, 30, 15, 31))
10035 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
10037 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
10038 // with PSHUFB. It is important to do this before we attempt to generate any
10039 // blends but after all of the single-input lowerings. If the single input
10040 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
10041 // want to preserve that and we can DAG combine any longer sequences into
10042 // a PSHUFB in the end. But once we start blending from multiple inputs,
10043 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
10044 // and there are *very* few patterns that would actually be faster than the
10045 // PSHUFB approach because of its ability to zero lanes.
10047 // FIXME: The only exceptions to the above are blends which are exact
10048 // interleavings with direct instructions supporting them. We currently don't
10049 // handle those well here.
10050 if (Subtarget->hasSSSE3()) {
10051 bool V1InUse = false;
10052 bool V2InUse = false;
10054 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
10055 DAG, V1InUse, V2InUse);
10057 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
10058 // do so. This avoids using them to handle blends-with-zero which is
10059 // important as a single pshufb is significantly faster for that.
10060 if (V1InUse && V2InUse) {
10061 if (Subtarget->hasSSE41())
10062 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
10063 Mask, Subtarget, DAG))
10066 // We can use an unpack to do the blending rather than an or in some
10067 // cases. Even though the or may be (very minorly) more efficient, we
10068 // preference this lowering because there are common cases where part of
10069 // the complexity of the shuffles goes away when we do the final blend as
10071 // FIXME: It might be worth trying to detect if the unpack-feeding
10072 // shuffles will both be pshufb, in which case we shouldn't bother with
10074 if (SDValue Unpack =
10075 lowerVectorShuffleAsUnpack(MVT::v16i8, DL, V1, V2, Mask, DAG))
10082 // There are special ways we can lower some single-element blends.
10083 if (NumV2Elements == 1)
10084 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2,
10085 Mask, Subtarget, DAG))
10088 if (SDValue BitBlend =
10089 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
10092 // Check whether a compaction lowering can be done. This handles shuffles
10093 // which take every Nth element for some even N. See the helper function for
10096 // We special case these as they can be particularly efficiently handled with
10097 // the PACKUSB instruction on x86 and they show up in common patterns of
10098 // rearranging bytes to truncate wide elements.
10099 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
10100 // NumEvenDrops is the power of two stride of the elements. Another way of
10101 // thinking about it is that we need to drop the even elements this many
10102 // times to get the original input.
10103 bool IsSingleInput = isSingleInputShuffleMask(Mask);
10105 // First we need to zero all the dropped bytes.
10106 assert(NumEvenDrops <= 3 &&
10107 "No support for dropping even elements more than 3 times.");
10108 // We use the mask type to pick which bytes are preserved based on how many
10109 // elements are dropped.
10110 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
10111 SDValue ByteClearMask =
10112 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
10113 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
10114 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
10115 if (!IsSingleInput)
10116 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
10118 // Now pack things back together.
10119 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
10120 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
10121 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
10122 for (int i = 1; i < NumEvenDrops; ++i) {
10123 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
10124 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
10130 // Handle multi-input cases by blending single-input shuffles.
10131 if (NumV2Elements > 0)
10132 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
10135 // The fallback path for single-input shuffles widens this into two v8i16
10136 // vectors with unpacks, shuffles those, and then pulls them back together
10140 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
10141 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
10142 for (int i = 0; i < 16; ++i)
10144 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
10146 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
10148 SDValue VLoHalf, VHiHalf;
10149 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
10150 // them out and avoid using UNPCK{L,H} to extract the elements of V as
10152 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
10153 [](int M) { return M >= 0 && M % 2 == 1; }) &&
10154 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
10155 [](int M) { return M >= 0 && M % 2 == 1; })) {
10156 // Use a mask to drop the high bytes.
10157 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
10158 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
10159 DAG.getConstant(0x00FF, MVT::v8i16));
10161 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
10162 VHiHalf = DAG.getUNDEF(MVT::v8i16);
10164 // Squash the masks to point directly into VLoHalf.
10165 for (int &M : LoBlendMask)
10168 for (int &M : HiBlendMask)
10172 // Otherwise just unpack the low half of V into VLoHalf and the high half into
10173 // VHiHalf so that we can blend them as i16s.
10174 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
10175 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
10176 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
10177 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
10180 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
10181 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
10183 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
10186 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
10188 /// This routine breaks down the specific type of 128-bit shuffle and
10189 /// dispatches to the lowering routines accordingly.
10190 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10191 MVT VT, const X86Subtarget *Subtarget,
10192 SelectionDAG &DAG) {
10193 switch (VT.SimpleTy) {
10195 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10197 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10199 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10201 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10203 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10205 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10208 llvm_unreachable("Unimplemented!");
10212 /// \brief Helper function to test whether a shuffle mask could be
10213 /// simplified by widening the elements being shuffled.
10215 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
10216 /// leaves it in an unspecified state.
10218 /// NOTE: This must handle normal vector shuffle masks and *target* vector
10219 /// shuffle masks. The latter have the special property of a '-2' representing
10220 /// a zero-ed lane of a vector.
10221 static bool canWidenShuffleElements(ArrayRef<int> Mask,
10222 SmallVectorImpl<int> &WidenedMask) {
10223 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
10224 // If both elements are undef, its trivial.
10225 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
10226 WidenedMask.push_back(SM_SentinelUndef);
10230 // Check for an undef mask and a mask value properly aligned to fit with
10231 // a pair of values. If we find such a case, use the non-undef mask's value.
10232 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
10233 WidenedMask.push_back(Mask[i + 1] / 2);
10236 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
10237 WidenedMask.push_back(Mask[i] / 2);
10241 // When zeroing, we need to spread the zeroing across both lanes to widen.
10242 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
10243 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
10244 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
10245 WidenedMask.push_back(SM_SentinelZero);
10251 // Finally check if the two mask values are adjacent and aligned with
10253 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
10254 WidenedMask.push_back(Mask[i] / 2);
10258 // Otherwise we can't safely widen the elements used in this shuffle.
10261 assert(WidenedMask.size() == Mask.size() / 2 &&
10262 "Incorrect size of mask after widening the elements!");
10267 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
10269 /// This routine just extracts two subvectors, shuffles them independently, and
10270 /// then concatenates them back together. This should work effectively with all
10271 /// AVX vector shuffle types.
10272 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10273 SDValue V2, ArrayRef<int> Mask,
10274 SelectionDAG &DAG) {
10275 assert(VT.getSizeInBits() >= 256 &&
10276 "Only for 256-bit or wider vector shuffles!");
10277 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
10278 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
10280 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
10281 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
10283 int NumElements = VT.getVectorNumElements();
10284 int SplitNumElements = NumElements / 2;
10285 MVT ScalarVT = VT.getScalarType();
10286 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
10288 // Rather than splitting build-vectors, just build two narrower build
10289 // vectors. This helps shuffling with splats and zeros.
10290 auto SplitVector = [&](SDValue V) {
10291 while (V.getOpcode() == ISD::BITCAST)
10292 V = V->getOperand(0);
10294 MVT OrigVT = V.getSimpleValueType();
10295 int OrigNumElements = OrigVT.getVectorNumElements();
10296 int OrigSplitNumElements = OrigNumElements / 2;
10297 MVT OrigScalarVT = OrigVT.getScalarType();
10298 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
10302 auto *BV = dyn_cast<BuildVectorSDNode>(V);
10304 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10305 DAG.getIntPtrConstant(0));
10306 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
10307 DAG.getIntPtrConstant(OrigSplitNumElements));
10310 SmallVector<SDValue, 16> LoOps, HiOps;
10311 for (int i = 0; i < OrigSplitNumElements; ++i) {
10312 LoOps.push_back(BV->getOperand(i));
10313 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
10315 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
10316 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
10318 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
10319 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
10322 SDValue LoV1, HiV1, LoV2, HiV2;
10323 std::tie(LoV1, HiV1) = SplitVector(V1);
10324 std::tie(LoV2, HiV2) = SplitVector(V2);
10326 // Now create two 4-way blends of these half-width vectors.
10327 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
10328 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
10329 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
10330 for (int i = 0; i < SplitNumElements; ++i) {
10331 int M = HalfMask[i];
10332 if (M >= NumElements) {
10333 if (M >= NumElements + SplitNumElements)
10337 V2BlendMask.push_back(M - NumElements);
10338 V1BlendMask.push_back(-1);
10339 BlendMask.push_back(SplitNumElements + i);
10340 } else if (M >= 0) {
10341 if (M >= SplitNumElements)
10345 V2BlendMask.push_back(-1);
10346 V1BlendMask.push_back(M);
10347 BlendMask.push_back(i);
10349 V2BlendMask.push_back(-1);
10350 V1BlendMask.push_back(-1);
10351 BlendMask.push_back(-1);
10355 // Because the lowering happens after all combining takes place, we need to
10356 // manually combine these blend masks as much as possible so that we create
10357 // a minimal number of high-level vector shuffle nodes.
10359 // First try just blending the halves of V1 or V2.
10360 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10361 return DAG.getUNDEF(SplitVT);
10362 if (!UseLoV2 && !UseHiV2)
10363 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10364 if (!UseLoV1 && !UseHiV1)
10365 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10367 SDValue V1Blend, V2Blend;
10368 if (UseLoV1 && UseHiV1) {
10370 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10372 // We only use half of V1 so map the usage down into the final blend mask.
10373 V1Blend = UseLoV1 ? LoV1 : HiV1;
10374 for (int i = 0; i < SplitNumElements; ++i)
10375 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10376 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10378 if (UseLoV2 && UseHiV2) {
10380 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10382 // We only use half of V2 so map the usage down into the final blend mask.
10383 V2Blend = UseLoV2 ? LoV2 : HiV2;
10384 for (int i = 0; i < SplitNumElements; ++i)
10385 if (BlendMask[i] >= SplitNumElements)
10386 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10388 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10390 SDValue Lo = HalfBlend(LoMask);
10391 SDValue Hi = HalfBlend(HiMask);
10392 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10395 /// \brief Either split a vector in halves or decompose the shuffles and the
10398 /// This is provided as a good fallback for many lowerings of non-single-input
10399 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10400 /// between splitting the shuffle into 128-bit components and stitching those
10401 /// back together vs. extracting the single-input shuffles and blending those
10403 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10404 SDValue V2, ArrayRef<int> Mask,
10405 SelectionDAG &DAG) {
10406 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10407 "lower single-input shuffles as it "
10408 "could then recurse on itself.");
10409 int Size = Mask.size();
10411 // If this can be modeled as a broadcast of two elements followed by a blend,
10412 // prefer that lowering. This is especially important because broadcasts can
10413 // often fold with memory operands.
10414 auto DoBothBroadcast = [&] {
10415 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10418 if (V2BroadcastIdx == -1)
10419 V2BroadcastIdx = M - Size;
10420 else if (M - Size != V2BroadcastIdx)
10422 } else if (M >= 0) {
10423 if (V1BroadcastIdx == -1)
10424 V1BroadcastIdx = M;
10425 else if (M != V1BroadcastIdx)
10430 if (DoBothBroadcast())
10431 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10434 // If the inputs all stem from a single 128-bit lane of each input, then we
10435 // split them rather than blending because the split will decompose to
10436 // unusually few instructions.
10437 int LaneCount = VT.getSizeInBits() / 128;
10438 int LaneSize = Size / LaneCount;
10439 SmallBitVector LaneInputs[2];
10440 LaneInputs[0].resize(LaneCount, false);
10441 LaneInputs[1].resize(LaneCount, false);
10442 for (int i = 0; i < Size; ++i)
10444 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10445 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10446 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10448 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10449 // that the decomposed single-input shuffles don't end up here.
10450 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10453 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10454 /// a permutation and blend of those lanes.
10456 /// This essentially blends the out-of-lane inputs to each lane into the lane
10457 /// from a permuted copy of the vector. This lowering strategy results in four
10458 /// instructions in the worst case for a single-input cross lane shuffle which
10459 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10460 /// of. Special cases for each particular shuffle pattern should be handled
10461 /// prior to trying this lowering.
10462 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10463 SDValue V1, SDValue V2,
10464 ArrayRef<int> Mask,
10465 SelectionDAG &DAG) {
10466 // FIXME: This should probably be generalized for 512-bit vectors as well.
10467 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
10468 int LaneSize = Mask.size() / 2;
10470 // If there are only inputs from one 128-bit lane, splitting will in fact be
10471 // less expensive. The flags track wether the given lane contains an element
10472 // that crosses to another lane.
10473 bool LaneCrossing[2] = {false, false};
10474 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10475 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10476 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10477 if (!LaneCrossing[0] || !LaneCrossing[1])
10478 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10480 if (isSingleInputShuffleMask(Mask)) {
10481 SmallVector<int, 32> FlippedBlendMask;
10482 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10483 FlippedBlendMask.push_back(
10484 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10486 : Mask[i] % LaneSize +
10487 (i / LaneSize) * LaneSize + Size));
10489 // Flip the vector, and blend the results which should now be in-lane. The
10490 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10491 // 5 for the high source. The value 3 selects the high half of source 2 and
10492 // the value 2 selects the low half of source 2. We only use source 2 to
10493 // allow folding it into a memory operand.
10494 unsigned PERMMask = 3 | 2 << 4;
10495 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10496 V1, DAG.getConstant(PERMMask, MVT::i8));
10497 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10500 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10501 // will be handled by the above logic and a blend of the results, much like
10502 // other patterns in AVX.
10503 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10506 /// \brief Handle lowering 2-lane 128-bit shuffles.
10507 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10508 SDValue V2, ArrayRef<int> Mask,
10509 const X86Subtarget *Subtarget,
10510 SelectionDAG &DAG) {
10511 // Blends are faster and handle all the non-lane-crossing cases.
10512 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10516 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10517 VT.getVectorNumElements() / 2);
10518 // Check for patterns which can be matched with a single insert of a 128-bit
10520 if (isShuffleEquivalent(V1, V2, Mask, 0, 1, 0, 1) ||
10521 isShuffleEquivalent(V1, V2, Mask, 0, 1, 4, 5)) {
10522 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10523 DAG.getIntPtrConstant(0));
10524 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10525 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
10526 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10528 if (isShuffleEquivalent(V1, V2, Mask, 0, 1, 6, 7)) {
10529 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10530 DAG.getIntPtrConstant(0));
10531 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
10532 DAG.getIntPtrConstant(2));
10533 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10536 // Otherwise form a 128-bit permutation.
10537 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
10538 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
10539 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10540 DAG.getConstant(PermMask, MVT::i8));
10543 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10544 /// shuffling each lane.
10546 /// This will only succeed when the result of fixing the 128-bit lanes results
10547 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10548 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10549 /// the lane crosses early and then use simpler shuffles within each lane.
10551 /// FIXME: It might be worthwhile at some point to support this without
10552 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10553 /// in x86 only floating point has interesting non-repeating shuffles, and even
10554 /// those are still *marginally* more expensive.
10555 static SDValue lowerVectorShuffleByMerging128BitLanes(
10556 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10557 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10558 assert(!isSingleInputShuffleMask(Mask) &&
10559 "This is only useful with multiple inputs.");
10561 int Size = Mask.size();
10562 int LaneSize = 128 / VT.getScalarSizeInBits();
10563 int NumLanes = Size / LaneSize;
10564 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10566 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10567 // check whether the in-128-bit lane shuffles share a repeating pattern.
10568 SmallVector<int, 4> Lanes;
10569 Lanes.resize(NumLanes, -1);
10570 SmallVector<int, 4> InLaneMask;
10571 InLaneMask.resize(LaneSize, -1);
10572 for (int i = 0; i < Size; ++i) {
10576 int j = i / LaneSize;
10578 if (Lanes[j] < 0) {
10579 // First entry we've seen for this lane.
10580 Lanes[j] = Mask[i] / LaneSize;
10581 } else if (Lanes[j] != Mask[i] / LaneSize) {
10582 // This doesn't match the lane selected previously!
10586 // Check that within each lane we have a consistent shuffle mask.
10587 int k = i % LaneSize;
10588 if (InLaneMask[k] < 0) {
10589 InLaneMask[k] = Mask[i] % LaneSize;
10590 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10591 // This doesn't fit a repeating in-lane mask.
10596 // First shuffle the lanes into place.
10597 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10598 VT.getSizeInBits() / 64);
10599 SmallVector<int, 8> LaneMask;
10600 LaneMask.resize(NumLanes * 2, -1);
10601 for (int i = 0; i < NumLanes; ++i)
10602 if (Lanes[i] >= 0) {
10603 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10604 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10607 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
10608 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
10609 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10611 // Cast it back to the type we actually want.
10612 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
10614 // Now do a simple shuffle that isn't lane crossing.
10615 SmallVector<int, 8> NewMask;
10616 NewMask.resize(Size, -1);
10617 for (int i = 0; i < Size; ++i)
10619 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10620 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10621 "Must not introduce lane crosses at this point!");
10623 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10626 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10629 /// This returns true if the elements from a particular input are already in the
10630 /// slot required by the given mask and require no permutation.
10631 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10632 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10633 int Size = Mask.size();
10634 for (int i = 0; i < Size; ++i)
10635 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10641 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10643 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10644 /// isn't available.
10645 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10646 const X86Subtarget *Subtarget,
10647 SelectionDAG &DAG) {
10649 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10650 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10651 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10652 ArrayRef<int> Mask = SVOp->getMask();
10653 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10655 SmallVector<int, 4> WidenedMask;
10656 if (canWidenShuffleElements(Mask, WidenedMask))
10657 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10660 if (isSingleInputShuffleMask(Mask)) {
10661 // Check for being able to broadcast a single element.
10662 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1,
10663 Mask, Subtarget, DAG))
10666 // Use low duplicate instructions for masks that match their pattern.
10667 if (isShuffleEquivalent(V1, V2, Mask, 0, 0, 2, 2))
10668 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10670 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10671 // Non-half-crossing single input shuffles can be lowerid with an
10672 // interleaved permutation.
10673 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10674 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10675 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10676 DAG.getConstant(VPERMILPMask, MVT::i8));
10679 // With AVX2 we have direct support for this permutation.
10680 if (Subtarget->hasAVX2())
10681 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10682 getV4X86ShuffleImm8ForMask(Mask, DAG));
10684 // Otherwise, fall back.
10685 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10689 // X86 has dedicated unpack instructions that can handle specific blend
10690 // operations: UNPCKH and UNPCKL.
10691 if (isShuffleEquivalent(V1, V2, Mask, 0, 4, 2, 6))
10692 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10693 if (isShuffleEquivalent(V1, V2, Mask, 1, 5, 3, 7))
10694 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10695 if (isShuffleEquivalent(V1, V2, Mask, 4, 0, 6, 2))
10696 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
10697 if (isShuffleEquivalent(V1, V2, Mask, 5, 1, 7, 3))
10698 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
10700 // If we have a single input to the zero element, insert that into V1 if we
10701 // can do so cheaply.
10702 int NumV2Elements =
10703 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
10704 if (NumV2Elements == 1 && Mask[0] >= 4)
10705 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10706 MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
10709 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10713 // Check if the blend happens to exactly fit that of SHUFPD.
10714 if ((Mask[0] == -1 || Mask[0] < 2) &&
10715 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
10716 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
10717 (Mask[3] == -1 || Mask[3] >= 6)) {
10718 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
10719 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
10720 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
10721 DAG.getConstant(SHUFPDMask, MVT::i8));
10723 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
10724 (Mask[1] == -1 || Mask[1] < 2) &&
10725 (Mask[2] == -1 || Mask[2] >= 6) &&
10726 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
10727 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
10728 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
10729 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
10730 DAG.getConstant(SHUFPDMask, MVT::i8));
10733 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10734 // shuffle. However, if we have AVX2 and either inputs are already in place,
10735 // we will be able to shuffle even across lanes the other input in a single
10736 // instruction so skip this pattern.
10737 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10738 isShuffleMaskInputInPlace(1, Mask))))
10739 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10740 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10743 // If we have AVX2 then we always want to lower with a blend because an v4 we
10744 // can fully permute the elements.
10745 if (Subtarget->hasAVX2())
10746 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10749 // Otherwise fall back on generic lowering.
10750 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10753 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10755 /// This routine is only called when we have AVX2 and thus a reasonable
10756 /// instruction set for v4i64 shuffling..
10757 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10758 const X86Subtarget *Subtarget,
10759 SelectionDAG &DAG) {
10761 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10762 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10763 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10764 ArrayRef<int> Mask = SVOp->getMask();
10765 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10766 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10768 SmallVector<int, 4> WidenedMask;
10769 if (canWidenShuffleElements(Mask, WidenedMask))
10770 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10773 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10777 // Check for being able to broadcast a single element.
10778 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1,
10779 Mask, Subtarget, DAG))
10782 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10783 // use lower latency instructions that will operate on both 128-bit lanes.
10784 SmallVector<int, 2> RepeatedMask;
10785 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10786 if (isSingleInputShuffleMask(Mask)) {
10787 int PSHUFDMask[] = {-1, -1, -1, -1};
10788 for (int i = 0; i < 2; ++i)
10789 if (RepeatedMask[i] >= 0) {
10790 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10791 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10793 return DAG.getNode(
10794 ISD::BITCAST, DL, MVT::v4i64,
10795 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10796 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
10797 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
10801 // AVX2 provides a direct instruction for permuting a single input across
10803 if (isSingleInputShuffleMask(Mask))
10804 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10805 getV4X86ShuffleImm8ForMask(Mask, DAG));
10807 // Try to use shift instructions.
10808 if (SDValue Shift =
10809 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10812 // Use dedicated unpack instructions for masks that match their pattern.
10813 if (isShuffleEquivalent(V1, V2, Mask, 0, 4, 2, 6))
10814 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10815 if (isShuffleEquivalent(V1, V2, Mask, 1, 5, 3, 7))
10816 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10817 if (isShuffleEquivalent(V1, V2, Mask, 4, 0, 6, 2))
10818 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
10819 if (isShuffleEquivalent(V1, V2, Mask, 5, 1, 7, 3))
10820 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
10822 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10823 // shuffle. However, if we have AVX2 and either inputs are already in place,
10824 // we will be able to shuffle even across lanes the other input in a single
10825 // instruction so skip this pattern.
10826 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10827 isShuffleMaskInputInPlace(1, Mask))))
10828 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10829 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10832 // Otherwise fall back on generic blend lowering.
10833 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10837 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10839 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10840 /// isn't available.
10841 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10842 const X86Subtarget *Subtarget,
10843 SelectionDAG &DAG) {
10845 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10846 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10847 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10848 ArrayRef<int> Mask = SVOp->getMask();
10849 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10851 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10855 // Check for being able to broadcast a single element.
10856 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1,
10857 Mask, Subtarget, DAG))
10860 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10861 // options to efficiently lower the shuffle.
10862 SmallVector<int, 4> RepeatedMask;
10863 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10864 assert(RepeatedMask.size() == 4 &&
10865 "Repeated masks must be half the mask width!");
10867 // Use even/odd duplicate instructions for masks that match their pattern.
10868 if (isShuffleEquivalent(V1, V2, Mask, 0, 0, 2, 2, 4, 4, 6, 6))
10869 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10870 if (isShuffleEquivalent(V1, V2, Mask, 1, 1, 3, 3, 5, 5, 7, 7))
10871 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10873 if (isSingleInputShuffleMask(Mask))
10874 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10875 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10877 // Use dedicated unpack instructions for masks that match their pattern.
10878 if (isShuffleEquivalent(V1, V2, Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10879 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10880 if (isShuffleEquivalent(V1, V2, Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10881 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10882 if (isShuffleEquivalent(V1, V2, Mask, 8, 0, 9, 1, 12, 4, 13, 5))
10883 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10884 if (isShuffleEquivalent(V1, V2, Mask, 10, 2, 11, 3, 14, 6, 15, 7))
10885 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10887 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10888 // have already handled any direct blends. We also need to squash the
10889 // repeated mask into a simulated v4f32 mask.
10890 for (int i = 0; i < 4; ++i)
10891 if (RepeatedMask[i] >= 8)
10892 RepeatedMask[i] -= 4;
10893 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10896 // If we have a single input shuffle with different shuffle patterns in the
10897 // two 128-bit lanes use the variable mask to VPERMILPS.
10898 if (isSingleInputShuffleMask(Mask)) {
10899 SDValue VPermMask[8];
10900 for (int i = 0; i < 8; ++i)
10901 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10902 : DAG.getConstant(Mask[i], MVT::i32);
10903 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10904 return DAG.getNode(
10905 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10906 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10908 if (Subtarget->hasAVX2())
10909 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
10910 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
10911 DAG.getNode(ISD::BUILD_VECTOR, DL,
10912 MVT::v8i32, VPermMask)),
10915 // Otherwise, fall back.
10916 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10920 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10922 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10923 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10926 // If we have AVX2 then we always want to lower with a blend because at v8 we
10927 // can fully permute the elements.
10928 if (Subtarget->hasAVX2())
10929 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10932 // Otherwise fall back on generic lowering.
10933 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10936 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10938 /// This routine is only called when we have AVX2 and thus a reasonable
10939 /// instruction set for v8i32 shuffling..
10940 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10941 const X86Subtarget *Subtarget,
10942 SelectionDAG &DAG) {
10944 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10945 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10946 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10947 ArrayRef<int> Mask = SVOp->getMask();
10948 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10949 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10951 // Whenever we can lower this as a zext, that instruction is strictly faster
10952 // than any alternative. It also allows us to fold memory operands into the
10953 // shuffle in many cases.
10954 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10955 Mask, Subtarget, DAG))
10958 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10962 // Check for being able to broadcast a single element.
10963 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1,
10964 Mask, Subtarget, DAG))
10967 // If the shuffle mask is repeated in each 128-bit lane we can use more
10968 // efficient instructions that mirror the shuffles across the two 128-bit
10970 SmallVector<int, 4> RepeatedMask;
10971 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10972 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10973 if (isSingleInputShuffleMask(Mask))
10974 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10975 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10977 // Use dedicated unpack instructions for masks that match their pattern.
10978 if (isShuffleEquivalent(V1, V2, Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10979 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10980 if (isShuffleEquivalent(V1, V2, Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10981 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10982 if (isShuffleEquivalent(V1, V2, Mask, 8, 0, 9, 1, 12, 4, 13, 5))
10983 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10984 if (isShuffleEquivalent(V1, V2, Mask, 10, 2, 11, 3, 14, 6, 15, 7))
10985 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10988 // Try to use shift instructions.
10989 if (SDValue Shift =
10990 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10993 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10994 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10997 // If the shuffle patterns aren't repeated but it is a single input, directly
10998 // generate a cross-lane VPERMD instruction.
10999 if (isSingleInputShuffleMask(Mask)) {
11000 SDValue VPermMask[8];
11001 for (int i = 0; i < 8; ++i)
11002 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
11003 : DAG.getConstant(Mask[i], MVT::i32);
11004 return DAG.getNode(
11005 X86ISD::VPERMV, DL, MVT::v8i32,
11006 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
11009 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11011 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11012 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
11015 // Otherwise fall back on generic blend lowering.
11016 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
11020 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
11022 /// This routine is only called when we have AVX2 and thus a reasonable
11023 /// instruction set for v16i16 shuffling..
11024 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11025 const X86Subtarget *Subtarget,
11026 SelectionDAG &DAG) {
11028 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
11029 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
11030 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11031 ArrayRef<int> Mask = SVOp->getMask();
11032 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11033 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
11035 // Whenever we can lower this as a zext, that instruction is strictly faster
11036 // than any alternative. It also allows us to fold memory operands into the
11037 // shuffle in many cases.
11038 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
11039 Mask, Subtarget, DAG))
11042 // Check for being able to broadcast a single element.
11043 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1,
11044 Mask, Subtarget, DAG))
11047 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
11051 // Use dedicated unpack instructions for masks that match their pattern.
11052 if (isShuffleEquivalent(V1, V2, Mask,
11053 // First 128-bit lane:
11054 0, 16, 1, 17, 2, 18, 3, 19,
11055 // Second 128-bit lane:
11056 8, 24, 9, 25, 10, 26, 11, 27))
11057 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
11058 if (isShuffleEquivalent(V1, V2, Mask,
11059 // First 128-bit lane:
11060 4, 20, 5, 21, 6, 22, 7, 23,
11061 // Second 128-bit lane:
11062 12, 28, 13, 29, 14, 30, 15, 31))
11063 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
11065 // Try to use shift instructions.
11066 if (SDValue Shift =
11067 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
11070 // Try to use byte rotation instructions.
11071 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
11072 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
11075 if (isSingleInputShuffleMask(Mask)) {
11076 // There are no generalized cross-lane shuffle operations available on i16
11078 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
11079 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
11082 SDValue PSHUFBMask[32];
11083 for (int i = 0; i < 16; ++i) {
11084 if (Mask[i] == -1) {
11085 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
11089 int M = i < 8 ? Mask[i] : Mask[i] - 8;
11090 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
11091 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
11092 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
11094 return DAG.getNode(
11095 ISD::BITCAST, DL, MVT::v16i16,
11097 X86ISD::PSHUFB, DL, MVT::v32i8,
11098 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
11099 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
11102 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11104 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11105 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
11108 // Otherwise fall back on generic lowering.
11109 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
11112 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
11114 /// This routine is only called when we have AVX2 and thus a reasonable
11115 /// instruction set for v32i8 shuffling..
11116 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11117 const X86Subtarget *Subtarget,
11118 SelectionDAG &DAG) {
11120 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
11121 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
11122 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11123 ArrayRef<int> Mask = SVOp->getMask();
11124 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11125 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
11127 // Whenever we can lower this as a zext, that instruction is strictly faster
11128 // than any alternative. It also allows us to fold memory operands into the
11129 // shuffle in many cases.
11130 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
11131 Mask, Subtarget, DAG))
11134 // Check for being able to broadcast a single element.
11135 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1,
11136 Mask, Subtarget, DAG))
11139 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
11143 // Use dedicated unpack instructions for masks that match their pattern.
11144 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
11146 if (isShuffleEquivalent(
11148 // First 128-bit lane:
11149 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
11150 // Second 128-bit lane:
11151 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55))
11152 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
11153 if (isShuffleEquivalent(
11155 // First 128-bit lane:
11156 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
11157 // Second 128-bit lane:
11158 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63))
11159 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
11161 // Try to use shift instructions.
11162 if (SDValue Shift =
11163 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
11166 // Try to use byte rotation instructions.
11167 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
11168 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
11171 if (isSingleInputShuffleMask(Mask)) {
11172 // There are no generalized cross-lane shuffle operations available on i8
11174 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
11175 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
11178 SDValue PSHUFBMask[32];
11179 for (int i = 0; i < 32; ++i)
11182 ? DAG.getUNDEF(MVT::i8)
11183 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
11185 return DAG.getNode(
11186 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
11187 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
11190 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11192 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11193 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
11196 // Otherwise fall back on generic lowering.
11197 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
11200 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
11202 /// This routine either breaks down the specific type of a 256-bit x86 vector
11203 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
11204 /// together based on the available instructions.
11205 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11206 MVT VT, const X86Subtarget *Subtarget,
11207 SelectionDAG &DAG) {
11209 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11210 ArrayRef<int> Mask = SVOp->getMask();
11212 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
11213 // check for those subtargets here and avoid much of the subtarget querying in
11214 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
11215 // ability to manipulate a 256-bit vector with integer types. Since we'll use
11216 // floating point types there eventually, just immediately cast everything to
11217 // a float and operate entirely in that domain.
11218 if (VT.isInteger() && !Subtarget->hasAVX2()) {
11219 int ElementBits = VT.getScalarSizeInBits();
11220 if (ElementBits < 32)
11221 // No floating point type available, decompose into 128-bit vectors.
11222 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11224 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
11225 VT.getVectorNumElements());
11226 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
11227 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
11228 return DAG.getNode(ISD::BITCAST, DL, VT,
11229 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
11232 switch (VT.SimpleTy) {
11234 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11236 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11238 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11240 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11242 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11244 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11247 llvm_unreachable("Not a valid 256-bit x86 vector type!");
11251 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
11252 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11253 const X86Subtarget *Subtarget,
11254 SelectionDAG &DAG) {
11256 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11257 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11258 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11259 ArrayRef<int> Mask = SVOp->getMask();
11260 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11262 // X86 has dedicated unpack instructions that can handle specific blend
11263 // operations: UNPCKH and UNPCKL.
11264 if (isShuffleEquivalent(V1, V2, Mask, 0, 8, 2, 10, 4, 12, 6, 14))
11265 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
11266 if (isShuffleEquivalent(V1, V2, Mask, 1, 9, 3, 11, 5, 13, 7, 15))
11267 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
11269 // FIXME: Implement direct support for this type!
11270 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
11273 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
11274 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11275 const X86Subtarget *Subtarget,
11276 SelectionDAG &DAG) {
11278 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11279 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11280 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11281 ArrayRef<int> Mask = SVOp->getMask();
11282 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11284 // Use dedicated unpack instructions for masks that match their pattern.
11285 if (isShuffleEquivalent(V1, V2, Mask,
11286 0, 16, 1, 17, 4, 20, 5, 21,
11287 8, 24, 9, 25, 12, 28, 13, 29))
11288 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
11289 if (isShuffleEquivalent(V1, V2, Mask,
11290 2, 18, 3, 19, 6, 22, 7, 23,
11291 10, 26, 11, 27, 14, 30, 15, 31))
11292 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
11294 // FIXME: Implement direct support for this type!
11295 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
11298 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11299 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11300 const X86Subtarget *Subtarget,
11301 SelectionDAG &DAG) {
11303 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11304 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11305 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11306 ArrayRef<int> Mask = SVOp->getMask();
11307 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11309 // X86 has dedicated unpack instructions that can handle specific blend
11310 // operations: UNPCKH and UNPCKL.
11311 if (isShuffleEquivalent(V1, V2, Mask, 0, 8, 2, 10, 4, 12, 6, 14))
11312 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
11313 if (isShuffleEquivalent(V1, V2, Mask, 1, 9, 3, 11, 5, 13, 7, 15))
11314 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
11316 // FIXME: Implement direct support for this type!
11317 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
11320 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11321 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11322 const X86Subtarget *Subtarget,
11323 SelectionDAG &DAG) {
11325 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11326 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11327 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11328 ArrayRef<int> Mask = SVOp->getMask();
11329 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11331 // Use dedicated unpack instructions for masks that match their pattern.
11332 if (isShuffleEquivalent(V1, V2, Mask,
11333 0, 16, 1, 17, 4, 20, 5, 21,
11334 8, 24, 9, 25, 12, 28, 13, 29))
11335 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
11336 if (isShuffleEquivalent(V1, V2, Mask,
11337 2, 18, 3, 19, 6, 22, 7, 23,
11338 10, 26, 11, 27, 14, 30, 15, 31))
11339 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
11341 // FIXME: Implement direct support for this type!
11342 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
11345 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11346 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11347 const X86Subtarget *Subtarget,
11348 SelectionDAG &DAG) {
11350 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11351 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11352 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11353 ArrayRef<int> Mask = SVOp->getMask();
11354 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11355 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11357 // FIXME: Implement direct support for this type!
11358 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
11361 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11362 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11363 const X86Subtarget *Subtarget,
11364 SelectionDAG &DAG) {
11366 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11367 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11368 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11369 ArrayRef<int> Mask = SVOp->getMask();
11370 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11371 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11373 // FIXME: Implement direct support for this type!
11374 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11377 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11379 /// This routine either breaks down the specific type of a 512-bit x86 vector
11380 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11381 /// together based on the available instructions.
11382 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11383 MVT VT, const X86Subtarget *Subtarget,
11384 SelectionDAG &DAG) {
11386 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11387 ArrayRef<int> Mask = SVOp->getMask();
11388 assert(Subtarget->hasAVX512() &&
11389 "Cannot lower 512-bit vectors w/ basic ISA!");
11391 // Check for being able to broadcast a single element.
11392 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
11393 Mask, Subtarget, DAG))
11396 // Dispatch to each element type for lowering. If we don't have supprot for
11397 // specific element type shuffles at 512 bits, immediately split them and
11398 // lower them. Each lowering routine of a given type is allowed to assume that
11399 // the requisite ISA extensions for that element type are available.
11400 switch (VT.SimpleTy) {
11402 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11404 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11406 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11408 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11410 if (Subtarget->hasBWI())
11411 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11414 if (Subtarget->hasBWI())
11415 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11419 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11422 // Otherwise fall back on splitting.
11423 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11426 /// \brief Top-level lowering for x86 vector shuffles.
11428 /// This handles decomposition, canonicalization, and lowering of all x86
11429 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11430 /// above in helper routines. The canonicalization attempts to widen shuffles
11431 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11432 /// s.t. only one of the two inputs needs to be tested, etc.
11433 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11434 SelectionDAG &DAG) {
11435 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11436 ArrayRef<int> Mask = SVOp->getMask();
11437 SDValue V1 = Op.getOperand(0);
11438 SDValue V2 = Op.getOperand(1);
11439 MVT VT = Op.getSimpleValueType();
11440 int NumElements = VT.getVectorNumElements();
11443 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
11445 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11446 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11447 if (V1IsUndef && V2IsUndef)
11448 return DAG.getUNDEF(VT);
11450 // When we create a shuffle node we put the UNDEF node to second operand,
11451 // but in some cases the first operand may be transformed to UNDEF.
11452 // In this case we should just commute the node.
11454 return DAG.getCommutedVectorShuffle(*SVOp);
11456 // Check for non-undef masks pointing at an undef vector and make the masks
11457 // undef as well. This makes it easier to match the shuffle based solely on
11461 if (M >= NumElements) {
11462 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11463 for (int &M : NewMask)
11464 if (M >= NumElements)
11466 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11469 // We actually see shuffles that are entirely re-arrangements of a set of
11470 // zero inputs. This mostly happens while decomposing complex shuffles into
11471 // simple ones. Directly lower these as a buildvector of zeros.
11472 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11473 if (Zeroable.all())
11474 return getZeroVector(VT, Subtarget, DAG, dl);
11476 // Try to collapse shuffles into using a vector type with fewer elements but
11477 // wider element types. We cap this to not form integers or floating point
11478 // elements wider than 64 bits, but it might be interesting to form i128
11479 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11480 SmallVector<int, 16> WidenedMask;
11481 if (VT.getScalarSizeInBits() < 64 &&
11482 canWidenShuffleElements(Mask, WidenedMask)) {
11483 MVT NewEltVT = VT.isFloatingPoint()
11484 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11485 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11486 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11487 // Make sure that the new vector type is legal. For example, v2f64 isn't
11489 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11490 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
11491 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
11492 return DAG.getNode(ISD::BITCAST, dl, VT,
11493 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11497 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11498 for (int M : SVOp->getMask())
11500 ++NumUndefElements;
11501 else if (M < NumElements)
11506 // Commute the shuffle as needed such that more elements come from V1 than
11507 // V2. This allows us to match the shuffle pattern strictly on how many
11508 // elements come from V1 without handling the symmetric cases.
11509 if (NumV2Elements > NumV1Elements)
11510 return DAG.getCommutedVectorShuffle(*SVOp);
11512 // When the number of V1 and V2 elements are the same, try to minimize the
11513 // number of uses of V2 in the low half of the vector. When that is tied,
11514 // ensure that the sum of indices for V1 is equal to or lower than the sum
11515 // indices for V2. When those are equal, try to ensure that the number of odd
11516 // indices for V1 is lower than the number of odd indices for V2.
11517 if (NumV1Elements == NumV2Elements) {
11518 int LowV1Elements = 0, LowV2Elements = 0;
11519 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11520 if (M >= NumElements)
11524 if (LowV2Elements > LowV1Elements) {
11525 return DAG.getCommutedVectorShuffle(*SVOp);
11526 } else if (LowV2Elements == LowV1Elements) {
11527 int SumV1Indices = 0, SumV2Indices = 0;
11528 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11529 if (SVOp->getMask()[i] >= NumElements)
11531 else if (SVOp->getMask()[i] >= 0)
11533 if (SumV2Indices < SumV1Indices) {
11534 return DAG.getCommutedVectorShuffle(*SVOp);
11535 } else if (SumV2Indices == SumV1Indices) {
11536 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11537 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11538 if (SVOp->getMask()[i] >= NumElements)
11539 NumV2OddIndices += i % 2;
11540 else if (SVOp->getMask()[i] >= 0)
11541 NumV1OddIndices += i % 2;
11542 if (NumV2OddIndices < NumV1OddIndices)
11543 return DAG.getCommutedVectorShuffle(*SVOp);
11548 // For each vector width, delegate to a specialized lowering routine.
11549 if (VT.getSizeInBits() == 128)
11550 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11552 if (VT.getSizeInBits() == 256)
11553 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11555 // Force AVX-512 vectors to be scalarized for now.
11556 // FIXME: Implement AVX-512 support!
11557 if (VT.getSizeInBits() == 512)
11558 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11560 llvm_unreachable("Unimplemented!");
11564 //===----------------------------------------------------------------------===//
11565 // Legacy vector shuffle lowering
11567 // This code is the legacy code handling vector shuffles until the above
11568 // replaces its functionality and performance.
11569 //===----------------------------------------------------------------------===//
11571 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
11572 bool hasInt256, unsigned *MaskOut = nullptr) {
11573 MVT EltVT = VT.getVectorElementType();
11575 // There is no blend with immediate in AVX-512.
11576 if (VT.is512BitVector())
11579 if (!hasSSE41 || EltVT == MVT::i8)
11581 if (!hasInt256 && VT == MVT::v16i16)
11584 unsigned MaskValue = 0;
11585 unsigned NumElems = VT.getVectorNumElements();
11586 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11587 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11588 unsigned NumElemsInLane = NumElems / NumLanes;
11590 // Blend for v16i16 should be symmetric for both lanes.
11591 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11593 int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
11594 int EltIdx = MaskVals[i];
11596 if ((EltIdx < 0 || EltIdx == (int)i) &&
11597 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
11600 if (((unsigned)EltIdx == (i + NumElems)) &&
11601 (SndLaneEltIdx < 0 ||
11602 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
11603 MaskValue |= (1 << i);
11609 *MaskOut = MaskValue;
11613 // Try to lower a shuffle node into a simple blend instruction.
11614 // This function assumes isBlendMask returns true for this
11615 // SuffleVectorSDNode
11616 static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
11617 unsigned MaskValue,
11618 const X86Subtarget *Subtarget,
11619 SelectionDAG &DAG) {
11620 MVT VT = SVOp->getSimpleValueType(0);
11621 MVT EltVT = VT.getVectorElementType();
11622 assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
11623 Subtarget->hasInt256() && "Trying to lower a "
11624 "VECTOR_SHUFFLE to a Blend but "
11625 "with the wrong mask"));
11626 SDValue V1 = SVOp->getOperand(0);
11627 SDValue V2 = SVOp->getOperand(1);
11629 unsigned NumElems = VT.getVectorNumElements();
11631 // Convert i32 vectors to floating point if it is not AVX2.
11632 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
11634 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
11635 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
11637 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
11638 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
11641 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
11642 DAG.getConstant(MaskValue, MVT::i32));
11643 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
11646 /// In vector type \p VT, return true if the element at index \p InputIdx
11647 /// falls on a different 128-bit lane than \p OutputIdx.
11648 static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
11649 unsigned OutputIdx) {
11650 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
11651 return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
11654 /// Generate a PSHUFB if possible. Selects elements from \p V1 according to
11655 /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
11656 /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
11657 /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
11659 static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
11660 SelectionDAG &DAG) {
11661 MVT VT = V1.getSimpleValueType();
11662 assert(VT.is128BitVector() || VT.is256BitVector());
11664 MVT EltVT = VT.getVectorElementType();
11665 unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
11666 unsigned NumElts = VT.getVectorNumElements();
11668 SmallVector<SDValue, 32> PshufbMask;
11669 for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
11670 int InputIdx = MaskVals[OutputIdx];
11671 unsigned InputByteIdx;
11673 if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
11674 InputByteIdx = 0x80;
11676 // Cross lane is not allowed.
11677 if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
11679 InputByteIdx = InputIdx * EltSizeInBytes;
11680 // Index is an byte offset within the 128-bit lane.
11681 InputByteIdx &= 0xf;
11684 for (unsigned j = 0; j < EltSizeInBytes; ++j) {
11685 PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
11686 if (InputByteIdx != 0x80)
11691 MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
11693 V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
11694 return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
11695 DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
11698 // v8i16 shuffles - Prefer shuffles in the following order:
11699 // 1. [all] pshuflw, pshufhw, optional move
11700 // 2. [ssse3] 1 x pshufb
11701 // 3. [ssse3] 2 x pshufb + 1 x por
11702 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
11704 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
11705 SelectionDAG &DAG) {
11706 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11707 SDValue V1 = SVOp->getOperand(0);
11708 SDValue V2 = SVOp->getOperand(1);
11710 SmallVector<int, 8> MaskVals;
11712 // Determine if more than 1 of the words in each of the low and high quadwords
11713 // of the result come from the same quadword of one of the two inputs. Undef
11714 // mask values count as coming from any quadword, for better codegen.
11716 // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
11717 // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
11718 unsigned LoQuad[] = { 0, 0, 0, 0 };
11719 unsigned HiQuad[] = { 0, 0, 0, 0 };
11720 // Indices of quads used.
11721 std::bitset<4> InputQuads;
11722 for (unsigned i = 0; i < 8; ++i) {
11723 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
11724 int EltIdx = SVOp->getMaskElt(i);
11725 MaskVals.push_back(EltIdx);
11733 ++Quad[EltIdx / 4];
11734 InputQuads.set(EltIdx / 4);
11737 int BestLoQuad = -1;
11738 unsigned MaxQuad = 1;
11739 for (unsigned i = 0; i < 4; ++i) {
11740 if (LoQuad[i] > MaxQuad) {
11742 MaxQuad = LoQuad[i];
11746 int BestHiQuad = -1;
11748 for (unsigned i = 0; i < 4; ++i) {
11749 if (HiQuad[i] > MaxQuad) {
11751 MaxQuad = HiQuad[i];
11755 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
11756 // of the two input vectors, shuffle them into one input vector so only a
11757 // single pshufb instruction is necessary. If there are more than 2 input
11758 // quads, disable the next transformation since it does not help SSSE3.
11759 bool V1Used = InputQuads[0] || InputQuads[1];
11760 bool V2Used = InputQuads[2] || InputQuads[3];
11761 if (Subtarget->hasSSSE3()) {
11762 if (InputQuads.count() == 2 && V1Used && V2Used) {
11763 BestLoQuad = InputQuads[0] ? 0 : 1;
11764 BestHiQuad = InputQuads[2] ? 2 : 3;
11766 if (InputQuads.count() > 2) {
11772 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
11773 // the shuffle mask. If a quad is scored as -1, that means that it contains
11774 // words from all 4 input quadwords.
11776 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
11778 BestLoQuad < 0 ? 0 : BestLoQuad,
11779 BestHiQuad < 0 ? 1 : BestHiQuad
11781 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
11782 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
11783 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
11784 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
11786 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
11787 // source words for the shuffle, to aid later transformations.
11788 bool AllWordsInNewV = true;
11789 bool InOrder[2] = { true, true };
11790 for (unsigned i = 0; i != 8; ++i) {
11791 int idx = MaskVals[i];
11793 InOrder[i/4] = false;
11794 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
11796 AllWordsInNewV = false;
11800 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
11801 if (AllWordsInNewV) {
11802 for (int i = 0; i != 8; ++i) {
11803 int idx = MaskVals[i];
11806 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
11807 if ((idx != i) && idx < 4)
11809 if ((idx != i) && idx > 3)
11818 // If we've eliminated the use of V2, and the new mask is a pshuflw or
11819 // pshufhw, that's as cheap as it gets. Return the new shuffle.
11820 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
11821 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
11822 unsigned TargetMask = 0;
11823 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
11824 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
11825 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11826 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
11827 getShufflePSHUFLWImmediate(SVOp);
11828 V1 = NewV.getOperand(0);
11829 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
11833 // Promote splats to a larger type which usually leads to more efficient code.
11834 // FIXME: Is this true if pshufb is available?
11835 if (SVOp->isSplat())
11836 return PromoteSplat(SVOp, DAG);
11838 // If we have SSSE3, and all words of the result are from 1 input vector,
11839 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
11840 // is present, fall back to case 4.
11841 if (Subtarget->hasSSSE3()) {
11842 SmallVector<SDValue,16> pshufbMask;
11844 // If we have elements from both input vectors, set the high bit of the
11845 // shuffle mask element to zero out elements that come from V2 in the V1
11846 // mask, and elements that come from V1 in the V2 mask, so that the two
11847 // results can be OR'd together.
11848 bool TwoInputs = V1Used && V2Used;
11849 V1 = getPSHUFB(MaskVals, V1, dl, DAG);
11851 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11853 // Calculate the shuffle mask for the second input, shuffle it, and
11854 // OR it with the first shuffled input.
11855 CommuteVectorShuffleMask(MaskVals, 8);
11856 V2 = getPSHUFB(MaskVals, V2, dl, DAG);
11857 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
11858 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11861 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
11862 // and update MaskVals with new element order.
11863 std::bitset<8> InOrder;
11864 if (BestLoQuad >= 0) {
11865 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
11866 for (int i = 0; i != 4; ++i) {
11867 int idx = MaskVals[i];
11870 } else if ((idx / 4) == BestLoQuad) {
11871 MaskV[i] = idx & 3;
11875 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11878 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11879 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11880 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
11881 NewV.getOperand(0),
11882 getShufflePSHUFLWImmediate(SVOp), DAG);
11886 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
11887 // and update MaskVals with the new element order.
11888 if (BestHiQuad >= 0) {
11889 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
11890 for (unsigned i = 4; i != 8; ++i) {
11891 int idx = MaskVals[i];
11894 } else if ((idx / 4) == BestHiQuad) {
11895 MaskV[i] = (idx & 3) + 4;
11899 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11902 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11903 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11904 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
11905 NewV.getOperand(0),
11906 getShufflePSHUFHWImmediate(SVOp), DAG);
11910 // In case BestHi & BestLo were both -1, which means each quadword has a word
11911 // from each of the four input quadwords, calculate the InOrder bitvector now
11912 // before falling through to the insert/extract cleanup.
11913 if (BestLoQuad == -1 && BestHiQuad == -1) {
11915 for (int i = 0; i != 8; ++i)
11916 if (MaskVals[i] < 0 || MaskVals[i] == i)
11920 // The other elements are put in the right place using pextrw and pinsrw.
11921 for (unsigned i = 0; i != 8; ++i) {
11924 int EltIdx = MaskVals[i];
11927 SDValue ExtOp = (EltIdx < 8) ?
11928 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
11929 DAG.getIntPtrConstant(EltIdx)) :
11930 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
11931 DAG.getIntPtrConstant(EltIdx - 8));
11932 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
11933 DAG.getIntPtrConstant(i));
11938 /// \brief v16i16 shuffles
11940 /// FIXME: We only support generation of a single pshufb currently. We can
11941 /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
11942 /// well (e.g 2 x pshufb + 1 x por).
11944 LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
11945 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11946 SDValue V1 = SVOp->getOperand(0);
11947 SDValue V2 = SVOp->getOperand(1);
11950 if (V2.getOpcode() != ISD::UNDEF)
11953 SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
11954 return getPSHUFB(MaskVals, V1, dl, DAG);
11957 // v16i8 shuffles - Prefer shuffles in the following order:
11958 // 1. [ssse3] 1 x pshufb
11959 // 2. [ssse3] 2 x pshufb + 1 x por
11960 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
11961 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
11962 const X86Subtarget* Subtarget,
11963 SelectionDAG &DAG) {
11964 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11965 SDValue V1 = SVOp->getOperand(0);
11966 SDValue V2 = SVOp->getOperand(1);
11968 ArrayRef<int> MaskVals = SVOp->getMask();
11970 // Promote splats to a larger type which usually leads to more efficient code.
11971 // FIXME: Is this true if pshufb is available?
11972 if (SVOp->isSplat())
11973 return PromoteSplat(SVOp, DAG);
11975 // If we have SSSE3, case 1 is generated when all result bytes come from
11976 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
11977 // present, fall back to case 3.
11979 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
11980 if (Subtarget->hasSSSE3()) {
11981 SmallVector<SDValue,16> pshufbMask;
11983 // If all result elements are from one input vector, then only translate
11984 // undef mask values to 0x80 (zero out result) in the pshufb mask.
11986 // Otherwise, we have elements from both input vectors, and must zero out
11987 // elements that come from V2 in the first mask, and V1 in the second mask
11988 // so that we can OR them together.
11989 for (unsigned i = 0; i != 16; ++i) {
11990 int EltIdx = MaskVals[i];
11991 if (EltIdx < 0 || EltIdx >= 16)
11993 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
11995 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
11996 DAG.getNode(ISD::BUILD_VECTOR, dl,
11997 MVT::v16i8, pshufbMask));
11999 // As PSHUFB will zero elements with negative indices, it's safe to ignore
12000 // the 2nd operand if it's undefined or zero.
12001 if (V2.getOpcode() == ISD::UNDEF ||
12002 ISD::isBuildVectorAllZeros(V2.getNode()))
12005 // Calculate the shuffle mask for the second input, shuffle it, and
12006 // OR it with the first shuffled input.
12007 pshufbMask.clear();
12008 for (unsigned i = 0; i != 16; ++i) {
12009 int EltIdx = MaskVals[i];
12010 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
12011 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
12013 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
12014 DAG.getNode(ISD::BUILD_VECTOR, dl,
12015 MVT::v16i8, pshufbMask));
12016 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
12019 // No SSSE3 - Calculate in place words and then fix all out of place words
12020 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
12021 // the 16 different words that comprise the two doublequadword input vectors.
12022 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
12023 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
12025 for (int i = 0; i != 8; ++i) {
12026 int Elt0 = MaskVals[i*2];
12027 int Elt1 = MaskVals[i*2+1];
12029 // This word of the result is all undef, skip it.
12030 if (Elt0 < 0 && Elt1 < 0)
12033 // This word of the result is already in the correct place, skip it.
12034 if ((Elt0 == i*2) && (Elt1 == i*2+1))
12037 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
12038 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
12041 // If Elt0 and Elt1 are defined, are consecutive, and can be load
12042 // using a single extract together, load it and store it.
12043 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
12044 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
12045 DAG.getIntPtrConstant(Elt1 / 2));
12046 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
12047 DAG.getIntPtrConstant(i));
12051 // If Elt1 is defined, extract it from the appropriate source. If the
12052 // source byte is not also odd, shift the extracted word left 8 bits
12053 // otherwise clear the bottom 8 bits if we need to do an or.
12055 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
12056 DAG.getIntPtrConstant(Elt1 / 2));
12057 if ((Elt1 & 1) == 0)
12058 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
12060 TLI.getShiftAmountTy(InsElt.getValueType())));
12061 else if (Elt0 >= 0)
12062 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
12063 DAG.getConstant(0xFF00, MVT::i16));
12065 // If Elt0 is defined, extract it from the appropriate source. If the
12066 // source byte is not also even, shift the extracted word right 8 bits. If
12067 // Elt1 was also defined, OR the extracted values together before
12068 // inserting them in the result.
12070 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
12071 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
12072 if ((Elt0 & 1) != 0)
12073 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
12075 TLI.getShiftAmountTy(InsElt0.getValueType())));
12076 else if (Elt1 >= 0)
12077 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
12078 DAG.getConstant(0x00FF, MVT::i16));
12079 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
12082 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
12083 DAG.getIntPtrConstant(i));
12085 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
12088 // v32i8 shuffles - Translate to VPSHUFB if possible.
12090 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
12091 const X86Subtarget *Subtarget,
12092 SelectionDAG &DAG) {
12093 MVT VT = SVOp->getSimpleValueType(0);
12094 SDValue V1 = SVOp->getOperand(0);
12095 SDValue V2 = SVOp->getOperand(1);
12097 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
12099 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
12100 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
12101 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
12103 // VPSHUFB may be generated if
12104 // (1) one of input vector is undefined or zeroinitializer.
12105 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
12106 // And (2) the mask indexes don't cross the 128-bit lane.
12107 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
12108 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
12111 if (V1IsAllZero && !V2IsAllZero) {
12112 CommuteVectorShuffleMask(MaskVals, 32);
12115 return getPSHUFB(MaskVals, V1, dl, DAG);
12118 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
12119 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
12120 /// done when every pair / quad of shuffle mask elements point to elements in
12121 /// the right sequence. e.g.
12122 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
12124 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
12125 SelectionDAG &DAG) {
12126 MVT VT = SVOp->getSimpleValueType(0);
12128 unsigned NumElems = VT.getVectorNumElements();
12131 switch (VT.SimpleTy) {
12132 default: llvm_unreachable("Unexpected!");
12135 return SDValue(SVOp, 0);
12136 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
12137 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
12138 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
12139 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
12140 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
12141 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
12144 SmallVector<int, 8> MaskVec;
12145 for (unsigned i = 0; i != NumElems; i += Scale) {
12147 for (unsigned j = 0; j != Scale; ++j) {
12148 int EltIdx = SVOp->getMaskElt(i+j);
12152 StartIdx = (EltIdx / Scale);
12153 if (EltIdx != (int)(StartIdx*Scale + j))
12156 MaskVec.push_back(StartIdx);
12159 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
12160 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
12161 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
12164 /// getVZextMovL - Return a zero-extending vector move low node.
12166 static SDValue getVZextMovL(MVT VT, MVT OpVT,
12167 SDValue SrcOp, SelectionDAG &DAG,
12168 const X86Subtarget *Subtarget, SDLoc dl) {
12169 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
12170 LoadSDNode *LD = nullptr;
12171 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
12172 LD = dyn_cast<LoadSDNode>(SrcOp);
12174 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
12176 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
12177 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
12178 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
12179 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
12180 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
12182 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
12183 return DAG.getNode(ISD::BITCAST, dl, VT,
12184 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
12185 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
12187 SrcOp.getOperand(0)
12193 return DAG.getNode(ISD::BITCAST, dl, VT,
12194 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
12195 DAG.getNode(ISD::BITCAST, dl,
12199 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
12200 /// which could not be matched by any known target speficic shuffle
12202 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
12204 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
12205 if (NewOp.getNode())
12208 MVT VT = SVOp->getSimpleValueType(0);
12210 unsigned NumElems = VT.getVectorNumElements();
12211 unsigned NumLaneElems = NumElems / 2;
12214 MVT EltVT = VT.getVectorElementType();
12215 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
12218 SmallVector<int, 16> Mask;
12219 for (unsigned l = 0; l < 2; ++l) {
12220 // Build a shuffle mask for the output, discovering on the fly which
12221 // input vectors to use as shuffle operands (recorded in InputUsed).
12222 // If building a suitable shuffle vector proves too hard, then bail
12223 // out with UseBuildVector set.
12224 bool UseBuildVector = false;
12225 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
12226 unsigned LaneStart = l * NumLaneElems;
12227 for (unsigned i = 0; i != NumLaneElems; ++i) {
12228 // The mask element. This indexes into the input.
12229 int Idx = SVOp->getMaskElt(i+LaneStart);
12231 // the mask element does not index into any input vector.
12232 Mask.push_back(-1);
12236 // The input vector this mask element indexes into.
12237 int Input = Idx / NumLaneElems;
12239 // Turn the index into an offset from the start of the input vector.
12240 Idx -= Input * NumLaneElems;
12242 // Find or create a shuffle vector operand to hold this input.
12244 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
12245 if (InputUsed[OpNo] == Input)
12246 // This input vector is already an operand.
12248 if (InputUsed[OpNo] < 0) {
12249 // Create a new operand for this input vector.
12250 InputUsed[OpNo] = Input;
12255 if (OpNo >= array_lengthof(InputUsed)) {
12256 // More than two input vectors used! Give up on trying to create a
12257 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
12258 UseBuildVector = true;
12262 // Add the mask index for the new shuffle vector.
12263 Mask.push_back(Idx + OpNo * NumLaneElems);
12266 if (UseBuildVector) {
12267 SmallVector<SDValue, 16> SVOps;
12268 for (unsigned i = 0; i != NumLaneElems; ++i) {
12269 // The mask element. This indexes into the input.
12270 int Idx = SVOp->getMaskElt(i+LaneStart);
12272 SVOps.push_back(DAG.getUNDEF(EltVT));
12276 // The input vector this mask element indexes into.
12277 int Input = Idx / NumElems;
12279 // Turn the index into an offset from the start of the input vector.
12280 Idx -= Input * NumElems;
12282 // Extract the vector element by hand.
12283 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
12284 SVOp->getOperand(Input),
12285 DAG.getIntPtrConstant(Idx)));
12288 // Construct the output using a BUILD_VECTOR.
12289 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
12290 } else if (InputUsed[0] < 0) {
12291 // No input vectors were used! The result is undefined.
12292 Output[l] = DAG.getUNDEF(NVT);
12294 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
12295 (InputUsed[0] % 2) * NumLaneElems,
12297 // If only one input was used, use an undefined vector for the other.
12298 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
12299 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
12300 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
12301 // At least one input vector was used. Create a new shuffle vector.
12302 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
12308 // Concatenate the result back
12309 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
12312 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
12313 /// 4 elements, and match them with several different shuffle types.
12315 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
12316 SDValue V1 = SVOp->getOperand(0);
12317 SDValue V2 = SVOp->getOperand(1);
12319 MVT VT = SVOp->getSimpleValueType(0);
12321 assert(VT.is128BitVector() && "Unsupported vector size");
12323 std::pair<int, int> Locs[4];
12324 int Mask1[] = { -1, -1, -1, -1 };
12325 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
12327 unsigned NumHi = 0;
12328 unsigned NumLo = 0;
12329 for (unsigned i = 0; i != 4; ++i) {
12330 int Idx = PermMask[i];
12332 Locs[i] = std::make_pair(-1, -1);
12334 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
12336 Locs[i] = std::make_pair(0, NumLo);
12337 Mask1[NumLo] = Idx;
12340 Locs[i] = std::make_pair(1, NumHi);
12342 Mask1[2+NumHi] = Idx;
12348 if (NumLo <= 2 && NumHi <= 2) {
12349 // If no more than two elements come from either vector. This can be
12350 // implemented with two shuffles. First shuffle gather the elements.
12351 // The second shuffle, which takes the first shuffle as both of its
12352 // vector operands, put the elements into the right order.
12353 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12355 int Mask2[] = { -1, -1, -1, -1 };
12357 for (unsigned i = 0; i != 4; ++i)
12358 if (Locs[i].first != -1) {
12359 unsigned Idx = (i < 2) ? 0 : 4;
12360 Idx += Locs[i].first * 2 + Locs[i].second;
12364 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
12367 if (NumLo == 3 || NumHi == 3) {
12368 // Otherwise, we must have three elements from one vector, call it X, and
12369 // one element from the other, call it Y. First, use a shufps to build an
12370 // intermediate vector with the one element from Y and the element from X
12371 // that will be in the same half in the final destination (the indexes don't
12372 // matter). Then, use a shufps to build the final vector, taking the half
12373 // containing the element from Y from the intermediate, and the other half
12376 // Normalize it so the 3 elements come from V1.
12377 CommuteVectorShuffleMask(PermMask, 4);
12381 // Find the element from V2.
12383 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
12384 int Val = PermMask[HiIndex];
12391 Mask1[0] = PermMask[HiIndex];
12393 Mask1[2] = PermMask[HiIndex^1];
12395 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12397 if (HiIndex >= 2) {
12398 Mask1[0] = PermMask[0];
12399 Mask1[1] = PermMask[1];
12400 Mask1[2] = HiIndex & 1 ? 6 : 4;
12401 Mask1[3] = HiIndex & 1 ? 4 : 6;
12402 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12405 Mask1[0] = HiIndex & 1 ? 2 : 0;
12406 Mask1[1] = HiIndex & 1 ? 0 : 2;
12407 Mask1[2] = PermMask[2];
12408 Mask1[3] = PermMask[3];
12413 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
12416 // Break it into (shuffle shuffle_hi, shuffle_lo).
12417 int LoMask[] = { -1, -1, -1, -1 };
12418 int HiMask[] = { -1, -1, -1, -1 };
12420 int *MaskPtr = LoMask;
12421 unsigned MaskIdx = 0;
12422 unsigned LoIdx = 0;
12423 unsigned HiIdx = 2;
12424 for (unsigned i = 0; i != 4; ++i) {
12431 int Idx = PermMask[i];
12433 Locs[i] = std::make_pair(-1, -1);
12434 } else if (Idx < 4) {
12435 Locs[i] = std::make_pair(MaskIdx, LoIdx);
12436 MaskPtr[LoIdx] = Idx;
12439 Locs[i] = std::make_pair(MaskIdx, HiIdx);
12440 MaskPtr[HiIdx] = Idx;
12445 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
12446 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
12447 int MaskOps[] = { -1, -1, -1, -1 };
12448 for (unsigned i = 0; i != 4; ++i)
12449 if (Locs[i].first != -1)
12450 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
12451 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
12454 static bool MayFoldVectorLoad(SDValue V) {
12455 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
12456 V = V.getOperand(0);
12458 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
12459 V = V.getOperand(0);
12460 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
12461 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
12462 // BUILD_VECTOR (load), undef
12463 V = V.getOperand(0);
12465 return MayFoldLoad(V);
12469 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
12470 MVT VT = Op.getSimpleValueType();
12472 // Canonicalize to v2f64.
12473 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
12474 return DAG.getNode(ISD::BITCAST, dl, VT,
12475 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
12480 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
12482 SDValue V1 = Op.getOperand(0);
12483 SDValue V2 = Op.getOperand(1);
12484 MVT VT = Op.getSimpleValueType();
12486 assert(VT != MVT::v2i64 && "unsupported shuffle type");
12488 if (HasSSE2 && VT == MVT::v2f64)
12489 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
12491 // v4f32 or v4i32: canonicalize to v4f32 (which is legal for SSE1)
12492 return DAG.getNode(ISD::BITCAST, dl, VT,
12493 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
12494 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
12495 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
12499 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
12500 SDValue V1 = Op.getOperand(0);
12501 SDValue V2 = Op.getOperand(1);
12502 MVT VT = Op.getSimpleValueType();
12504 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
12505 "unsupported shuffle type");
12507 if (V2.getOpcode() == ISD::UNDEF)
12511 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
12515 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
12516 SDValue V1 = Op.getOperand(0);
12517 SDValue V2 = Op.getOperand(1);
12518 MVT VT = Op.getSimpleValueType();
12519 unsigned NumElems = VT.getVectorNumElements();
12521 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
12522 // operand of these instructions is only memory, so check if there's a
12523 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
12525 bool CanFoldLoad = false;
12527 // Trivial case, when V2 comes from a load.
12528 if (MayFoldVectorLoad(V2))
12529 CanFoldLoad = true;
12531 // When V1 is a load, it can be folded later into a store in isel, example:
12532 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
12534 // (MOVLPSmr addr:$src1, VR128:$src2)
12535 // So, recognize this potential and also use MOVLPS or MOVLPD
12536 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
12537 CanFoldLoad = true;
12539 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12541 if (HasSSE2 && NumElems == 2)
12542 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
12545 // If we don't care about the second element, proceed to use movss.
12546 if (SVOp->getMaskElt(1) != -1)
12547 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
12550 // movl and movlp will both match v2i64, but v2i64 is never matched by
12551 // movl earlier because we make it strict to avoid messing with the movlp load
12552 // folding logic (see the code above getMOVLP call). Match it here then,
12553 // this is horrible, but will stay like this until we move all shuffle
12554 // matching to x86 specific nodes. Note that for the 1st condition all
12555 // types are matched with movsd.
12557 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
12558 // as to remove this logic from here, as much as possible
12559 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
12560 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12561 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12564 assert(VT != MVT::v4i32 && "unsupported shuffle type");
12566 // Invert the operand order and use SHUFPS to match it.
12567 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
12568 getShuffleSHUFImmediate(SVOp), DAG);
12571 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
12572 SelectionDAG &DAG) {
12574 MVT VT = Load->getSimpleValueType(0);
12575 MVT EVT = VT.getVectorElementType();
12576 SDValue Addr = Load->getOperand(1);
12577 SDValue NewAddr = DAG.getNode(
12578 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
12579 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
12582 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
12583 DAG.getMachineFunction().getMachineMemOperand(
12584 Load->getMemOperand(), 0, EVT.getStoreSize()));
12588 // It is only safe to call this function if isINSERTPSMask is true for
12589 // this shufflevector mask.
12590 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
12591 SelectionDAG &DAG) {
12592 // Generate an insertps instruction when inserting an f32 from memory onto a
12593 // v4f32 or when copying a member from one v4f32 to another.
12594 // We also use it for transferring i32 from one register to another,
12595 // since it simply copies the same bits.
12596 // If we're transferring an i32 from memory to a specific element in a
12597 // register, we output a generic DAG that will match the PINSRD
12599 MVT VT = SVOp->getSimpleValueType(0);
12600 MVT EVT = VT.getVectorElementType();
12601 SDValue V1 = SVOp->getOperand(0);
12602 SDValue V2 = SVOp->getOperand(1);
12603 auto Mask = SVOp->getMask();
12604 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
12605 "unsupported vector type for insertps/pinsrd");
12607 auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
12608 auto FromV2Predicate = [](const int &i) { return i >= 4; };
12609 int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
12613 unsigned DestIndex;
12617 DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
12620 // If we have 1 element from each vector, we have to check if we're
12621 // changing V1's element's place. If so, we're done. Otherwise, we
12622 // should assume we're changing V2's element's place and behave
12624 int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
12625 assert(DestIndex <= INT32_MAX && "truncated destination index");
12626 if (FromV1 == FromV2 &&
12627 static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
12631 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12634 assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
12635 "More than one element from V1 and from V2, or no elements from one "
12636 "of the vectors. This case should not have returned true from "
12641 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12644 // Get an index into the source vector in the range [0,4) (the mask is
12645 // in the range [0,8) because it can address V1 and V2)
12646 unsigned SrcIndex = Mask[DestIndex] % 4;
12647 if (MayFoldLoad(From)) {
12648 // Trivial case, when From comes from a load and is only used by the
12649 // shuffle. Make it use insertps from the vector that we need from that
12652 NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
12653 if (!NewLoad.getNode())
12656 if (EVT == MVT::f32) {
12657 // Create this as a scalar to vector to match the instruction pattern.
12658 SDValue LoadScalarToVector =
12659 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
12660 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
12661 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
12663 } else { // EVT == MVT::i32
12664 // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
12665 // instruction, to match the PINSRD instruction, which loads an i32 to a
12666 // certain vector element.
12667 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
12668 DAG.getConstant(DestIndex, MVT::i32));
12672 // Vector-element-to-vector
12673 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
12674 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
12677 // Reduce a vector shuffle to zext.
12678 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
12679 SelectionDAG &DAG) {
12680 // PMOVZX is only available from SSE41.
12681 if (!Subtarget->hasSSE41())
12684 MVT VT = Op.getSimpleValueType();
12686 // Only AVX2 support 256-bit vector integer extending.
12687 if (!Subtarget->hasInt256() && VT.is256BitVector())
12690 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12692 SDValue V1 = Op.getOperand(0);
12693 SDValue V2 = Op.getOperand(1);
12694 unsigned NumElems = VT.getVectorNumElements();
12696 // Extending is an unary operation and the element type of the source vector
12697 // won't be equal to or larger than i64.
12698 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
12699 VT.getVectorElementType() == MVT::i64)
12702 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
12703 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
12704 while ((1U << Shift) < NumElems) {
12705 if (SVOp->getMaskElt(1U << Shift) == 1)
12708 // The maximal ratio is 8, i.e. from i8 to i64.
12713 // Check the shuffle mask.
12714 unsigned Mask = (1U << Shift) - 1;
12715 for (unsigned i = 0; i != NumElems; ++i) {
12716 int EltIdx = SVOp->getMaskElt(i);
12717 if ((i & Mask) != 0 && EltIdx != -1)
12719 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
12723 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
12724 MVT NeVT = MVT::getIntegerVT(NBits);
12725 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
12727 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
12730 return DAG.getNode(ISD::BITCAST, DL, VT,
12731 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
12734 static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
12735 SelectionDAG &DAG) {
12736 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12737 MVT VT = Op.getSimpleValueType();
12739 SDValue V1 = Op.getOperand(0);
12740 SDValue V2 = Op.getOperand(1);
12742 if (isZeroShuffle(SVOp))
12743 return getZeroVector(VT, Subtarget, DAG, dl);
12745 // Handle splat operations
12746 if (SVOp->isSplat()) {
12747 // Use vbroadcast whenever the splat comes from a foldable load
12748 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
12749 if (Broadcast.getNode())
12753 // Check integer expanding shuffles.
12754 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
12755 if (NewOp.getNode())
12758 // If the shuffle can be profitably rewritten as a narrower shuffle, then
12760 if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
12761 VT == MVT::v32i8) {
12762 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12763 if (NewOp.getNode())
12764 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
12765 } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
12766 // FIXME: Figure out a cleaner way to do this.
12767 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
12768 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12769 if (NewOp.getNode()) {
12770 MVT NewVT = NewOp.getSimpleValueType();
12771 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
12772 NewVT, true, false))
12773 return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
12776 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
12777 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12778 if (NewOp.getNode()) {
12779 MVT NewVT = NewOp.getSimpleValueType();
12780 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
12781 return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
12790 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
12791 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12792 SDValue V1 = Op.getOperand(0);
12793 SDValue V2 = Op.getOperand(1);
12794 MVT VT = Op.getSimpleValueType();
12796 unsigned NumElems = VT.getVectorNumElements();
12797 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
12798 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
12799 bool V1IsSplat = false;
12800 bool V2IsSplat = false;
12801 bool HasSSE2 = Subtarget->hasSSE2();
12802 bool HasFp256 = Subtarget->hasFp256();
12803 bool HasInt256 = Subtarget->hasInt256();
12804 MachineFunction &MF = DAG.getMachineFunction();
12806 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
12808 // Check if we should use the experimental vector shuffle lowering. If so,
12809 // delegate completely to that code path.
12810 if (ExperimentalVectorShuffleLowering)
12811 return lowerVectorShuffle(Op, Subtarget, DAG);
12813 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
12815 if (V1IsUndef && V2IsUndef)
12816 return DAG.getUNDEF(VT);
12818 // When we create a shuffle node we put the UNDEF node to second operand,
12819 // but in some cases the first operand may be transformed to UNDEF.
12820 // In this case we should just commute the node.
12822 return DAG.getCommutedVectorShuffle(*SVOp);
12824 // Vector shuffle lowering takes 3 steps:
12826 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
12827 // narrowing and commutation of operands should be handled.
12828 // 2) Matching of shuffles with known shuffle masks to x86 target specific
12830 // 3) Rewriting of unmatched masks into new generic shuffle operations,
12831 // so the shuffle can be broken into other shuffles and the legalizer can
12832 // try the lowering again.
12834 // The general idea is that no vector_shuffle operation should be left to
12835 // be matched during isel, all of them must be converted to a target specific
12838 // Normalize the input vectors. Here splats, zeroed vectors, profitable
12839 // narrowing and commutation of operands should be handled. The actual code
12840 // doesn't include all of those, work in progress...
12841 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
12842 if (NewOp.getNode())
12845 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
12847 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
12848 // unpckh_undef). Only use pshufd if speed is more important than size.
12849 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
12850 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12851 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
12852 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12854 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
12855 V2IsUndef && MayFoldVectorLoad(V1))
12856 return getMOVDDup(Op, dl, V1, DAG);
12858 if (isMOVHLPS_v_undef_Mask(M, VT))
12859 return getMOVHighToLow(Op, dl, DAG);
12861 // Use to match splats
12862 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
12863 (VT == MVT::v2f64 || VT == MVT::v2i64))
12864 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12866 if (isPSHUFDMask(M, VT)) {
12867 // The actual implementation will match the mask in the if above and then
12868 // during isel it can match several different instructions, not only pshufd
12869 // as its name says, sad but true, emulate the behavior for now...
12870 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
12871 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
12873 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
12875 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
12876 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
12878 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
12879 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask,
12882 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
12886 if (isPALIGNRMask(M, VT, Subtarget))
12887 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
12888 getShufflePALIGNRImmediate(SVOp),
12891 if (isVALIGNMask(M, VT, Subtarget))
12892 return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2,
12893 getShuffleVALIGNImmediate(SVOp),
12896 // Check if this can be converted into a logical shift.
12897 bool isLeft = false;
12898 unsigned ShAmt = 0;
12900 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
12901 if (isShift && ShVal.hasOneUse()) {
12902 // If the shifted value has multiple uses, it may be cheaper to use
12903 // v_set0 + movlhps or movhlps, etc.
12904 MVT EltVT = VT.getVectorElementType();
12905 ShAmt *= EltVT.getSizeInBits();
12906 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12909 if (isMOVLMask(M, VT)) {
12910 if (ISD::isBuildVectorAllZeros(V1.getNode()))
12911 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
12912 if (!isMOVLPMask(M, VT)) {
12913 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
12914 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12916 if (VT == MVT::v4i32 || VT == MVT::v4f32)
12917 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12921 // FIXME: fold these into legal mask.
12922 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
12923 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
12925 if (isMOVHLPSMask(M, VT))
12926 return getMOVHighToLow(Op, dl, DAG);
12928 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
12929 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
12931 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
12932 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
12934 if (isMOVLPMask(M, VT))
12935 return getMOVLP(Op, dl, DAG, HasSSE2);
12937 if (ShouldXformToMOVHLPS(M, VT) ||
12938 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
12939 return DAG.getCommutedVectorShuffle(*SVOp);
12942 // No better options. Use a vshldq / vsrldq.
12943 MVT EltVT = VT.getVectorElementType();
12944 ShAmt *= EltVT.getSizeInBits();
12945 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12948 bool Commuted = false;
12949 // FIXME: This should also accept a bitcast of a splat? Be careful, not
12950 // 1,1,1,1 -> v8i16 though.
12951 BitVector UndefElements;
12952 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
12953 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12955 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
12956 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12959 // Canonicalize the splat or undef, if present, to be on the RHS.
12960 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
12961 CommuteVectorShuffleMask(M, NumElems);
12963 std::swap(V1IsSplat, V2IsSplat);
12967 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
12968 // Shuffling low element of v1 into undef, just return v1.
12971 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
12972 // the instruction selector will not match, so get a canonical MOVL with
12973 // swapped operands to undo the commute.
12974 return getMOVL(DAG, dl, VT, V2, V1);
12977 if (isUNPCKLMask(M, VT, HasInt256))
12978 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12980 if (isUNPCKHMask(M, VT, HasInt256))
12981 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12984 // Normalize mask so all entries that point to V2 points to its first
12985 // element then try to match unpck{h|l} again. If match, return a
12986 // new vector_shuffle with the corrected mask.p
12987 SmallVector<int, 8> NewMask(M.begin(), M.end());
12988 NormalizeMask(NewMask, NumElems);
12989 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
12990 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12991 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
12992 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12996 // Commute is back and try unpck* again.
12997 // FIXME: this seems wrong.
12998 CommuteVectorShuffleMask(M, NumElems);
13000 std::swap(V1IsSplat, V2IsSplat);
13002 if (isUNPCKLMask(M, VT, HasInt256))
13003 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
13005 if (isUNPCKHMask(M, VT, HasInt256))
13006 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
13009 // Normalize the node to match x86 shuffle ops if needed
13010 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
13011 return DAG.getCommutedVectorShuffle(*SVOp);
13013 // The checks below are all present in isShuffleMaskLegal, but they are
13014 // inlined here right now to enable us to directly emit target specific
13015 // nodes, and remove one by one until they don't return Op anymore.
13017 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
13018 SVOp->getSplatIndex() == 0 && V2IsUndef) {
13019 if (VT == MVT::v2f64 || VT == MVT::v2i64)
13020 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
13023 if (isPSHUFHWMask(M, VT, HasInt256))
13024 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
13025 getShufflePSHUFHWImmediate(SVOp),
13028 if (isPSHUFLWMask(M, VT, HasInt256))
13029 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
13030 getShufflePSHUFLWImmediate(SVOp),
13033 unsigned MaskValue;
13034 if (isBlendMask(M, VT, Subtarget->hasSSE41(), HasInt256, &MaskValue))
13035 return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
13037 if (isSHUFPMask(M, VT))
13038 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
13039 getShuffleSHUFImmediate(SVOp), DAG);
13041 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
13042 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
13043 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
13044 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
13046 //===--------------------------------------------------------------------===//
13047 // Generate target specific nodes for 128 or 256-bit shuffles only
13048 // supported in the AVX instruction set.
13051 // Handle VMOVDDUPY permutations
13052 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
13053 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
13055 // Handle VPERMILPS/D* permutations
13056 if (isVPERMILPMask(M, VT)) {
13057 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
13058 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
13059 getShuffleSHUFImmediate(SVOp), DAG);
13060 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1,
13061 getShuffleSHUFImmediate(SVOp), DAG);
13065 if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
13066 return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
13067 Idx*(NumElems/2), DAG, dl);
13069 // Handle VPERM2F128/VPERM2I128 permutations
13070 if (isVPERM2X128Mask(M, VT, HasFp256))
13071 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
13072 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
13074 if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
13075 return getINSERTPS(SVOp, dl, DAG);
13078 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
13079 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
13081 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
13082 VT.is512BitVector()) {
13083 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
13084 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
13085 SmallVector<SDValue, 16> permclMask;
13086 for (unsigned i = 0; i != NumElems; ++i) {
13087 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
13090 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
13092 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
13093 return DAG.getNode(X86ISD::VPERMV, dl, VT,
13094 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
13095 return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
13096 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
13099 //===--------------------------------------------------------------------===//
13100 // Since no target specific shuffle was selected for this generic one,
13101 // lower it into other known shuffles. FIXME: this isn't true yet, but
13102 // this is the plan.
13105 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
13106 if (VT == MVT::v8i16) {
13107 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
13108 if (NewOp.getNode())
13112 if (VT == MVT::v16i16 && HasInt256) {
13113 SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
13114 if (NewOp.getNode())
13118 if (VT == MVT::v16i8) {
13119 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
13120 if (NewOp.getNode())
13124 if (VT == MVT::v32i8) {
13125 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
13126 if (NewOp.getNode())
13130 // Handle all 128-bit wide vectors with 4 elements, and match them with
13131 // several different shuffle types.
13132 if (NumElems == 4 && VT.is128BitVector())
13133 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
13135 // Handle general 256-bit shuffles
13136 if (VT.is256BitVector())
13137 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
13142 // This function assumes its argument is a BUILD_VECTOR of constants or
13143 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
13145 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
13146 unsigned &MaskValue) {
13148 unsigned NumElems = BuildVector->getNumOperands();
13149 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
13150 unsigned NumLanes = (NumElems - 1) / 8 + 1;
13151 unsigned NumElemsInLane = NumElems / NumLanes;
13153 // Blend for v16i16 should be symetric for the both lanes.
13154 for (unsigned i = 0; i < NumElemsInLane; ++i) {
13155 SDValue EltCond = BuildVector->getOperand(i);
13156 SDValue SndLaneEltCond =
13157 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
13159 int Lane1Cond = -1, Lane2Cond = -1;
13160 if (isa<ConstantSDNode>(EltCond))
13161 Lane1Cond = !isZero(EltCond);
13162 if (isa<ConstantSDNode>(SndLaneEltCond))
13163 Lane2Cond = !isZero(SndLaneEltCond);
13165 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
13166 // Lane1Cond != 0, means we want the first argument.
13167 // Lane1Cond == 0, means we want the second argument.
13168 // The encoding of this argument is 0 for the first argument, 1
13169 // for the second. Therefore, invert the condition.
13170 MaskValue |= !Lane1Cond << i;
13171 else if (Lane1Cond < 0)
13172 MaskValue |= !Lane2Cond << i;
13179 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
13180 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
13181 const X86Subtarget *Subtarget,
13182 SelectionDAG &DAG) {
13183 SDValue Cond = Op.getOperand(0);
13184 SDValue LHS = Op.getOperand(1);
13185 SDValue RHS = Op.getOperand(2);
13187 MVT VT = Op.getSimpleValueType();
13189 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
13191 auto *CondBV = cast<BuildVectorSDNode>(Cond);
13193 // Only non-legal VSELECTs reach this lowering, convert those into generic
13194 // shuffles and re-use the shuffle lowering path for blends.
13195 SmallVector<int, 32> Mask;
13196 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
13197 SDValue CondElt = CondBV->getOperand(i);
13199 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
13201 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
13204 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
13205 // A vselect where all conditions and data are constants can be optimized into
13206 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
13207 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
13208 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
13209 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
13212 // Try to lower this to a blend-style vector shuffle. This can handle all
13213 // constant condition cases.
13214 SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG);
13215 if (BlendOp.getNode())
13218 // Variable blends are only legal from SSE4.1 onward.
13219 if (!Subtarget->hasSSE41())
13222 // Some types for vselect were previously set to Expand, not Legal or
13223 // Custom. Return an empty SDValue so we fall-through to Expand, after
13224 // the Custom lowering phase.
13225 MVT VT = Op.getSimpleValueType();
13226 switch (VT.SimpleTy) {
13231 if (Subtarget->hasBWI() && Subtarget->hasVLX())
13236 // We couldn't create a "Blend with immediate" node.
13237 // This node should still be legal, but we'll have to emit a blendv*
13242 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
13243 MVT VT = Op.getSimpleValueType();
13246 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
13249 if (VT.getSizeInBits() == 8) {
13250 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
13251 Op.getOperand(0), Op.getOperand(1));
13252 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
13253 DAG.getValueType(VT));
13254 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
13257 if (VT.getSizeInBits() == 16) {
13258 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13259 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
13261 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
13262 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13263 DAG.getNode(ISD::BITCAST, dl,
13266 Op.getOperand(1)));
13267 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
13268 Op.getOperand(0), Op.getOperand(1));
13269 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
13270 DAG.getValueType(VT));
13271 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
13274 if (VT == MVT::f32) {
13275 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
13276 // the result back to FR32 register. It's only worth matching if the
13277 // result has a single use which is a store or a bitcast to i32. And in
13278 // the case of a store, it's not worth it if the index is a constant 0,
13279 // because a MOVSSmr can be used instead, which is smaller and faster.
13280 if (!Op.hasOneUse())
13282 SDNode *User = *Op.getNode()->use_begin();
13283 if ((User->getOpcode() != ISD::STORE ||
13284 (isa<ConstantSDNode>(Op.getOperand(1)) &&
13285 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
13286 (User->getOpcode() != ISD::BITCAST ||
13287 User->getValueType(0) != MVT::i32))
13289 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13290 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
13293 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
13296 if (VT == MVT::i32 || VT == MVT::i64) {
13297 // ExtractPS/pextrq works with constant index.
13298 if (isa<ConstantSDNode>(Op.getOperand(1)))
13304 /// Extract one bit from mask vector, like v16i1 or v8i1.
13305 /// AVX-512 feature.
13307 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
13308 SDValue Vec = Op.getOperand(0);
13310 MVT VecVT = Vec.getSimpleValueType();
13311 SDValue Idx = Op.getOperand(1);
13312 MVT EltVT = Op.getSimpleValueType();
13314 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
13315 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
13316 "Unexpected vector type in ExtractBitFromMaskVector");
13318 // variable index can't be handled in mask registers,
13319 // extend vector to VR512
13320 if (!isa<ConstantSDNode>(Idx)) {
13321 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
13322 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
13323 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
13324 ExtVT.getVectorElementType(), Ext, Idx);
13325 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
13328 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13329 const TargetRegisterClass* rc = getRegClassFor(VecVT);
13330 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
13331 rc = getRegClassFor(MVT::v16i1);
13332 unsigned MaxSift = rc->getSize()*8 - 1;
13333 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
13334 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
13335 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
13336 DAG.getConstant(MaxSift, MVT::i8));
13337 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
13338 DAG.getIntPtrConstant(0));
13342 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
13343 SelectionDAG &DAG) const {
13345 SDValue Vec = Op.getOperand(0);
13346 MVT VecVT = Vec.getSimpleValueType();
13347 SDValue Idx = Op.getOperand(1);
13349 if (Op.getSimpleValueType() == MVT::i1)
13350 return ExtractBitFromMaskVector(Op, DAG);
13352 if (!isa<ConstantSDNode>(Idx)) {
13353 if (VecVT.is512BitVector() ||
13354 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
13355 VecVT.getVectorElementType().getSizeInBits() == 32)) {
13358 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
13359 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
13360 MaskEltVT.getSizeInBits());
13362 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
13363 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
13364 getZeroVector(MaskVT, Subtarget, DAG, dl),
13365 Idx, DAG.getConstant(0, getPointerTy()));
13366 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
13367 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
13368 Perm, DAG.getConstant(0, getPointerTy()));
13373 // If this is a 256-bit vector result, first extract the 128-bit vector and
13374 // then extract the element from the 128-bit vector.
13375 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
13377 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13378 // Get the 128-bit vector.
13379 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
13380 MVT EltVT = VecVT.getVectorElementType();
13382 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
13384 //if (IdxVal >= NumElems/2)
13385 // IdxVal -= NumElems/2;
13386 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
13387 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
13388 DAG.getConstant(IdxVal, MVT::i32));
13391 assert(VecVT.is128BitVector() && "Unexpected vector length");
13393 if (Subtarget->hasSSE41()) {
13394 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
13399 MVT VT = Op.getSimpleValueType();
13400 // TODO: handle v16i8.
13401 if (VT.getSizeInBits() == 16) {
13402 SDValue Vec = Op.getOperand(0);
13403 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13405 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
13406 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13407 DAG.getNode(ISD::BITCAST, dl,
13409 Op.getOperand(1)));
13410 // Transform it so it match pextrw which produces a 32-bit result.
13411 MVT EltVT = MVT::i32;
13412 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
13413 Op.getOperand(0), Op.getOperand(1));
13414 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
13415 DAG.getValueType(VT));
13416 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
13419 if (VT.getSizeInBits() == 32) {
13420 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13424 // SHUFPS the element to the lowest double word, then movss.
13425 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
13426 MVT VVT = Op.getOperand(0).getSimpleValueType();
13427 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
13428 DAG.getUNDEF(VVT), Mask);
13429 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
13430 DAG.getIntPtrConstant(0));
13433 if (VT.getSizeInBits() == 64) {
13434 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
13435 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
13436 // to match extract_elt for f64.
13437 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13441 // UNPCKHPD the element to the lowest double word, then movsd.
13442 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
13443 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
13444 int Mask[2] = { 1, -1 };
13445 MVT VVT = Op.getOperand(0).getSimpleValueType();
13446 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
13447 DAG.getUNDEF(VVT), Mask);
13448 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
13449 DAG.getIntPtrConstant(0));
13455 /// Insert one bit to mask vector, like v16i1 or v8i1.
13456 /// AVX-512 feature.
13458 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
13460 SDValue Vec = Op.getOperand(0);
13461 SDValue Elt = Op.getOperand(1);
13462 SDValue Idx = Op.getOperand(2);
13463 MVT VecVT = Vec.getSimpleValueType();
13465 if (!isa<ConstantSDNode>(Idx)) {
13466 // Non constant index. Extend source and destination,
13467 // insert element and then truncate the result.
13468 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
13469 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
13470 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
13471 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
13472 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
13473 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
13476 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13477 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
13478 if (Vec.getOpcode() == ISD::UNDEF)
13479 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
13480 DAG.getConstant(IdxVal, MVT::i8));
13481 const TargetRegisterClass* rc = getRegClassFor(VecVT);
13482 unsigned MaxSift = rc->getSize()*8 - 1;
13483 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
13484 DAG.getConstant(MaxSift, MVT::i8));
13485 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
13486 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
13487 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
13490 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
13491 SelectionDAG &DAG) const {
13492 MVT VT = Op.getSimpleValueType();
13493 MVT EltVT = VT.getVectorElementType();
13495 if (EltVT == MVT::i1)
13496 return InsertBitToMaskVector(Op, DAG);
13499 SDValue N0 = Op.getOperand(0);
13500 SDValue N1 = Op.getOperand(1);
13501 SDValue N2 = Op.getOperand(2);
13502 if (!isa<ConstantSDNode>(N2))
13504 auto *N2C = cast<ConstantSDNode>(N2);
13505 unsigned IdxVal = N2C->getZExtValue();
13507 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
13508 // into that, and then insert the subvector back into the result.
13509 if (VT.is256BitVector() || VT.is512BitVector()) {
13510 // Get the desired 128-bit vector half.
13511 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
13513 // Insert the element into the desired half.
13514 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
13515 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
13517 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
13518 DAG.getConstant(IdxIn128, MVT::i32));
13520 // Insert the changed part back to the 256-bit vector
13521 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
13523 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
13525 if (Subtarget->hasSSE41()) {
13526 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
13528 if (VT == MVT::v8i16) {
13529 Opc = X86ISD::PINSRW;
13531 assert(VT == MVT::v16i8);
13532 Opc = X86ISD::PINSRB;
13535 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
13537 if (N1.getValueType() != MVT::i32)
13538 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13539 if (N2.getValueType() != MVT::i32)
13540 N2 = DAG.getIntPtrConstant(IdxVal);
13541 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
13544 if (EltVT == MVT::f32) {
13545 // Bits [7:6] of the constant are the source select. This will always be
13546 // zero here. The DAG Combiner may combine an extract_elt index into
13548 // bits. For example (insert (extract, 3), 2) could be matched by
13550 // the '3' into bits [7:6] of X86ISD::INSERTPS.
13551 // Bits [5:4] of the constant are the destination select. This is the
13552 // value of the incoming immediate.
13553 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
13554 // combine either bitwise AND or insert of float 0.0 to set these bits.
13555 N2 = DAG.getIntPtrConstant(IdxVal << 4);
13556 // Create this as a scalar to vector..
13557 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
13558 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
13561 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
13562 // PINSR* works with constant index.
13567 if (EltVT == MVT::i8)
13570 if (EltVT.getSizeInBits() == 16) {
13571 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
13572 // as its second argument.
13573 if (N1.getValueType() != MVT::i32)
13574 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13575 if (N2.getValueType() != MVT::i32)
13576 N2 = DAG.getIntPtrConstant(IdxVal);
13577 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
13582 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
13584 MVT OpVT = Op.getSimpleValueType();
13586 // If this is a 256-bit vector result, first insert into a 128-bit
13587 // vector and then insert into the 256-bit vector.
13588 if (!OpVT.is128BitVector()) {
13589 // Insert into a 128-bit vector.
13590 unsigned SizeFactor = OpVT.getSizeInBits()/128;
13591 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
13592 OpVT.getVectorNumElements() / SizeFactor);
13594 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
13596 // Insert the 128-bit vector.
13597 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
13600 if (OpVT == MVT::v1i64 &&
13601 Op.getOperand(0).getValueType() == MVT::i64)
13602 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
13604 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
13605 assert(OpVT.is128BitVector() && "Expected an SSE type!");
13606 return DAG.getNode(ISD::BITCAST, dl, OpVT,
13607 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
13610 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
13611 // a simple subregister reference or explicit instructions to grab
13612 // upper bits of a vector.
13613 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13614 SelectionDAG &DAG) {
13616 SDValue In = Op.getOperand(0);
13617 SDValue Idx = Op.getOperand(1);
13618 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13619 MVT ResVT = Op.getSimpleValueType();
13620 MVT InVT = In.getSimpleValueType();
13622 if (Subtarget->hasFp256()) {
13623 if (ResVT.is128BitVector() &&
13624 (InVT.is256BitVector() || InVT.is512BitVector()) &&
13625 isa<ConstantSDNode>(Idx)) {
13626 return Extract128BitVector(In, IdxVal, DAG, dl);
13628 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
13629 isa<ConstantSDNode>(Idx)) {
13630 return Extract256BitVector(In, IdxVal, DAG, dl);
13636 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
13637 // simple superregister reference or explicit instructions to insert
13638 // the upper bits of a vector.
13639 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13640 SelectionDAG &DAG) {
13641 if (!Subtarget->hasAVX())
13645 SDValue Vec = Op.getOperand(0);
13646 SDValue SubVec = Op.getOperand(1);
13647 SDValue Idx = Op.getOperand(2);
13649 if (!isa<ConstantSDNode>(Idx))
13652 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13653 MVT OpVT = Op.getSimpleValueType();
13654 MVT SubVecVT = SubVec.getSimpleValueType();
13656 // Fold two 16-byte subvector loads into one 32-byte load:
13657 // (insert_subvector (insert_subvector undef, (load addr), 0),
13658 // (load addr + 16), Elts/2)
13660 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
13661 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
13662 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
13663 !Subtarget->isUnalignedMem32Slow()) {
13664 SDValue SubVec2 = Vec.getOperand(1);
13665 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
13666 if (Idx2->getZExtValue() == 0) {
13667 SDValue Ops[] = { SubVec2, SubVec };
13668 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
13675 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
13676 SubVecVT.is128BitVector())
13677 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
13679 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
13680 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
13685 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
13686 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
13687 // one of the above mentioned nodes. It has to be wrapped because otherwise
13688 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
13689 // be used to form addressing mode. These wrapped nodes will be selected
13692 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
13693 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
13695 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13696 // global base reg.
13697 unsigned char OpFlag = 0;
13698 unsigned WrapperKind = X86ISD::Wrapper;
13699 CodeModel::Model M = DAG.getTarget().getCodeModel();
13701 if (Subtarget->isPICStyleRIPRel() &&
13702 (M == CodeModel::Small || M == CodeModel::Kernel))
13703 WrapperKind = X86ISD::WrapperRIP;
13704 else if (Subtarget->isPICStyleGOT())
13705 OpFlag = X86II::MO_GOTOFF;
13706 else if (Subtarget->isPICStyleStubPIC())
13707 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13709 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
13710 CP->getAlignment(),
13711 CP->getOffset(), OpFlag);
13713 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13714 // With PIC, the address is actually $g + Offset.
13716 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13717 DAG.getNode(X86ISD::GlobalBaseReg,
13718 SDLoc(), getPointerTy()),
13725 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
13726 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
13728 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13729 // global base reg.
13730 unsigned char OpFlag = 0;
13731 unsigned WrapperKind = X86ISD::Wrapper;
13732 CodeModel::Model M = DAG.getTarget().getCodeModel();
13734 if (Subtarget->isPICStyleRIPRel() &&
13735 (M == CodeModel::Small || M == CodeModel::Kernel))
13736 WrapperKind = X86ISD::WrapperRIP;
13737 else if (Subtarget->isPICStyleGOT())
13738 OpFlag = X86II::MO_GOTOFF;
13739 else if (Subtarget->isPICStyleStubPIC())
13740 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13742 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
13745 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13747 // With PIC, the address is actually $g + Offset.
13749 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13750 DAG.getNode(X86ISD::GlobalBaseReg,
13751 SDLoc(), getPointerTy()),
13758 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
13759 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
13761 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13762 // global base reg.
13763 unsigned char OpFlag = 0;
13764 unsigned WrapperKind = X86ISD::Wrapper;
13765 CodeModel::Model M = DAG.getTarget().getCodeModel();
13767 if (Subtarget->isPICStyleRIPRel() &&
13768 (M == CodeModel::Small || M == CodeModel::Kernel)) {
13769 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
13770 OpFlag = X86II::MO_GOTPCREL;
13771 WrapperKind = X86ISD::WrapperRIP;
13772 } else if (Subtarget->isPICStyleGOT()) {
13773 OpFlag = X86II::MO_GOT;
13774 } else if (Subtarget->isPICStyleStubPIC()) {
13775 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
13776 } else if (Subtarget->isPICStyleStubNoDynamic()) {
13777 OpFlag = X86II::MO_DARWIN_NONLAZY;
13780 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
13783 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13785 // With PIC, the address is actually $g + Offset.
13786 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
13787 !Subtarget->is64Bit()) {
13788 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13789 DAG.getNode(X86ISD::GlobalBaseReg,
13790 SDLoc(), getPointerTy()),
13794 // For symbols that require a load from a stub to get the address, emit the
13796 if (isGlobalStubReference(OpFlag))
13797 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
13798 MachinePointerInfo::getGOT(), false, false, false, 0);
13804 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
13805 // Create the TargetBlockAddressAddress node.
13806 unsigned char OpFlags =
13807 Subtarget->ClassifyBlockAddressReference();
13808 CodeModel::Model M = DAG.getTarget().getCodeModel();
13809 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
13810 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
13812 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
13815 if (Subtarget->isPICStyleRIPRel() &&
13816 (M == CodeModel::Small || M == CodeModel::Kernel))
13817 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13819 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13821 // With PIC, the address is actually $g + Offset.
13822 if (isGlobalRelativeToPICBase(OpFlags)) {
13823 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13824 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13832 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
13833 int64_t Offset, SelectionDAG &DAG) const {
13834 // Create the TargetGlobalAddress node, folding in the constant
13835 // offset if it is legal.
13836 unsigned char OpFlags =
13837 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
13838 CodeModel::Model M = DAG.getTarget().getCodeModel();
13840 if (OpFlags == X86II::MO_NO_FLAG &&
13841 X86::isOffsetSuitableForCodeModel(Offset, M)) {
13842 // A direct static reference to a global.
13843 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
13846 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
13849 if (Subtarget->isPICStyleRIPRel() &&
13850 (M == CodeModel::Small || M == CodeModel::Kernel))
13851 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13853 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13855 // With PIC, the address is actually $g + Offset.
13856 if (isGlobalRelativeToPICBase(OpFlags)) {
13857 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13858 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13862 // For globals that require a load from a stub to get the address, emit the
13864 if (isGlobalStubReference(OpFlags))
13865 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
13866 MachinePointerInfo::getGOT(), false, false, false, 0);
13868 // If there was a non-zero offset that we didn't fold, create an explicit
13869 // addition for it.
13871 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
13872 DAG.getConstant(Offset, getPointerTy()));
13878 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
13879 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
13880 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
13881 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
13885 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
13886 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
13887 unsigned char OperandFlags, bool LocalDynamic = false) {
13888 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13889 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13891 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13892 GA->getValueType(0),
13896 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
13900 SDValue Ops[] = { Chain, TGA, *InFlag };
13901 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13903 SDValue Ops[] = { Chain, TGA };
13904 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13907 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
13908 MFI->setAdjustsStack(true);
13909 MFI->setHasCalls(true);
13911 SDValue Flag = Chain.getValue(1);
13912 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
13915 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
13917 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13920 SDLoc dl(GA); // ? function entry point might be better
13921 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13922 DAG.getNode(X86ISD::GlobalBaseReg,
13923 SDLoc(), PtrVT), InFlag);
13924 InFlag = Chain.getValue(1);
13926 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
13929 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
13931 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13933 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
13934 X86::RAX, X86II::MO_TLSGD);
13937 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
13943 // Get the start address of the TLS block for this module.
13944 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
13945 .getInfo<X86MachineFunctionInfo>();
13946 MFI->incNumLocalDynamicTLSAccesses();
13950 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
13951 X86II::MO_TLSLD, /*LocalDynamic=*/true);
13954 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13955 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
13956 InFlag = Chain.getValue(1);
13957 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
13958 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
13961 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
13965 unsigned char OperandFlags = X86II::MO_DTPOFF;
13966 unsigned WrapperKind = X86ISD::Wrapper;
13967 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13968 GA->getValueType(0),
13969 GA->getOffset(), OperandFlags);
13970 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
13972 // Add x@dtpoff with the base.
13973 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
13976 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
13977 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13978 const EVT PtrVT, TLSModel::Model model,
13979 bool is64Bit, bool isPIC) {
13982 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
13983 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
13984 is64Bit ? 257 : 256));
13986 SDValue ThreadPointer =
13987 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
13988 MachinePointerInfo(Ptr), false, false, false, 0);
13990 unsigned char OperandFlags = 0;
13991 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
13993 unsigned WrapperKind = X86ISD::Wrapper;
13994 if (model == TLSModel::LocalExec) {
13995 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
13996 } else if (model == TLSModel::InitialExec) {
13998 OperandFlags = X86II::MO_GOTTPOFF;
13999 WrapperKind = X86ISD::WrapperRIP;
14001 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
14004 llvm_unreachable("Unexpected model");
14007 // emit "addl x@ntpoff,%eax" (local exec)
14008 // or "addl x@indntpoff,%eax" (initial exec)
14009 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
14011 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
14012 GA->getOffset(), OperandFlags);
14013 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
14015 if (model == TLSModel::InitialExec) {
14016 if (isPIC && !is64Bit) {
14017 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
14018 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
14022 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
14023 MachinePointerInfo::getGOT(), false, false, false, 0);
14026 // The address of the thread local variable is the add of the thread
14027 // pointer with the offset of the variable.
14028 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
14032 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
14034 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
14035 const GlobalValue *GV = GA->getGlobal();
14037 if (Subtarget->isTargetELF()) {
14038 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
14041 case TLSModel::GeneralDynamic:
14042 if (Subtarget->is64Bit())
14043 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
14044 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
14045 case TLSModel::LocalDynamic:
14046 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
14047 Subtarget->is64Bit());
14048 case TLSModel::InitialExec:
14049 case TLSModel::LocalExec:
14050 return LowerToTLSExecModel(
14051 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
14052 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
14054 llvm_unreachable("Unknown TLS model.");
14057 if (Subtarget->isTargetDarwin()) {
14058 // Darwin only has one model of TLS. Lower to that.
14059 unsigned char OpFlag = 0;
14060 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
14061 X86ISD::WrapperRIP : X86ISD::Wrapper;
14063 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
14064 // global base reg.
14065 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
14066 !Subtarget->is64Bit();
14068 OpFlag = X86II::MO_TLVP_PIC_BASE;
14070 OpFlag = X86II::MO_TLVP;
14072 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
14073 GA->getValueType(0),
14074 GA->getOffset(), OpFlag);
14075 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
14077 // With PIC32, the address is actually $g + Offset.
14079 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14080 DAG.getNode(X86ISD::GlobalBaseReg,
14081 SDLoc(), getPointerTy()),
14084 // Lowering the machine isd will make sure everything is in the right
14086 SDValue Chain = DAG.getEntryNode();
14087 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14088 SDValue Args[] = { Chain, Offset };
14089 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
14091 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
14092 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
14093 MFI->setAdjustsStack(true);
14095 // And our return value (tls address) is in the standard call return value
14097 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
14098 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
14099 Chain.getValue(1));
14102 if (Subtarget->isTargetKnownWindowsMSVC() ||
14103 Subtarget->isTargetWindowsGNU()) {
14104 // Just use the implicit TLS architecture
14105 // Need to generate someting similar to:
14106 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
14108 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
14109 // mov rcx, qword [rdx+rcx*8]
14110 // mov eax, .tls$:tlsvar
14111 // [rax+rcx] contains the address
14112 // Windows 64bit: gs:0x58
14113 // Windows 32bit: fs:__tls_array
14116 SDValue Chain = DAG.getEntryNode();
14118 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
14119 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
14120 // use its literal value of 0x2C.
14121 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
14122 ? Type::getInt8PtrTy(*DAG.getContext(),
14124 : Type::getInt32PtrTy(*DAG.getContext(),
14128 Subtarget->is64Bit()
14129 ? DAG.getIntPtrConstant(0x58)
14130 : (Subtarget->isTargetWindowsGNU()
14131 ? DAG.getIntPtrConstant(0x2C)
14132 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
14134 SDValue ThreadPointer =
14135 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
14136 MachinePointerInfo(Ptr), false, false, false, 0);
14138 // Load the _tls_index variable
14139 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
14140 if (Subtarget->is64Bit())
14141 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
14142 IDX, MachinePointerInfo(), MVT::i32,
14143 false, false, false, 0);
14145 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
14146 false, false, false, 0);
14148 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
14150 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
14152 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
14153 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
14154 false, false, false, 0);
14156 // Get the offset of start of .tls section
14157 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
14158 GA->getValueType(0),
14159 GA->getOffset(), X86II::MO_SECREL);
14160 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
14162 // The address of the thread local variable is the add of the thread
14163 // pointer with the offset of the variable.
14164 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
14167 llvm_unreachable("TLS not implemented for this target.");
14170 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
14171 /// and take a 2 x i32 value to shift plus a shift amount.
14172 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
14173 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
14174 MVT VT = Op.getSimpleValueType();
14175 unsigned VTBits = VT.getSizeInBits();
14177 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
14178 SDValue ShOpLo = Op.getOperand(0);
14179 SDValue ShOpHi = Op.getOperand(1);
14180 SDValue ShAmt = Op.getOperand(2);
14181 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
14182 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
14184 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
14185 DAG.getConstant(VTBits - 1, MVT::i8));
14186 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
14187 DAG.getConstant(VTBits - 1, MVT::i8))
14188 : DAG.getConstant(0, VT);
14190 SDValue Tmp2, Tmp3;
14191 if (Op.getOpcode() == ISD::SHL_PARTS) {
14192 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
14193 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
14195 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
14196 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
14199 // If the shift amount is larger or equal than the width of a part we can't
14200 // rely on the results of shld/shrd. Insert a test and select the appropriate
14201 // values for large shift amounts.
14202 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
14203 DAG.getConstant(VTBits, MVT::i8));
14204 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14205 AndNode, DAG.getConstant(0, MVT::i8));
14208 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
14209 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
14210 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
14212 if (Op.getOpcode() == ISD::SHL_PARTS) {
14213 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
14214 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
14216 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
14217 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
14220 SDValue Ops[2] = { Lo, Hi };
14221 return DAG.getMergeValues(Ops, dl);
14224 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
14225 SelectionDAG &DAG) const {
14226 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
14229 if (SrcVT.isVector()) {
14230 if (SrcVT.getVectorElementType() == MVT::i1) {
14231 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
14232 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
14233 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
14234 Op.getOperand(0)));
14239 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
14240 "Unknown SINT_TO_FP to lower!");
14242 // These are really Legal; return the operand so the caller accepts it as
14244 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
14246 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
14247 Subtarget->is64Bit()) {
14251 unsigned Size = SrcVT.getSizeInBits()/8;
14252 MachineFunction &MF = DAG.getMachineFunction();
14253 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
14254 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14255 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14257 MachinePointerInfo::getFixedStack(SSFI),
14259 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
14262 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
14264 SelectionDAG &DAG) const {
14268 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
14270 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
14272 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
14274 unsigned ByteSize = SrcVT.getSizeInBits()/8;
14276 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
14277 MachineMemOperand *MMO;
14279 int SSFI = FI->getIndex();
14281 DAG.getMachineFunction()
14282 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14283 MachineMemOperand::MOLoad, ByteSize, ByteSize);
14285 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
14286 StackSlot = StackSlot.getOperand(1);
14288 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
14289 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
14291 Tys, Ops, SrcVT, MMO);
14294 Chain = Result.getValue(1);
14295 SDValue InFlag = Result.getValue(2);
14297 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
14298 // shouldn't be necessary except that RFP cannot be live across
14299 // multiple blocks. When stackifier is fixed, they can be uncoupled.
14300 MachineFunction &MF = DAG.getMachineFunction();
14301 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
14302 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
14303 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14304 Tys = DAG.getVTList(MVT::Other);
14306 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
14308 MachineMemOperand *MMO =
14309 DAG.getMachineFunction()
14310 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14311 MachineMemOperand::MOStore, SSFISize, SSFISize);
14313 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
14314 Ops, Op.getValueType(), MMO);
14315 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
14316 MachinePointerInfo::getFixedStack(SSFI),
14317 false, false, false, 0);
14323 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
14324 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
14325 SelectionDAG &DAG) const {
14326 // This algorithm is not obvious. Here it is what we're trying to output:
14329 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
14330 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
14332 haddpd %xmm0, %xmm0
14334 pshufd $0x4e, %xmm0, %xmm1
14340 LLVMContext *Context = DAG.getContext();
14342 // Build some magic constants.
14343 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
14344 Constant *C0 = ConstantDataVector::get(*Context, CV0);
14345 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
14347 SmallVector<Constant*,2> CV1;
14349 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
14350 APInt(64, 0x4330000000000000ULL))));
14352 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
14353 APInt(64, 0x4530000000000000ULL))));
14354 Constant *C1 = ConstantVector::get(CV1);
14355 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
14357 // Load the 64-bit value into an XMM register.
14358 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
14360 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
14361 MachinePointerInfo::getConstantPool(),
14362 false, false, false, 16);
14363 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
14364 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
14367 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
14368 MachinePointerInfo::getConstantPool(),
14369 false, false, false, 16);
14370 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
14371 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
14374 if (Subtarget->hasSSE3()) {
14375 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
14376 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
14378 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
14379 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
14381 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
14382 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
14386 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
14387 DAG.getIntPtrConstant(0));
14390 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
14391 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
14392 SelectionDAG &DAG) const {
14394 // FP constant to bias correct the final result.
14395 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
14398 // Load the 32-bit value into an XMM register.
14399 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
14402 // Zero out the upper parts of the register.
14403 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
14405 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
14406 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
14407 DAG.getIntPtrConstant(0));
14409 // Or the load with the bias.
14410 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
14411 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
14412 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
14413 MVT::v2f64, Load)),
14414 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
14415 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
14416 MVT::v2f64, Bias)));
14417 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
14418 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
14419 DAG.getIntPtrConstant(0));
14421 // Subtract the bias.
14422 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
14424 // Handle final rounding.
14425 EVT DestVT = Op.getValueType();
14427 if (DestVT.bitsLT(MVT::f64))
14428 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
14429 DAG.getIntPtrConstant(0));
14430 if (DestVT.bitsGT(MVT::f64))
14431 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
14433 // Handle final rounding.
14437 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
14438 const X86Subtarget &Subtarget) {
14439 // The algorithm is the following:
14440 // #ifdef __SSE4_1__
14441 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
14442 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
14443 // (uint4) 0x53000000, 0xaa);
14445 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
14446 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
14448 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
14449 // return (float4) lo + fhi;
14452 SDValue V = Op->getOperand(0);
14453 EVT VecIntVT = V.getValueType();
14454 bool Is128 = VecIntVT == MVT::v4i32;
14455 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
14456 // If we convert to something else than the supported type, e.g., to v4f64,
14458 if (VecFloatVT != Op->getValueType(0))
14461 unsigned NumElts = VecIntVT.getVectorNumElements();
14462 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
14463 "Unsupported custom type");
14464 assert(NumElts <= 8 && "The size of the constant array must be fixed");
14466 // In the #idef/#else code, we have in common:
14467 // - The vector of constants:
14473 // Create the splat vector for 0x4b000000.
14474 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
14475 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
14476 CstLow, CstLow, CstLow, CstLow};
14477 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14478 makeArrayRef(&CstLowArray[0], NumElts));
14479 // Create the splat vector for 0x53000000.
14480 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
14481 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
14482 CstHigh, CstHigh, CstHigh, CstHigh};
14483 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14484 makeArrayRef(&CstHighArray[0], NumElts));
14486 // Create the right shift.
14487 SDValue CstShift = DAG.getConstant(16, MVT::i32);
14488 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
14489 CstShift, CstShift, CstShift, CstShift};
14490 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14491 makeArrayRef(&CstShiftArray[0], NumElts));
14492 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
14495 if (Subtarget.hasSSE41()) {
14496 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
14497 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
14498 SDValue VecCstLowBitcast =
14499 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
14500 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
14501 // Low will be bitcasted right away, so do not bother bitcasting back to its
14503 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
14504 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
14505 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
14506 // (uint4) 0x53000000, 0xaa);
14507 SDValue VecCstHighBitcast =
14508 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
14509 SDValue VecShiftBitcast =
14510 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
14511 // High will be bitcasted right away, so do not bother bitcasting back to
14512 // its original type.
14513 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
14514 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
14516 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
14517 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
14518 CstMask, CstMask, CstMask);
14519 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
14520 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
14521 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
14523 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
14524 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
14527 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
14528 SDValue CstFAdd = DAG.getConstantFP(
14529 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
14530 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
14531 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
14532 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
14533 makeArrayRef(&CstFAddArray[0], NumElts));
14535 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
14536 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
14538 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
14539 // return (float4) lo + fhi;
14540 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
14541 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
14544 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
14545 SelectionDAG &DAG) const {
14546 SDValue N0 = Op.getOperand(0);
14547 MVT SVT = N0.getSimpleValueType();
14550 switch (SVT.SimpleTy) {
14552 llvm_unreachable("Custom UINT_TO_FP is not supported!");
14557 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
14558 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
14559 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
14563 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
14565 llvm_unreachable(nullptr);
14568 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
14569 SelectionDAG &DAG) const {
14570 SDValue N0 = Op.getOperand(0);
14573 if (Op.getValueType().isVector())
14574 return lowerUINT_TO_FP_vec(Op, DAG);
14576 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
14577 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
14578 // the optimization here.
14579 if (DAG.SignBitIsZero(N0))
14580 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
14582 MVT SrcVT = N0.getSimpleValueType();
14583 MVT DstVT = Op.getSimpleValueType();
14584 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
14585 return LowerUINT_TO_FP_i64(Op, DAG);
14586 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
14587 return LowerUINT_TO_FP_i32(Op, DAG);
14588 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
14591 // Make a 64-bit buffer, and use it to build an FILD.
14592 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
14593 if (SrcVT == MVT::i32) {
14594 SDValue WordOff = DAG.getConstant(4, getPointerTy());
14595 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
14596 getPointerTy(), StackSlot, WordOff);
14597 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14598 StackSlot, MachinePointerInfo(),
14600 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
14601 OffsetSlot, MachinePointerInfo(),
14603 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
14607 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
14608 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14609 StackSlot, MachinePointerInfo(),
14611 // For i64 source, we need to add the appropriate power of 2 if the input
14612 // was negative. This is the same as the optimization in
14613 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
14614 // we must be careful to do the computation in x87 extended precision, not
14615 // in SSE. (The generic code can't know it's OK to do this, or how to.)
14616 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
14617 MachineMemOperand *MMO =
14618 DAG.getMachineFunction()
14619 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14620 MachineMemOperand::MOLoad, 8, 8);
14622 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
14623 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
14624 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
14627 APInt FF(32, 0x5F800000ULL);
14629 // Check whether the sign bit is set.
14630 SDValue SignSet = DAG.getSetCC(dl,
14631 getSetCCResultType(*DAG.getContext(), MVT::i64),
14632 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
14635 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
14636 SDValue FudgePtr = DAG.getConstantPool(
14637 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
14640 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
14641 SDValue Zero = DAG.getIntPtrConstant(0);
14642 SDValue Four = DAG.getIntPtrConstant(4);
14643 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
14645 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
14647 // Load the value out, extending it from f32 to f80.
14648 // FIXME: Avoid the extend by constructing the right constant pool?
14649 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
14650 FudgePtr, MachinePointerInfo::getConstantPool(),
14651 MVT::f32, false, false, false, 4);
14652 // Extend everything to 80 bits to force it to be done on x87.
14653 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
14654 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
14657 std::pair<SDValue,SDValue>
14658 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
14659 bool IsSigned, bool IsReplace) const {
14662 EVT DstTy = Op.getValueType();
14664 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
14665 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
14669 assert(DstTy.getSimpleVT() <= MVT::i64 &&
14670 DstTy.getSimpleVT() >= MVT::i16 &&
14671 "Unknown FP_TO_INT to lower!");
14673 // These are really Legal.
14674 if (DstTy == MVT::i32 &&
14675 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14676 return std::make_pair(SDValue(), SDValue());
14677 if (Subtarget->is64Bit() &&
14678 DstTy == MVT::i64 &&
14679 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14680 return std::make_pair(SDValue(), SDValue());
14682 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
14683 // stack slot, or into the FTOL runtime function.
14684 MachineFunction &MF = DAG.getMachineFunction();
14685 unsigned MemSize = DstTy.getSizeInBits()/8;
14686 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14687 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14690 if (!IsSigned && isIntegerTypeFTOL(DstTy))
14691 Opc = X86ISD::WIN_FTOL;
14693 switch (DstTy.getSimpleVT().SimpleTy) {
14694 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
14695 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
14696 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
14697 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
14700 SDValue Chain = DAG.getEntryNode();
14701 SDValue Value = Op.getOperand(0);
14702 EVT TheVT = Op.getOperand(0).getValueType();
14703 // FIXME This causes a redundant load/store if the SSE-class value is already
14704 // in memory, such as if it is on the callstack.
14705 if (isScalarFPTypeInSSEReg(TheVT)) {
14706 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
14707 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
14708 MachinePointerInfo::getFixedStack(SSFI),
14710 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
14712 Chain, StackSlot, DAG.getValueType(TheVT)
14715 MachineMemOperand *MMO =
14716 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14717 MachineMemOperand::MOLoad, MemSize, MemSize);
14718 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
14719 Chain = Value.getValue(1);
14720 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14721 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14724 MachineMemOperand *MMO =
14725 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14726 MachineMemOperand::MOStore, MemSize, MemSize);
14728 if (Opc != X86ISD::WIN_FTOL) {
14729 // Build the FP_TO_INT*_IN_MEM
14730 SDValue Ops[] = { Chain, Value, StackSlot };
14731 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
14733 return std::make_pair(FIST, StackSlot);
14735 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
14736 DAG.getVTList(MVT::Other, MVT::Glue),
14738 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
14739 MVT::i32, ftol.getValue(1));
14740 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
14741 MVT::i32, eax.getValue(2));
14742 SDValue Ops[] = { eax, edx };
14743 SDValue pair = IsReplace
14744 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
14745 : DAG.getMergeValues(Ops, DL);
14746 return std::make_pair(pair, SDValue());
14750 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
14751 const X86Subtarget *Subtarget) {
14752 MVT VT = Op->getSimpleValueType(0);
14753 SDValue In = Op->getOperand(0);
14754 MVT InVT = In.getSimpleValueType();
14757 // Optimize vectors in AVX mode:
14760 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
14761 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
14762 // Concat upper and lower parts.
14765 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
14766 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
14767 // Concat upper and lower parts.
14770 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
14771 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
14772 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
14775 if (Subtarget->hasInt256())
14776 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
14778 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
14779 SDValue Undef = DAG.getUNDEF(InVT);
14780 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
14781 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14782 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14784 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
14785 VT.getVectorNumElements()/2);
14787 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
14788 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
14790 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14793 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
14794 SelectionDAG &DAG) {
14795 MVT VT = Op->getSimpleValueType(0);
14796 SDValue In = Op->getOperand(0);
14797 MVT InVT = In.getSimpleValueType();
14799 unsigned int NumElts = VT.getVectorNumElements();
14800 if (NumElts != 8 && NumElts != 16)
14803 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
14804 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
14806 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
14807 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14808 // Now we have only mask extension
14809 assert(InVT.getVectorElementType() == MVT::i1);
14810 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
14811 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14812 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
14813 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14814 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14815 MachinePointerInfo::getConstantPool(),
14816 false, false, false, Alignment);
14818 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
14819 if (VT.is512BitVector())
14821 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
14824 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14825 SelectionDAG &DAG) {
14826 if (Subtarget->hasFp256()) {
14827 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14835 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14836 SelectionDAG &DAG) {
14838 MVT VT = Op.getSimpleValueType();
14839 SDValue In = Op.getOperand(0);
14840 MVT SVT = In.getSimpleValueType();
14842 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
14843 return LowerZERO_EXTEND_AVX512(Op, DAG);
14845 if (Subtarget->hasFp256()) {
14846 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14851 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
14852 VT.getVectorNumElements() != SVT.getVectorNumElements());
14856 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
14858 MVT VT = Op.getSimpleValueType();
14859 SDValue In = Op.getOperand(0);
14860 MVT InVT = In.getSimpleValueType();
14862 if (VT == MVT::i1) {
14863 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
14864 "Invalid scalar TRUNCATE operation");
14865 if (InVT.getSizeInBits() >= 32)
14867 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
14868 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
14870 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
14871 "Invalid TRUNCATE operation");
14873 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
14874 if (VT.getVectorElementType().getSizeInBits() >=8)
14875 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
14877 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14878 unsigned NumElts = InVT.getVectorNumElements();
14879 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
14880 if (InVT.getSizeInBits() < 512) {
14881 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
14882 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
14886 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
14887 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14888 SDValue CP = DAG.getConstantPool(C, getPointerTy());
14889 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14890 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14891 MachinePointerInfo::getConstantPool(),
14892 false, false, false, Alignment);
14893 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
14894 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
14895 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
14898 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
14899 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
14900 if (Subtarget->hasInt256()) {
14901 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
14902 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
14903 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
14905 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
14906 DAG.getIntPtrConstant(0));
14909 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14910 DAG.getIntPtrConstant(0));
14911 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14912 DAG.getIntPtrConstant(2));
14913 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14914 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14915 static const int ShufMask[] = {0, 2, 4, 6};
14916 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
14919 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
14920 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
14921 if (Subtarget->hasInt256()) {
14922 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
14924 SmallVector<SDValue,32> pshufbMask;
14925 for (unsigned i = 0; i < 2; ++i) {
14926 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
14927 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
14928 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
14929 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
14930 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
14931 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
14932 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
14933 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
14934 for (unsigned j = 0; j < 8; ++j)
14935 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
14937 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
14938 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
14939 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
14941 static const int ShufMask[] = {0, 2, -1, -1};
14942 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
14944 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14945 DAG.getIntPtrConstant(0));
14946 return DAG.getNode(ISD::BITCAST, DL, VT, In);
14949 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14950 DAG.getIntPtrConstant(0));
14952 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14953 DAG.getIntPtrConstant(4));
14955 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
14956 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
14958 // The PSHUFB mask:
14959 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
14960 -1, -1, -1, -1, -1, -1, -1, -1};
14962 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
14963 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
14964 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
14966 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14967 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14969 // The MOVLHPS Mask:
14970 static const int ShufMask2[] = {0, 1, 4, 5};
14971 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
14972 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
14975 // Handle truncation of V256 to V128 using shuffles.
14976 if (!VT.is128BitVector() || !InVT.is256BitVector())
14979 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
14981 unsigned NumElems = VT.getVectorNumElements();
14982 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
14984 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
14985 // Prepare truncation shuffle mask
14986 for (unsigned i = 0; i != NumElems; ++i)
14987 MaskVec[i] = i * 2;
14988 SDValue V = DAG.getVectorShuffle(NVT, DL,
14989 DAG.getNode(ISD::BITCAST, DL, NVT, In),
14990 DAG.getUNDEF(NVT), &MaskVec[0]);
14991 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
14992 DAG.getIntPtrConstant(0));
14995 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
14996 SelectionDAG &DAG) const {
14997 assert(!Op.getSimpleValueType().isVector());
14999 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
15000 /*IsSigned=*/ true, /*IsReplace=*/ false);
15001 SDValue FIST = Vals.first, StackSlot = Vals.second;
15002 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
15003 if (!FIST.getNode()) return Op;
15005 if (StackSlot.getNode())
15006 // Load the result.
15007 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
15008 FIST, StackSlot, MachinePointerInfo(),
15009 false, false, false, 0);
15011 // The node is the result.
15015 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
15016 SelectionDAG &DAG) const {
15017 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
15018 /*IsSigned=*/ false, /*IsReplace=*/ false);
15019 SDValue FIST = Vals.first, StackSlot = Vals.second;
15020 assert(FIST.getNode() && "Unexpected failure");
15022 if (StackSlot.getNode())
15023 // Load the result.
15024 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
15025 FIST, StackSlot, MachinePointerInfo(),
15026 false, false, false, 0);
15028 // The node is the result.
15032 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
15034 MVT VT = Op.getSimpleValueType();
15035 SDValue In = Op.getOperand(0);
15036 MVT SVT = In.getSimpleValueType();
15038 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
15040 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
15041 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
15042 In, DAG.getUNDEF(SVT)));
15045 /// The only differences between FABS and FNEG are the mask and the logic op.
15046 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
15047 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
15048 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
15049 "Wrong opcode for lowering FABS or FNEG.");
15051 bool IsFABS = (Op.getOpcode() == ISD::FABS);
15053 // If this is a FABS and it has an FNEG user, bail out to fold the combination
15054 // into an FNABS. We'll lower the FABS after that if it is still in use.
15056 for (SDNode *User : Op->uses())
15057 if (User->getOpcode() == ISD::FNEG)
15060 SDValue Op0 = Op.getOperand(0);
15061 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
15064 MVT VT = Op.getSimpleValueType();
15065 // Assume scalar op for initialization; update for vector if needed.
15066 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
15067 // generate a 16-byte vector constant and logic op even for the scalar case.
15068 // Using a 16-byte mask allows folding the load of the mask with
15069 // the logic op, so it can save (~4 bytes) on code size.
15071 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
15072 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
15073 // decide if we should generate a 16-byte constant mask when we only need 4 or
15074 // 8 bytes for the scalar case.
15075 if (VT.isVector()) {
15076 EltVT = VT.getVectorElementType();
15077 NumElts = VT.getVectorNumElements();
15080 unsigned EltBits = EltVT.getSizeInBits();
15081 LLVMContext *Context = DAG.getContext();
15082 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
15084 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
15085 Constant *C = ConstantInt::get(*Context, MaskElt);
15086 C = ConstantVector::getSplat(NumElts, C);
15087 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15088 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
15089 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
15090 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
15091 MachinePointerInfo::getConstantPool(),
15092 false, false, false, Alignment);
15094 if (VT.isVector()) {
15095 // For a vector, cast operands to a vector type, perform the logic op,
15096 // and cast the result back to the original value type.
15097 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
15098 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
15099 SDValue Operand = IsFNABS ?
15100 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
15101 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
15102 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
15103 return DAG.getNode(ISD::BITCAST, dl, VT,
15104 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
15107 // If not vector, then scalar.
15108 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
15109 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
15110 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
15113 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
15114 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15115 LLVMContext *Context = DAG.getContext();
15116 SDValue Op0 = Op.getOperand(0);
15117 SDValue Op1 = Op.getOperand(1);
15119 MVT VT = Op.getSimpleValueType();
15120 MVT SrcVT = Op1.getSimpleValueType();
15122 // If second operand is smaller, extend it first.
15123 if (SrcVT.bitsLT(VT)) {
15124 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
15127 // And if it is bigger, shrink it first.
15128 if (SrcVT.bitsGT(VT)) {
15129 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
15133 // At this point the operands and the result should have the same
15134 // type, and that won't be f80 since that is not custom lowered.
15136 const fltSemantics &Sem =
15137 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
15138 const unsigned SizeInBits = VT.getSizeInBits();
15140 SmallVector<Constant *, 4> CV(
15141 VT == MVT::f64 ? 2 : 4,
15142 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
15144 // First, clear all bits but the sign bit from the second operand (sign).
15145 CV[0] = ConstantFP::get(*Context,
15146 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
15147 Constant *C = ConstantVector::get(CV);
15148 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
15149 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
15150 MachinePointerInfo::getConstantPool(),
15151 false, false, false, 16);
15152 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
15154 // Next, clear the sign bit from the first operand (magnitude).
15155 // If it's a constant, we can clear it here.
15156 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
15157 APFloat APF = Op0CN->getValueAPF();
15158 // If the magnitude is a positive zero, the sign bit alone is enough.
15159 if (APF.isPosZero())
15162 CV[0] = ConstantFP::get(*Context, APF);
15164 CV[0] = ConstantFP::get(
15166 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
15168 C = ConstantVector::get(CV);
15169 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
15170 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
15171 MachinePointerInfo::getConstantPool(),
15172 false, false, false, 16);
15173 // If the magnitude operand wasn't a constant, we need to AND out the sign.
15174 if (!isa<ConstantFPSDNode>(Op0))
15175 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
15177 // OR the magnitude value with the sign bit.
15178 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
15181 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
15182 SDValue N0 = Op.getOperand(0);
15184 MVT VT = Op.getSimpleValueType();
15186 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
15187 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
15188 DAG.getConstant(1, VT));
15189 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
15192 // Check whether an OR'd tree is PTEST-able.
15193 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
15194 SelectionDAG &DAG) {
15195 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
15197 if (!Subtarget->hasSSE41())
15200 if (!Op->hasOneUse())
15203 SDNode *N = Op.getNode();
15206 SmallVector<SDValue, 8> Opnds;
15207 DenseMap<SDValue, unsigned> VecInMap;
15208 SmallVector<SDValue, 8> VecIns;
15209 EVT VT = MVT::Other;
15211 // Recognize a special case where a vector is casted into wide integer to
15213 Opnds.push_back(N->getOperand(0));
15214 Opnds.push_back(N->getOperand(1));
15216 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
15217 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
15218 // BFS traverse all OR'd operands.
15219 if (I->getOpcode() == ISD::OR) {
15220 Opnds.push_back(I->getOperand(0));
15221 Opnds.push_back(I->getOperand(1));
15222 // Re-evaluate the number of nodes to be traversed.
15223 e += 2; // 2 more nodes (LHS and RHS) are pushed.
15227 // Quit if a non-EXTRACT_VECTOR_ELT
15228 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
15231 // Quit if without a constant index.
15232 SDValue Idx = I->getOperand(1);
15233 if (!isa<ConstantSDNode>(Idx))
15236 SDValue ExtractedFromVec = I->getOperand(0);
15237 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
15238 if (M == VecInMap.end()) {
15239 VT = ExtractedFromVec.getValueType();
15240 // Quit if not 128/256-bit vector.
15241 if (!VT.is128BitVector() && !VT.is256BitVector())
15243 // Quit if not the same type.
15244 if (VecInMap.begin() != VecInMap.end() &&
15245 VT != VecInMap.begin()->first.getValueType())
15247 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
15248 VecIns.push_back(ExtractedFromVec);
15250 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
15253 assert((VT.is128BitVector() || VT.is256BitVector()) &&
15254 "Not extracted from 128-/256-bit vector.");
15256 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
15258 for (DenseMap<SDValue, unsigned>::const_iterator
15259 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
15260 // Quit if not all elements are used.
15261 if (I->second != FullMask)
15265 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
15267 // Cast all vectors into TestVT for PTEST.
15268 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
15269 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
15271 // If more than one full vectors are evaluated, OR them first before PTEST.
15272 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
15273 // Each iteration will OR 2 nodes and append the result until there is only
15274 // 1 node left, i.e. the final OR'd value of all vectors.
15275 SDValue LHS = VecIns[Slot];
15276 SDValue RHS = VecIns[Slot + 1];
15277 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
15280 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
15281 VecIns.back(), VecIns.back());
15284 /// \brief return true if \c Op has a use that doesn't just read flags.
15285 static bool hasNonFlagsUse(SDValue Op) {
15286 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
15288 SDNode *User = *UI;
15289 unsigned UOpNo = UI.getOperandNo();
15290 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
15291 // Look pass truncate.
15292 UOpNo = User->use_begin().getOperandNo();
15293 User = *User->use_begin();
15296 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
15297 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
15303 /// Emit nodes that will be selected as "test Op0,Op0", or something
15305 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
15306 SelectionDAG &DAG) const {
15307 if (Op.getValueType() == MVT::i1) {
15308 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
15309 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
15310 DAG.getConstant(0, MVT::i8));
15312 // CF and OF aren't always set the way we want. Determine which
15313 // of these we need.
15314 bool NeedCF = false;
15315 bool NeedOF = false;
15318 case X86::COND_A: case X86::COND_AE:
15319 case X86::COND_B: case X86::COND_BE:
15322 case X86::COND_G: case X86::COND_GE:
15323 case X86::COND_L: case X86::COND_LE:
15324 case X86::COND_O: case X86::COND_NO: {
15325 // Check if we really need to set the
15326 // Overflow flag. If NoSignedWrap is present
15327 // that is not actually needed.
15328 switch (Op->getOpcode()) {
15333 const BinaryWithFlagsSDNode *BinNode =
15334 cast<BinaryWithFlagsSDNode>(Op.getNode());
15335 if (BinNode->hasNoSignedWrap())
15345 // See if we can use the EFLAGS value from the operand instead of
15346 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
15347 // we prove that the arithmetic won't overflow, we can't use OF or CF.
15348 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
15349 // Emit a CMP with 0, which is the TEST pattern.
15350 //if (Op.getValueType() == MVT::i1)
15351 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
15352 // DAG.getConstant(0, MVT::i1));
15353 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
15354 DAG.getConstant(0, Op.getValueType()));
15356 unsigned Opcode = 0;
15357 unsigned NumOperands = 0;
15359 // Truncate operations may prevent the merge of the SETCC instruction
15360 // and the arithmetic instruction before it. Attempt to truncate the operands
15361 // of the arithmetic instruction and use a reduced bit-width instruction.
15362 bool NeedTruncation = false;
15363 SDValue ArithOp = Op;
15364 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
15365 SDValue Arith = Op->getOperand(0);
15366 // Both the trunc and the arithmetic op need to have one user each.
15367 if (Arith->hasOneUse())
15368 switch (Arith.getOpcode()) {
15375 NeedTruncation = true;
15381 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
15382 // which may be the result of a CAST. We use the variable 'Op', which is the
15383 // non-casted variable when we check for possible users.
15384 switch (ArithOp.getOpcode()) {
15386 // Due to an isel shortcoming, be conservative if this add is likely to be
15387 // selected as part of a load-modify-store instruction. When the root node
15388 // in a match is a store, isel doesn't know how to remap non-chain non-flag
15389 // uses of other nodes in the match, such as the ADD in this case. This
15390 // leads to the ADD being left around and reselected, with the result being
15391 // two adds in the output. Alas, even if none our users are stores, that
15392 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
15393 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
15394 // climbing the DAG back to the root, and it doesn't seem to be worth the
15396 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
15397 UE = Op.getNode()->use_end(); UI != UE; ++UI)
15398 if (UI->getOpcode() != ISD::CopyToReg &&
15399 UI->getOpcode() != ISD::SETCC &&
15400 UI->getOpcode() != ISD::STORE)
15403 if (ConstantSDNode *C =
15404 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
15405 // An add of one will be selected as an INC.
15406 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
15407 Opcode = X86ISD::INC;
15412 // An add of negative one (subtract of one) will be selected as a DEC.
15413 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
15414 Opcode = X86ISD::DEC;
15420 // Otherwise use a regular EFLAGS-setting add.
15421 Opcode = X86ISD::ADD;
15426 // If we have a constant logical shift that's only used in a comparison
15427 // against zero turn it into an equivalent AND. This allows turning it into
15428 // a TEST instruction later.
15429 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
15430 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
15431 EVT VT = Op.getValueType();
15432 unsigned BitWidth = VT.getSizeInBits();
15433 unsigned ShAmt = Op->getConstantOperandVal(1);
15434 if (ShAmt >= BitWidth) // Avoid undefined shifts.
15436 APInt Mask = ArithOp.getOpcode() == ISD::SRL
15437 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
15438 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
15439 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
15441 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
15442 DAG.getConstant(Mask, VT));
15443 DAG.ReplaceAllUsesWith(Op, New);
15449 // If the primary and result isn't used, don't bother using X86ISD::AND,
15450 // because a TEST instruction will be better.
15451 if (!hasNonFlagsUse(Op))
15457 // Due to the ISEL shortcoming noted above, be conservative if this op is
15458 // likely to be selected as part of a load-modify-store instruction.
15459 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
15460 UE = Op.getNode()->use_end(); UI != UE; ++UI)
15461 if (UI->getOpcode() == ISD::STORE)
15464 // Otherwise use a regular EFLAGS-setting instruction.
15465 switch (ArithOp.getOpcode()) {
15466 default: llvm_unreachable("unexpected operator!");
15467 case ISD::SUB: Opcode = X86ISD::SUB; break;
15468 case ISD::XOR: Opcode = X86ISD::XOR; break;
15469 case ISD::AND: Opcode = X86ISD::AND; break;
15471 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
15472 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
15473 if (EFLAGS.getNode())
15476 Opcode = X86ISD::OR;
15490 return SDValue(Op.getNode(), 1);
15496 // If we found that truncation is beneficial, perform the truncation and
15498 if (NeedTruncation) {
15499 EVT VT = Op.getValueType();
15500 SDValue WideVal = Op->getOperand(0);
15501 EVT WideVT = WideVal.getValueType();
15502 unsigned ConvertedOp = 0;
15503 // Use a target machine opcode to prevent further DAGCombine
15504 // optimizations that may separate the arithmetic operations
15505 // from the setcc node.
15506 switch (WideVal.getOpcode()) {
15508 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
15509 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
15510 case ISD::AND: ConvertedOp = X86ISD::AND; break;
15511 case ISD::OR: ConvertedOp = X86ISD::OR; break;
15512 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
15516 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15517 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
15518 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
15519 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
15520 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
15526 // Emit a CMP with 0, which is the TEST pattern.
15527 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
15528 DAG.getConstant(0, Op.getValueType()));
15530 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15531 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
15533 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
15534 DAG.ReplaceAllUsesWith(Op, New);
15535 return SDValue(New.getNode(), 1);
15538 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
15540 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
15541 SDLoc dl, SelectionDAG &DAG) const {
15542 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
15543 if (C->getAPIntValue() == 0)
15544 return EmitTest(Op0, X86CC, dl, DAG);
15546 if (Op0.getValueType() == MVT::i1)
15547 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
15550 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
15551 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
15552 // Do the comparison at i32 if it's smaller, besides the Atom case.
15553 // This avoids subregister aliasing issues. Keep the smaller reference
15554 // if we're optimizing for size, however, as that'll allow better folding
15555 // of memory operations.
15556 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
15557 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
15558 Attribute::MinSize) &&
15559 !Subtarget->isAtom()) {
15560 unsigned ExtendOp =
15561 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
15562 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
15563 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
15565 // Use SUB instead of CMP to enable CSE between SUB and CMP.
15566 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
15567 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
15569 return SDValue(Sub.getNode(), 1);
15571 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
15574 /// Convert a comparison if required by the subtarget.
15575 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
15576 SelectionDAG &DAG) const {
15577 // If the subtarget does not support the FUCOMI instruction, floating-point
15578 // comparisons have to be converted.
15579 if (Subtarget->hasCMov() ||
15580 Cmp.getOpcode() != X86ISD::CMP ||
15581 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
15582 !Cmp.getOperand(1).getValueType().isFloatingPoint())
15585 // The instruction selector will select an FUCOM instruction instead of
15586 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
15587 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
15588 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
15590 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
15591 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
15592 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
15593 DAG.getConstant(8, MVT::i8));
15594 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
15595 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
15598 /// The minimum architected relative accuracy is 2^-12. We need one
15599 /// Newton-Raphson step to have a good float result (24 bits of precision).
15600 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
15601 DAGCombinerInfo &DCI,
15602 unsigned &RefinementSteps,
15603 bool &UseOneConstNR) const {
15604 // FIXME: We should use instruction latency models to calculate the cost of
15605 // each potential sequence, but this is very hard to do reliably because
15606 // at least Intel's Core* chips have variable timing based on the number of
15607 // significant digits in the divisor and/or sqrt operand.
15608 if (!Subtarget->useSqrtEst())
15611 EVT VT = Op.getValueType();
15613 // SSE1 has rsqrtss and rsqrtps.
15614 // TODO: Add support for AVX512 (v16f32).
15615 // It is likely not profitable to do this for f64 because a double-precision
15616 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
15617 // instructions: convert to single, rsqrtss, convert back to double, refine
15618 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
15619 // along with FMA, this could be a throughput win.
15620 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15621 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15622 RefinementSteps = 1;
15623 UseOneConstNR = false;
15624 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
15629 /// The minimum architected relative accuracy is 2^-12. We need one
15630 /// Newton-Raphson step to have a good float result (24 bits of precision).
15631 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
15632 DAGCombinerInfo &DCI,
15633 unsigned &RefinementSteps) const {
15634 // FIXME: We should use instruction latency models to calculate the cost of
15635 // each potential sequence, but this is very hard to do reliably because
15636 // at least Intel's Core* chips have variable timing based on the number of
15637 // significant digits in the divisor.
15638 if (!Subtarget->useReciprocalEst())
15641 EVT VT = Op.getValueType();
15643 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
15644 // TODO: Add support for AVX512 (v16f32).
15645 // It is likely not profitable to do this for f64 because a double-precision
15646 // reciprocal estimate with refinement on x86 prior to FMA requires
15647 // 15 instructions: convert to single, rcpss, convert back to double, refine
15648 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
15649 // along with FMA, this could be a throughput win.
15650 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15651 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15652 RefinementSteps = ReciprocalEstimateRefinementSteps;
15653 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
15658 static bool isAllOnes(SDValue V) {
15659 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
15660 return C && C->isAllOnesValue();
15663 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
15664 /// if it's possible.
15665 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
15666 SDLoc dl, SelectionDAG &DAG) const {
15667 SDValue Op0 = And.getOperand(0);
15668 SDValue Op1 = And.getOperand(1);
15669 if (Op0.getOpcode() == ISD::TRUNCATE)
15670 Op0 = Op0.getOperand(0);
15671 if (Op1.getOpcode() == ISD::TRUNCATE)
15672 Op1 = Op1.getOperand(0);
15675 if (Op1.getOpcode() == ISD::SHL)
15676 std::swap(Op0, Op1);
15677 if (Op0.getOpcode() == ISD::SHL) {
15678 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
15679 if (And00C->getZExtValue() == 1) {
15680 // If we looked past a truncate, check that it's only truncating away
15682 unsigned BitWidth = Op0.getValueSizeInBits();
15683 unsigned AndBitWidth = And.getValueSizeInBits();
15684 if (BitWidth > AndBitWidth) {
15686 DAG.computeKnownBits(Op0, Zeros, Ones);
15687 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
15691 RHS = Op0.getOperand(1);
15693 } else if (Op1.getOpcode() == ISD::Constant) {
15694 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
15695 uint64_t AndRHSVal = AndRHS->getZExtValue();
15696 SDValue AndLHS = Op0;
15698 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
15699 LHS = AndLHS.getOperand(0);
15700 RHS = AndLHS.getOperand(1);
15703 // Use BT if the immediate can't be encoded in a TEST instruction.
15704 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
15706 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
15710 if (LHS.getNode()) {
15711 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
15712 // instruction. Since the shift amount is in-range-or-undefined, we know
15713 // that doing a bittest on the i32 value is ok. We extend to i32 because
15714 // the encoding for the i16 version is larger than the i32 version.
15715 // Also promote i16 to i32 for performance / code size reason.
15716 if (LHS.getValueType() == MVT::i8 ||
15717 LHS.getValueType() == MVT::i16)
15718 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
15720 // If the operand types disagree, extend the shift amount to match. Since
15721 // BT ignores high bits (like shifts) we can use anyextend.
15722 if (LHS.getValueType() != RHS.getValueType())
15723 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
15725 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
15726 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
15727 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15728 DAG.getConstant(Cond, MVT::i8), BT);
15734 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
15736 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
15741 // SSE Condition code mapping:
15750 switch (SetCCOpcode) {
15751 default: llvm_unreachable("Unexpected SETCC condition");
15753 case ISD::SETEQ: SSECC = 0; break;
15755 case ISD::SETGT: Swap = true; // Fallthrough
15757 case ISD::SETOLT: SSECC = 1; break;
15759 case ISD::SETGE: Swap = true; // Fallthrough
15761 case ISD::SETOLE: SSECC = 2; break;
15762 case ISD::SETUO: SSECC = 3; break;
15764 case ISD::SETNE: SSECC = 4; break;
15765 case ISD::SETULE: Swap = true; // Fallthrough
15766 case ISD::SETUGE: SSECC = 5; break;
15767 case ISD::SETULT: Swap = true; // Fallthrough
15768 case ISD::SETUGT: SSECC = 6; break;
15769 case ISD::SETO: SSECC = 7; break;
15771 case ISD::SETONE: SSECC = 8; break;
15774 std::swap(Op0, Op1);
15779 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
15780 // ones, and then concatenate the result back.
15781 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
15782 MVT VT = Op.getSimpleValueType();
15784 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
15785 "Unsupported value type for operation");
15787 unsigned NumElems = VT.getVectorNumElements();
15789 SDValue CC = Op.getOperand(2);
15791 // Extract the LHS vectors
15792 SDValue LHS = Op.getOperand(0);
15793 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15794 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15796 // Extract the RHS vectors
15797 SDValue RHS = Op.getOperand(1);
15798 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15799 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15801 // Issue the operation on the smaller types and concatenate the result back
15802 MVT EltVT = VT.getVectorElementType();
15803 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15804 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15805 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
15806 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
15809 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
15810 const X86Subtarget *Subtarget) {
15811 SDValue Op0 = Op.getOperand(0);
15812 SDValue Op1 = Op.getOperand(1);
15813 SDValue CC = Op.getOperand(2);
15814 MVT VT = Op.getSimpleValueType();
15817 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
15818 Op.getValueType().getScalarType() == MVT::i1 &&
15819 "Cannot set masked compare for this operation");
15821 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15823 bool Unsigned = false;
15826 switch (SetCCOpcode) {
15827 default: llvm_unreachable("Unexpected SETCC condition");
15828 case ISD::SETNE: SSECC = 4; break;
15829 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
15830 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
15831 case ISD::SETLT: Swap = true; //fall-through
15832 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
15833 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
15834 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
15835 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
15836 case ISD::SETULE: Unsigned = true; //fall-through
15837 case ISD::SETLE: SSECC = 2; break;
15841 std::swap(Op0, Op1);
15843 return DAG.getNode(Opc, dl, VT, Op0, Op1);
15844 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
15845 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15846 DAG.getConstant(SSECC, MVT::i8));
15849 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
15850 /// operand \p Op1. If non-trivial (for example because it's not constant)
15851 /// return an empty value.
15852 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
15854 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
15858 MVT VT = Op1.getSimpleValueType();
15859 MVT EVT = VT.getVectorElementType();
15860 unsigned n = VT.getVectorNumElements();
15861 SmallVector<SDValue, 8> ULTOp1;
15863 for (unsigned i = 0; i < n; ++i) {
15864 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
15865 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
15868 // Avoid underflow.
15869 APInt Val = Elt->getAPIntValue();
15873 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
15876 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
15879 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
15880 SelectionDAG &DAG) {
15881 SDValue Op0 = Op.getOperand(0);
15882 SDValue Op1 = Op.getOperand(1);
15883 SDValue CC = Op.getOperand(2);
15884 MVT VT = Op.getSimpleValueType();
15885 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15886 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
15891 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
15892 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
15895 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
15896 unsigned Opc = X86ISD::CMPP;
15897 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
15898 assert(VT.getVectorNumElements() <= 16);
15899 Opc = X86ISD::CMPM;
15901 // In the two special cases we can't handle, emit two comparisons.
15904 unsigned CombineOpc;
15905 if (SetCCOpcode == ISD::SETUEQ) {
15906 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
15908 assert(SetCCOpcode == ISD::SETONE);
15909 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
15912 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15913 DAG.getConstant(CC0, MVT::i8));
15914 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15915 DAG.getConstant(CC1, MVT::i8));
15916 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
15918 // Handle all other FP comparisons here.
15919 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15920 DAG.getConstant(SSECC, MVT::i8));
15923 // Break 256-bit integer vector compare into smaller ones.
15924 if (VT.is256BitVector() && !Subtarget->hasInt256())
15925 return Lower256IntVSETCC(Op, DAG);
15927 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
15928 EVT OpVT = Op1.getValueType();
15929 if (Subtarget->hasAVX512()) {
15930 if (Op1.getValueType().is512BitVector() ||
15931 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
15932 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
15933 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
15935 // In AVX-512 architecture setcc returns mask with i1 elements,
15936 // But there is no compare instruction for i8 and i16 elements in KNL.
15937 // We are not talking about 512-bit operands in this case, these
15938 // types are illegal.
15940 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
15941 OpVT.getVectorElementType().getSizeInBits() >= 8))
15942 return DAG.getNode(ISD::TRUNCATE, dl, VT,
15943 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
15946 // We are handling one of the integer comparisons here. Since SSE only has
15947 // GT and EQ comparisons for integer, swapping operands and multiple
15948 // operations may be required for some comparisons.
15950 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
15951 bool Subus = false;
15953 switch (SetCCOpcode) {
15954 default: llvm_unreachable("Unexpected SETCC condition");
15955 case ISD::SETNE: Invert = true;
15956 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
15957 case ISD::SETLT: Swap = true;
15958 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
15959 case ISD::SETGE: Swap = true;
15960 case ISD::SETLE: Opc = X86ISD::PCMPGT;
15961 Invert = true; break;
15962 case ISD::SETULT: Swap = true;
15963 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
15964 FlipSigns = true; break;
15965 case ISD::SETUGE: Swap = true;
15966 case ISD::SETULE: Opc = X86ISD::PCMPGT;
15967 FlipSigns = true; Invert = true; break;
15970 // Special case: Use min/max operations for SETULE/SETUGE
15971 MVT VET = VT.getVectorElementType();
15973 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
15974 || (Subtarget->hasSSE2() && (VET == MVT::i8));
15977 switch (SetCCOpcode) {
15979 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
15980 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
15983 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
15986 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
15987 if (!MinMax && hasSubus) {
15988 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
15990 // t = psubus Op0, Op1
15991 // pcmpeq t, <0..0>
15992 switch (SetCCOpcode) {
15994 case ISD::SETULT: {
15995 // If the comparison is against a constant we can turn this into a
15996 // setule. With psubus, setule does not require a swap. This is
15997 // beneficial because the constant in the register is no longer
15998 // destructed as the destination so it can be hoisted out of a loop.
15999 // Only do this pre-AVX since vpcmp* is no longer destructive.
16000 if (Subtarget->hasAVX())
16002 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
16003 if (ULEOp1.getNode()) {
16005 Subus = true; Invert = false; Swap = false;
16009 // Psubus is better than flip-sign because it requires no inversion.
16010 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
16011 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
16015 Opc = X86ISD::SUBUS;
16021 std::swap(Op0, Op1);
16023 // Check that the operation in question is available (most are plain SSE2,
16024 // but PCMPGTQ and PCMPEQQ have different requirements).
16025 if (VT == MVT::v2i64) {
16026 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
16027 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
16029 // First cast everything to the right type.
16030 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
16031 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
16033 // Since SSE has no unsigned integer comparisons, we need to flip the sign
16034 // bits of the inputs before performing those operations. The lower
16035 // compare is always unsigned.
16038 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
16040 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
16041 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
16042 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
16043 Sign, Zero, Sign, Zero);
16045 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
16046 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
16048 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
16049 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
16050 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
16052 // Create masks for only the low parts/high parts of the 64 bit integers.
16053 static const int MaskHi[] = { 1, 1, 3, 3 };
16054 static const int MaskLo[] = { 0, 0, 2, 2 };
16055 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
16056 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
16057 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
16059 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
16060 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
16063 Result = DAG.getNOT(dl, Result, MVT::v4i32);
16065 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16068 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
16069 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
16070 // pcmpeqd + pshufd + pand.
16071 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
16073 // First cast everything to the right type.
16074 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
16075 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
16078 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
16080 // Make sure the lower and upper halves are both all-ones.
16081 static const int Mask[] = { 1, 0, 3, 2 };
16082 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
16083 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
16086 Result = DAG.getNOT(dl, Result, MVT::v4i32);
16088 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16092 // Since SSE has no unsigned integer comparisons, we need to flip the sign
16093 // bits of the inputs before performing those operations.
16095 EVT EltVT = VT.getVectorElementType();
16096 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
16097 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
16098 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
16101 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
16103 // If the logical-not of the result is required, perform that now.
16105 Result = DAG.getNOT(dl, Result, VT);
16108 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
16111 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
16112 getZeroVector(VT, Subtarget, DAG, dl));
16117 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
16119 MVT VT = Op.getSimpleValueType();
16121 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
16123 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
16124 && "SetCC type must be 8-bit or 1-bit integer");
16125 SDValue Op0 = Op.getOperand(0);
16126 SDValue Op1 = Op.getOperand(1);
16128 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
16130 // Optimize to BT if possible.
16131 // Lower (X & (1 << N)) == 0 to BT(X, N).
16132 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
16133 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
16134 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
16135 Op1.getOpcode() == ISD::Constant &&
16136 cast<ConstantSDNode>(Op1)->isNullValue() &&
16137 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
16138 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
16139 if (NewSetCC.getNode()) {
16141 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
16146 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
16148 if (Op1.getOpcode() == ISD::Constant &&
16149 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
16150 cast<ConstantSDNode>(Op1)->isNullValue()) &&
16151 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
16153 // If the input is a setcc, then reuse the input setcc or use a new one with
16154 // the inverted condition.
16155 if (Op0.getOpcode() == X86ISD::SETCC) {
16156 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
16157 bool Invert = (CC == ISD::SETNE) ^
16158 cast<ConstantSDNode>(Op1)->isNullValue();
16162 CCode = X86::GetOppositeBranchCondition(CCode);
16163 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16164 DAG.getConstant(CCode, MVT::i8),
16165 Op0.getOperand(1));
16167 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
16171 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
16172 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
16173 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
16175 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
16176 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
16179 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
16180 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
16181 if (X86CC == X86::COND_INVALID)
16184 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
16185 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
16186 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16187 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
16189 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
16193 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
16194 static bool isX86LogicalCmp(SDValue Op) {
16195 unsigned Opc = Op.getNode()->getOpcode();
16196 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
16197 Opc == X86ISD::SAHF)
16199 if (Op.getResNo() == 1 &&
16200 (Opc == X86ISD::ADD ||
16201 Opc == X86ISD::SUB ||
16202 Opc == X86ISD::ADC ||
16203 Opc == X86ISD::SBB ||
16204 Opc == X86ISD::SMUL ||
16205 Opc == X86ISD::UMUL ||
16206 Opc == X86ISD::INC ||
16207 Opc == X86ISD::DEC ||
16208 Opc == X86ISD::OR ||
16209 Opc == X86ISD::XOR ||
16210 Opc == X86ISD::AND))
16213 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
16219 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
16220 if (V.getOpcode() != ISD::TRUNCATE)
16223 SDValue VOp0 = V.getOperand(0);
16224 unsigned InBits = VOp0.getValueSizeInBits();
16225 unsigned Bits = V.getValueSizeInBits();
16226 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
16229 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
16230 bool addTest = true;
16231 SDValue Cond = Op.getOperand(0);
16232 SDValue Op1 = Op.getOperand(1);
16233 SDValue Op2 = Op.getOperand(2);
16235 EVT VT = Op1.getValueType();
16238 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
16239 // are available. Otherwise fp cmovs get lowered into a less efficient branch
16240 // sequence later on.
16241 if (Cond.getOpcode() == ISD::SETCC &&
16242 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
16243 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
16244 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
16245 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
16246 int SSECC = translateX86FSETCC(
16247 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
16250 if (Subtarget->hasAVX512()) {
16251 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
16252 DAG.getConstant(SSECC, MVT::i8));
16253 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
16255 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
16256 DAG.getConstant(SSECC, MVT::i8));
16257 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
16258 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
16259 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
16263 if (Cond.getOpcode() == ISD::SETCC) {
16264 SDValue NewCond = LowerSETCC(Cond, DAG);
16265 if (NewCond.getNode())
16269 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
16270 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
16271 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
16272 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
16273 if (Cond.getOpcode() == X86ISD::SETCC &&
16274 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
16275 isZero(Cond.getOperand(1).getOperand(1))) {
16276 SDValue Cmp = Cond.getOperand(1);
16278 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
16280 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
16281 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
16282 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
16284 SDValue CmpOp0 = Cmp.getOperand(0);
16285 // Apply further optimizations for special cases
16286 // (select (x != 0), -1, 0) -> neg & sbb
16287 // (select (x == 0), 0, -1) -> neg & sbb
16288 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
16289 if (YC->isNullValue() &&
16290 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
16291 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
16292 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
16293 DAG.getConstant(0, CmpOp0.getValueType()),
16295 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16296 DAG.getConstant(X86::COND_B, MVT::i8),
16297 SDValue(Neg.getNode(), 1));
16301 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
16302 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
16303 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16305 SDValue Res = // Res = 0 or -1.
16306 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16307 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
16309 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
16310 Res = DAG.getNOT(DL, Res, Res.getValueType());
16312 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
16313 if (!N2C || !N2C->isNullValue())
16314 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
16319 // Look past (and (setcc_carry (cmp ...)), 1).
16320 if (Cond.getOpcode() == ISD::AND &&
16321 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
16322 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
16323 if (C && C->getAPIntValue() == 1)
16324 Cond = Cond.getOperand(0);
16327 // If condition flag is set by a X86ISD::CMP, then use it as the condition
16328 // setting operand in place of the X86ISD::SETCC.
16329 unsigned CondOpcode = Cond.getOpcode();
16330 if (CondOpcode == X86ISD::SETCC ||
16331 CondOpcode == X86ISD::SETCC_CARRY) {
16332 CC = Cond.getOperand(0);
16334 SDValue Cmp = Cond.getOperand(1);
16335 unsigned Opc = Cmp.getOpcode();
16336 MVT VT = Op.getSimpleValueType();
16338 bool IllegalFPCMov = false;
16339 if (VT.isFloatingPoint() && !VT.isVector() &&
16340 !isScalarFPTypeInSSEReg(VT)) // FPStack?
16341 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
16343 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
16344 Opc == X86ISD::BT) { // FIXME
16348 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
16349 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
16350 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
16351 Cond.getOperand(0).getValueType() != MVT::i8)) {
16352 SDValue LHS = Cond.getOperand(0);
16353 SDValue RHS = Cond.getOperand(1);
16354 unsigned X86Opcode;
16357 switch (CondOpcode) {
16358 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
16359 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
16360 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
16361 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
16362 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
16363 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
16364 default: llvm_unreachable("unexpected overflowing operator");
16366 if (CondOpcode == ISD::UMULO)
16367 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
16370 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
16372 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
16374 if (CondOpcode == ISD::UMULO)
16375 Cond = X86Op.getValue(2);
16377 Cond = X86Op.getValue(1);
16379 CC = DAG.getConstant(X86Cond, MVT::i8);
16384 // Look pass the truncate if the high bits are known zero.
16385 if (isTruncWithZeroHighBitsInput(Cond, DAG))
16386 Cond = Cond.getOperand(0);
16388 // We know the result of AND is compared against zero. Try to match
16390 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
16391 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
16392 if (NewSetCC.getNode()) {
16393 CC = NewSetCC.getOperand(0);
16394 Cond = NewSetCC.getOperand(1);
16401 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16402 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
16405 // a < b ? -1 : 0 -> RES = ~setcc_carry
16406 // a < b ? 0 : -1 -> RES = setcc_carry
16407 // a >= b ? -1 : 0 -> RES = setcc_carry
16408 // a >= b ? 0 : -1 -> RES = ~setcc_carry
16409 if (Cond.getOpcode() == X86ISD::SUB) {
16410 Cond = ConvertCmpIfNecessary(Cond, DAG);
16411 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
16413 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
16414 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
16415 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16416 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
16417 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
16418 return DAG.getNOT(DL, Res, Res.getValueType());
16423 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
16424 // widen the cmov and push the truncate through. This avoids introducing a new
16425 // branch during isel and doesn't add any extensions.
16426 if (Op.getValueType() == MVT::i8 &&
16427 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
16428 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
16429 if (T1.getValueType() == T2.getValueType() &&
16430 // Blacklist CopyFromReg to avoid partial register stalls.
16431 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
16432 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
16433 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
16434 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
16438 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
16439 // condition is true.
16440 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
16441 SDValue Ops[] = { Op2, Op1, CC, Cond };
16442 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
16445 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
16446 SelectionDAG &DAG) {
16447 MVT VT = Op->getSimpleValueType(0);
16448 SDValue In = Op->getOperand(0);
16449 MVT InVT = In.getSimpleValueType();
16450 MVT VTElt = VT.getVectorElementType();
16451 MVT InVTElt = InVT.getVectorElementType();
16455 if ((InVTElt == MVT::i1) &&
16456 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
16457 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
16459 ((Subtarget->hasBWI() && VT.is512BitVector() &&
16460 VTElt.getSizeInBits() <= 16)) ||
16462 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
16463 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
16465 ((Subtarget->hasDQI() && VT.is512BitVector() &&
16466 VTElt.getSizeInBits() >= 32))))
16467 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16469 unsigned int NumElts = VT.getVectorNumElements();
16471 if (NumElts != 8 && NumElts != 16)
16474 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
16475 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
16476 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
16477 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16480 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16481 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
16483 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
16484 Constant *C = ConstantInt::get(*DAG.getContext(),
16485 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
16487 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
16488 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
16489 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
16490 MachinePointerInfo::getConstantPool(),
16491 false, false, false, Alignment);
16492 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
16493 if (VT.is512BitVector())
16495 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
16498 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
16499 SelectionDAG &DAG) {
16500 MVT VT = Op->getSimpleValueType(0);
16501 SDValue In = Op->getOperand(0);
16502 MVT InVT = In.getSimpleValueType();
16505 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
16506 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
16508 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
16509 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
16510 (VT != MVT::v16i16 || InVT != MVT::v16i8))
16513 if (Subtarget->hasInt256())
16514 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16516 // Optimize vectors in AVX mode
16517 // Sign extend v8i16 to v8i32 and
16520 // Divide input vector into two parts
16521 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
16522 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
16523 // concat the vectors to original VT
16525 unsigned NumElems = InVT.getVectorNumElements();
16526 SDValue Undef = DAG.getUNDEF(InVT);
16528 SmallVector<int,8> ShufMask1(NumElems, -1);
16529 for (unsigned i = 0; i != NumElems/2; ++i)
16532 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
16534 SmallVector<int,8> ShufMask2(NumElems, -1);
16535 for (unsigned i = 0; i != NumElems/2; ++i)
16536 ShufMask2[i] = i + NumElems/2;
16538 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
16540 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
16541 VT.getVectorNumElements()/2);
16543 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
16544 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
16546 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
16549 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
16550 // may emit an illegal shuffle but the expansion is still better than scalar
16551 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
16552 // we'll emit a shuffle and a arithmetic shift.
16553 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
16554 // TODO: It is possible to support ZExt by zeroing the undef values during
16555 // the shuffle phase or after the shuffle.
16556 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
16557 SelectionDAG &DAG) {
16558 MVT RegVT = Op.getSimpleValueType();
16559 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
16560 assert(RegVT.isInteger() &&
16561 "We only custom lower integer vector sext loads.");
16563 // Nothing useful we can do without SSE2 shuffles.
16564 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
16566 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
16568 EVT MemVT = Ld->getMemoryVT();
16569 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16570 unsigned RegSz = RegVT.getSizeInBits();
16572 ISD::LoadExtType Ext = Ld->getExtensionType();
16574 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
16575 && "Only anyext and sext are currently implemented.");
16576 assert(MemVT != RegVT && "Cannot extend to the same type");
16577 assert(MemVT.isVector() && "Must load a vector from memory");
16579 unsigned NumElems = RegVT.getVectorNumElements();
16580 unsigned MemSz = MemVT.getSizeInBits();
16581 assert(RegSz > MemSz && "Register size must be greater than the mem size");
16583 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
16584 // The only way in which we have a legal 256-bit vector result but not the
16585 // integer 256-bit operations needed to directly lower a sextload is if we
16586 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
16587 // a 128-bit vector and a normal sign_extend to 256-bits that should get
16588 // correctly legalized. We do this late to allow the canonical form of
16589 // sextload to persist throughout the rest of the DAG combiner -- it wants
16590 // to fold together any extensions it can, and so will fuse a sign_extend
16591 // of an sextload into a sextload targeting a wider value.
16593 if (MemSz == 128) {
16594 // Just switch this to a normal load.
16595 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
16596 "it must be a legal 128-bit vector "
16598 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
16599 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
16600 Ld->isInvariant(), Ld->getAlignment());
16602 assert(MemSz < 128 &&
16603 "Can't extend a type wider than 128 bits to a 256 bit vector!");
16604 // Do an sext load to a 128-bit vector type. We want to use the same
16605 // number of elements, but elements half as wide. This will end up being
16606 // recursively lowered by this routine, but will succeed as we definitely
16607 // have all the necessary features if we're using AVX1.
16609 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
16610 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
16612 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
16613 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
16614 Ld->isNonTemporal(), Ld->isInvariant(),
16615 Ld->getAlignment());
16618 // Replace chain users with the new chain.
16619 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
16620 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
16622 // Finally, do a normal sign-extend to the desired register.
16623 return DAG.getSExtOrTrunc(Load, dl, RegVT);
16626 // All sizes must be a power of two.
16627 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
16628 "Non-power-of-two elements are not custom lowered!");
16630 // Attempt to load the original value using scalar loads.
16631 // Find the largest scalar type that divides the total loaded size.
16632 MVT SclrLoadTy = MVT::i8;
16633 for (MVT Tp : MVT::integer_valuetypes()) {
16634 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
16639 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
16640 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
16642 SclrLoadTy = MVT::f64;
16644 // Calculate the number of scalar loads that we need to perform
16645 // in order to load our vector from memory.
16646 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
16648 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
16649 "Can only lower sext loads with a single scalar load!");
16651 unsigned loadRegZize = RegSz;
16652 if (Ext == ISD::SEXTLOAD && RegSz == 256)
16655 // Represent our vector as a sequence of elements which are the
16656 // largest scalar that we can load.
16657 EVT LoadUnitVecVT = EVT::getVectorVT(
16658 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
16660 // Represent the data using the same element type that is stored in
16661 // memory. In practice, we ''widen'' MemVT.
16663 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
16664 loadRegZize / MemVT.getScalarType().getSizeInBits());
16666 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
16667 "Invalid vector type");
16669 // We can't shuffle using an illegal type.
16670 assert(TLI.isTypeLegal(WideVecVT) &&
16671 "We only lower types that form legal widened vector types");
16673 SmallVector<SDValue, 8> Chains;
16674 SDValue Ptr = Ld->getBasePtr();
16675 SDValue Increment =
16676 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
16677 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
16679 for (unsigned i = 0; i < NumLoads; ++i) {
16680 // Perform a single load.
16681 SDValue ScalarLoad =
16682 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
16683 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
16684 Ld->getAlignment());
16685 Chains.push_back(ScalarLoad.getValue(1));
16686 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
16687 // another round of DAGCombining.
16689 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
16691 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
16692 ScalarLoad, DAG.getIntPtrConstant(i));
16694 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16697 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
16699 // Bitcast the loaded value to a vector of the original element type, in
16700 // the size of the target vector type.
16701 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
16702 unsigned SizeRatio = RegSz / MemSz;
16704 if (Ext == ISD::SEXTLOAD) {
16705 // If we have SSE4.1, we can directly emit a VSEXT node.
16706 if (Subtarget->hasSSE41()) {
16707 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
16708 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16712 // Otherwise we'll shuffle the small elements in the high bits of the
16713 // larger type and perform an arithmetic shift. If the shift is not legal
16714 // it's better to scalarize.
16715 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
16716 "We can't implement a sext load without an arithmetic right shift!");
16718 // Redistribute the loaded elements into the different locations.
16719 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16720 for (unsigned i = 0; i != NumElems; ++i)
16721 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
16723 SDValue Shuff = DAG.getVectorShuffle(
16724 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16726 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16728 // Build the arithmetic shift.
16729 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
16730 MemVT.getVectorElementType().getSizeInBits();
16732 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
16734 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16738 // Redistribute the loaded elements into the different locations.
16739 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16740 for (unsigned i = 0; i != NumElems; ++i)
16741 ShuffleVec[i * SizeRatio] = i;
16743 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
16744 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16746 // Bitcast to the requested type.
16747 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16748 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16752 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
16753 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
16754 // from the AND / OR.
16755 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
16756 Opc = Op.getOpcode();
16757 if (Opc != ISD::OR && Opc != ISD::AND)
16759 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16760 Op.getOperand(0).hasOneUse() &&
16761 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
16762 Op.getOperand(1).hasOneUse());
16765 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
16766 // 1 and that the SETCC node has a single use.
16767 static bool isXor1OfSetCC(SDValue Op) {
16768 if (Op.getOpcode() != ISD::XOR)
16770 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
16771 if (N1C && N1C->getAPIntValue() == 1) {
16772 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16773 Op.getOperand(0).hasOneUse();
16778 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
16779 bool addTest = true;
16780 SDValue Chain = Op.getOperand(0);
16781 SDValue Cond = Op.getOperand(1);
16782 SDValue Dest = Op.getOperand(2);
16785 bool Inverted = false;
16787 if (Cond.getOpcode() == ISD::SETCC) {
16788 // Check for setcc([su]{add,sub,mul}o == 0).
16789 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
16790 isa<ConstantSDNode>(Cond.getOperand(1)) &&
16791 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
16792 Cond.getOperand(0).getResNo() == 1 &&
16793 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
16794 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
16795 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
16796 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
16797 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
16798 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
16800 Cond = Cond.getOperand(0);
16802 SDValue NewCond = LowerSETCC(Cond, DAG);
16803 if (NewCond.getNode())
16808 // FIXME: LowerXALUO doesn't handle these!!
16809 else if (Cond.getOpcode() == X86ISD::ADD ||
16810 Cond.getOpcode() == X86ISD::SUB ||
16811 Cond.getOpcode() == X86ISD::SMUL ||
16812 Cond.getOpcode() == X86ISD::UMUL)
16813 Cond = LowerXALUO(Cond, DAG);
16816 // Look pass (and (setcc_carry (cmp ...)), 1).
16817 if (Cond.getOpcode() == ISD::AND &&
16818 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
16819 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
16820 if (C && C->getAPIntValue() == 1)
16821 Cond = Cond.getOperand(0);
16824 // If condition flag is set by a X86ISD::CMP, then use it as the condition
16825 // setting operand in place of the X86ISD::SETCC.
16826 unsigned CondOpcode = Cond.getOpcode();
16827 if (CondOpcode == X86ISD::SETCC ||
16828 CondOpcode == X86ISD::SETCC_CARRY) {
16829 CC = Cond.getOperand(0);
16831 SDValue Cmp = Cond.getOperand(1);
16832 unsigned Opc = Cmp.getOpcode();
16833 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
16834 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
16838 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
16842 // These can only come from an arithmetic instruction with overflow,
16843 // e.g. SADDO, UADDO.
16844 Cond = Cond.getNode()->getOperand(1);
16850 CondOpcode = Cond.getOpcode();
16851 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
16852 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
16853 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
16854 Cond.getOperand(0).getValueType() != MVT::i8)) {
16855 SDValue LHS = Cond.getOperand(0);
16856 SDValue RHS = Cond.getOperand(1);
16857 unsigned X86Opcode;
16860 // Keep this in sync with LowerXALUO, otherwise we might create redundant
16861 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
16863 switch (CondOpcode) {
16864 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
16866 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16868 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
16871 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
16872 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
16874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16876 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
16879 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
16880 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
16881 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
16882 default: llvm_unreachable("unexpected overflowing operator");
16885 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
16886 if (CondOpcode == ISD::UMULO)
16887 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
16890 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
16892 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
16894 if (CondOpcode == ISD::UMULO)
16895 Cond = X86Op.getValue(2);
16897 Cond = X86Op.getValue(1);
16899 CC = DAG.getConstant(X86Cond, MVT::i8);
16903 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
16904 SDValue Cmp = Cond.getOperand(0).getOperand(1);
16905 if (CondOpc == ISD::OR) {
16906 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
16907 // two branches instead of an explicit OR instruction with a
16909 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16910 isX86LogicalCmp(Cmp)) {
16911 CC = Cond.getOperand(0).getOperand(0);
16912 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16913 Chain, Dest, CC, Cmp);
16914 CC = Cond.getOperand(1).getOperand(0);
16918 } else { // ISD::AND
16919 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
16920 // two branches instead of an explicit AND instruction with a
16921 // separate test. However, we only do this if this block doesn't
16922 // have a fall-through edge, because this requires an explicit
16923 // jmp when the condition is false.
16924 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16925 isX86LogicalCmp(Cmp) &&
16926 Op.getNode()->hasOneUse()) {
16927 X86::CondCode CCode =
16928 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16929 CCode = X86::GetOppositeBranchCondition(CCode);
16930 CC = DAG.getConstant(CCode, MVT::i8);
16931 SDNode *User = *Op.getNode()->use_begin();
16932 // Look for an unconditional branch following this conditional branch.
16933 // We need this because we need to reverse the successors in order
16934 // to implement FCMP_OEQ.
16935 if (User->getOpcode() == ISD::BR) {
16936 SDValue FalseBB = User->getOperand(1);
16938 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16939 assert(NewBR == User);
16943 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16944 Chain, Dest, CC, Cmp);
16945 X86::CondCode CCode =
16946 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
16947 CCode = X86::GetOppositeBranchCondition(CCode);
16948 CC = DAG.getConstant(CCode, MVT::i8);
16954 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
16955 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
16956 // It should be transformed during dag combiner except when the condition
16957 // is set by a arithmetics with overflow node.
16958 X86::CondCode CCode =
16959 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16960 CCode = X86::GetOppositeBranchCondition(CCode);
16961 CC = DAG.getConstant(CCode, MVT::i8);
16962 Cond = Cond.getOperand(0).getOperand(1);
16964 } else if (Cond.getOpcode() == ISD::SETCC &&
16965 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
16966 // For FCMP_OEQ, we can emit
16967 // two branches instead of an explicit AND instruction with a
16968 // separate test. However, we only do this if this block doesn't
16969 // have a fall-through edge, because this requires an explicit
16970 // jmp when the condition is false.
16971 if (Op.getNode()->hasOneUse()) {
16972 SDNode *User = *Op.getNode()->use_begin();
16973 // Look for an unconditional branch following this conditional branch.
16974 // We need this because we need to reverse the successors in order
16975 // to implement FCMP_OEQ.
16976 if (User->getOpcode() == ISD::BR) {
16977 SDValue FalseBB = User->getOperand(1);
16979 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16980 assert(NewBR == User);
16984 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
16985 Cond.getOperand(0), Cond.getOperand(1));
16986 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16987 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16988 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16989 Chain, Dest, CC, Cmp);
16990 CC = DAG.getConstant(X86::COND_P, MVT::i8);
16995 } else if (Cond.getOpcode() == ISD::SETCC &&
16996 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
16997 // For FCMP_UNE, we can emit
16998 // two branches instead of an explicit AND instruction with a
16999 // separate test. However, we only do this if this block doesn't
17000 // have a fall-through edge, because this requires an explicit
17001 // jmp when the condition is false.
17002 if (Op.getNode()->hasOneUse()) {
17003 SDNode *User = *Op.getNode()->use_begin();
17004 // Look for an unconditional branch following this conditional branch.
17005 // We need this because we need to reverse the successors in order
17006 // to implement FCMP_UNE.
17007 if (User->getOpcode() == ISD::BR) {
17008 SDValue FalseBB = User->getOperand(1);
17010 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
17011 assert(NewBR == User);
17014 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
17015 Cond.getOperand(0), Cond.getOperand(1));
17016 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
17017 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
17018 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
17019 Chain, Dest, CC, Cmp);
17020 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
17030 // Look pass the truncate if the high bits are known zero.
17031 if (isTruncWithZeroHighBitsInput(Cond, DAG))
17032 Cond = Cond.getOperand(0);
17034 // We know the result of AND is compared against zero. Try to match
17036 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
17037 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
17038 if (NewSetCC.getNode()) {
17039 CC = NewSetCC.getOperand(0);
17040 Cond = NewSetCC.getOperand(1);
17047 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
17048 CC = DAG.getConstant(X86Cond, MVT::i8);
17049 Cond = EmitTest(Cond, X86Cond, dl, DAG);
17051 Cond = ConvertCmpIfNecessary(Cond, DAG);
17052 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
17053 Chain, Dest, CC, Cond);
17056 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
17057 // Calls to _alloca are needed to probe the stack when allocating more than 4k
17058 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
17059 // that the guard pages used by the OS virtual memory manager are allocated in
17060 // correct sequence.
17062 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
17063 SelectionDAG &DAG) const {
17064 MachineFunction &MF = DAG.getMachineFunction();
17065 bool SplitStack = MF.shouldSplitStack();
17066 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
17071 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17072 SDNode* Node = Op.getNode();
17074 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
17075 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
17076 " not tell us which reg is the stack pointer!");
17077 EVT VT = Node->getValueType(0);
17078 SDValue Tmp1 = SDValue(Node, 0);
17079 SDValue Tmp2 = SDValue(Node, 1);
17080 SDValue Tmp3 = Node->getOperand(2);
17081 SDValue Chain = Tmp1.getOperand(0);
17083 // Chain the dynamic stack allocation so that it doesn't modify the stack
17084 // pointer when other instructions are using the stack.
17085 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
17088 SDValue Size = Tmp2.getOperand(1);
17089 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
17090 Chain = SP.getValue(1);
17091 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
17092 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17093 unsigned StackAlign = TFI.getStackAlignment();
17094 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
17095 if (Align > StackAlign)
17096 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
17097 DAG.getConstant(-(uint64_t)Align, VT));
17098 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
17100 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
17101 DAG.getIntPtrConstant(0, true), SDValue(),
17104 SDValue Ops[2] = { Tmp1, Tmp2 };
17105 return DAG.getMergeValues(Ops, dl);
17109 SDValue Chain = Op.getOperand(0);
17110 SDValue Size = Op.getOperand(1);
17111 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
17112 EVT VT = Op.getNode()->getValueType(0);
17114 bool Is64Bit = Subtarget->is64Bit();
17115 EVT SPTy = getPointerTy();
17118 MachineRegisterInfo &MRI = MF.getRegInfo();
17121 // The 64 bit implementation of segmented stacks needs to clobber both r10
17122 // r11. This makes it impossible to use it along with nested parameters.
17123 const Function *F = MF.getFunction();
17125 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
17127 if (I->hasNestAttr())
17128 report_fatal_error("Cannot use segmented stacks with functions that "
17129 "have nested arguments.");
17132 const TargetRegisterClass *AddrRegClass =
17133 getRegClassFor(getPointerTy());
17134 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
17135 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
17136 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
17137 DAG.getRegister(Vreg, SPTy));
17138 SDValue Ops1[2] = { Value, Chain };
17139 return DAG.getMergeValues(Ops1, dl);
17142 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
17144 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
17145 Flag = Chain.getValue(1);
17146 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
17148 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
17150 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17151 unsigned SPReg = RegInfo->getStackRegister();
17152 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
17153 Chain = SP.getValue(1);
17156 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
17157 DAG.getConstant(-(uint64_t)Align, VT));
17158 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
17161 SDValue Ops1[2] = { SP, Chain };
17162 return DAG.getMergeValues(Ops1, dl);
17166 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
17167 MachineFunction &MF = DAG.getMachineFunction();
17168 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17170 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
17173 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
17174 // vastart just stores the address of the VarArgsFrameIndex slot into the
17175 // memory location argument.
17176 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
17178 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
17179 MachinePointerInfo(SV), false, false, 0);
17183 // gp_offset (0 - 6 * 8)
17184 // fp_offset (48 - 48 + 8 * 16)
17185 // overflow_arg_area (point to parameters coming in memory).
17187 SmallVector<SDValue, 8> MemOps;
17188 SDValue FIN = Op.getOperand(1);
17190 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
17191 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
17193 FIN, MachinePointerInfo(SV), false, false, 0);
17194 MemOps.push_back(Store);
17197 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
17198 FIN, DAG.getIntPtrConstant(4));
17199 Store = DAG.getStore(Op.getOperand(0), DL,
17200 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
17202 FIN, MachinePointerInfo(SV, 4), false, false, 0);
17203 MemOps.push_back(Store);
17205 // Store ptr to overflow_arg_area
17206 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
17207 FIN, DAG.getIntPtrConstant(4));
17208 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
17210 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
17211 MachinePointerInfo(SV, 8),
17213 MemOps.push_back(Store);
17215 // Store ptr to reg_save_area.
17216 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
17217 FIN, DAG.getIntPtrConstant(8));
17218 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
17220 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
17221 MachinePointerInfo(SV, 16), false, false, 0);
17222 MemOps.push_back(Store);
17223 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
17226 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
17227 assert(Subtarget->is64Bit() &&
17228 "LowerVAARG only handles 64-bit va_arg!");
17229 assert((Subtarget->isTargetLinux() ||
17230 Subtarget->isTargetDarwin()) &&
17231 "Unhandled target in LowerVAARG");
17232 assert(Op.getNode()->getNumOperands() == 4);
17233 SDValue Chain = Op.getOperand(0);
17234 SDValue SrcPtr = Op.getOperand(1);
17235 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
17236 unsigned Align = Op.getConstantOperandVal(3);
17239 EVT ArgVT = Op.getNode()->getValueType(0);
17240 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17241 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
17244 // Decide which area this value should be read from.
17245 // TODO: Implement the AMD64 ABI in its entirety. This simple
17246 // selection mechanism works only for the basic types.
17247 if (ArgVT == MVT::f80) {
17248 llvm_unreachable("va_arg for f80 not yet implemented");
17249 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
17250 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
17251 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
17252 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
17254 llvm_unreachable("Unhandled argument type in LowerVAARG");
17257 if (ArgMode == 2) {
17258 // Sanity Check: Make sure using fp_offset makes sense.
17259 assert(!DAG.getTarget().Options.UseSoftFloat &&
17260 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
17261 Attribute::NoImplicitFloat)) &&
17262 Subtarget->hasSSE1());
17265 // Insert VAARG_64 node into the DAG
17266 // VAARG_64 returns two values: Variable Argument Address, Chain
17267 SmallVector<SDValue, 11> InstOps;
17268 InstOps.push_back(Chain);
17269 InstOps.push_back(SrcPtr);
17270 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
17271 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
17272 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
17273 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
17274 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
17275 VTs, InstOps, MVT::i64,
17276 MachinePointerInfo(SV),
17278 /*Volatile=*/false,
17280 /*WriteMem=*/true);
17281 Chain = VAARG.getValue(1);
17283 // Load the next argument and return it
17284 return DAG.getLoad(ArgVT, dl,
17287 MachinePointerInfo(),
17288 false, false, false, 0);
17291 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
17292 SelectionDAG &DAG) {
17293 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
17294 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
17295 SDValue Chain = Op.getOperand(0);
17296 SDValue DstPtr = Op.getOperand(1);
17297 SDValue SrcPtr = Op.getOperand(2);
17298 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
17299 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17302 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
17303 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
17305 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
17308 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
17309 // amount is a constant. Takes immediate version of shift as input.
17310 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
17311 SDValue SrcOp, uint64_t ShiftAmt,
17312 SelectionDAG &DAG) {
17313 MVT ElementType = VT.getVectorElementType();
17315 // Fold this packed shift into its first operand if ShiftAmt is 0.
17319 // Check for ShiftAmt >= element width
17320 if (ShiftAmt >= ElementType.getSizeInBits()) {
17321 if (Opc == X86ISD::VSRAI)
17322 ShiftAmt = ElementType.getSizeInBits() - 1;
17324 return DAG.getConstant(0, VT);
17327 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
17328 && "Unknown target vector shift-by-constant node");
17330 // Fold this packed vector shift into a build vector if SrcOp is a
17331 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
17332 if (VT == SrcOp.getSimpleValueType() &&
17333 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
17334 SmallVector<SDValue, 8> Elts;
17335 unsigned NumElts = SrcOp->getNumOperands();
17336 ConstantSDNode *ND;
17339 default: llvm_unreachable(nullptr);
17340 case X86ISD::VSHLI:
17341 for (unsigned i=0; i!=NumElts; ++i) {
17342 SDValue CurrentOp = SrcOp->getOperand(i);
17343 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17344 Elts.push_back(CurrentOp);
17347 ND = cast<ConstantSDNode>(CurrentOp);
17348 const APInt &C = ND->getAPIntValue();
17349 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
17352 case X86ISD::VSRLI:
17353 for (unsigned i=0; i!=NumElts; ++i) {
17354 SDValue CurrentOp = SrcOp->getOperand(i);
17355 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17356 Elts.push_back(CurrentOp);
17359 ND = cast<ConstantSDNode>(CurrentOp);
17360 const APInt &C = ND->getAPIntValue();
17361 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
17364 case X86ISD::VSRAI:
17365 for (unsigned i=0; i!=NumElts; ++i) {
17366 SDValue CurrentOp = SrcOp->getOperand(i);
17367 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17368 Elts.push_back(CurrentOp);
17371 ND = cast<ConstantSDNode>(CurrentOp);
17372 const APInt &C = ND->getAPIntValue();
17373 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
17378 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17381 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
17384 // getTargetVShiftNode - Handle vector element shifts where the shift amount
17385 // may or may not be a constant. Takes immediate version of shift as input.
17386 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
17387 SDValue SrcOp, SDValue ShAmt,
17388 SelectionDAG &DAG) {
17389 MVT SVT = ShAmt.getSimpleValueType();
17390 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
17392 // Catch shift-by-constant.
17393 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
17394 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
17395 CShAmt->getZExtValue(), DAG);
17397 // Change opcode to non-immediate version
17399 default: llvm_unreachable("Unknown target vector shift node");
17400 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
17401 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
17402 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
17405 const X86Subtarget &Subtarget =
17406 static_cast<const X86Subtarget &>(DAG.getSubtarget());
17407 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
17408 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
17409 // Let the shuffle legalizer expand this shift amount node.
17410 SDValue Op0 = ShAmt.getOperand(0);
17411 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
17412 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
17414 // Need to build a vector containing shift amount.
17415 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
17416 SmallVector<SDValue, 4> ShOps;
17417 ShOps.push_back(ShAmt);
17418 if (SVT == MVT::i32) {
17419 ShOps.push_back(DAG.getConstant(0, SVT));
17420 ShOps.push_back(DAG.getUNDEF(SVT));
17422 ShOps.push_back(DAG.getUNDEF(SVT));
17424 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
17425 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
17428 // The return type has to be a 128-bit type with the same element
17429 // type as the input type.
17430 MVT EltVT = VT.getVectorElementType();
17431 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
17433 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
17434 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
17437 /// \brief Return (and \p Op, \p Mask) for compare instructions or
17438 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
17439 /// necessary casting for \p Mask when lowering masking intrinsics.
17440 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
17441 SDValue PreservedSrc,
17442 const X86Subtarget *Subtarget,
17443 SelectionDAG &DAG) {
17444 EVT VT = Op.getValueType();
17445 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
17446 MVT::i1, VT.getVectorNumElements());
17447 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17448 Mask.getValueType().getSizeInBits());
17451 assert(MaskVT.isSimple() && "invalid mask type");
17453 if (isAllOnes(Mask))
17456 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17457 // are extracted by EXTRACT_SUBVECTOR.
17458 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17459 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17460 DAG.getIntPtrConstant(0));
17462 switch (Op.getOpcode()) {
17464 case X86ISD::PCMPEQM:
17465 case X86ISD::PCMPGTM:
17467 case X86ISD::CMPMU:
17468 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
17470 if (PreservedSrc.getOpcode() == ISD::UNDEF)
17471 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
17472 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
17475 /// \brief Creates an SDNode for a predicated scalar operation.
17476 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
17477 /// The mask is comming as MVT::i8 and it should be truncated
17478 /// to MVT::i1 while lowering masking intrinsics.
17479 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
17480 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
17481 /// a scalar instruction.
17482 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
17483 SDValue PreservedSrc,
17484 const X86Subtarget *Subtarget,
17485 SelectionDAG &DAG) {
17486 if (isAllOnes(Mask))
17489 EVT VT = Op.getValueType();
17491 // The mask should be of type MVT::i1
17492 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
17494 if (PreservedSrc.getOpcode() == ISD::UNDEF)
17495 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
17496 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
17499 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17500 SelectionDAG &DAG) {
17502 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17503 EVT VT = Op.getValueType();
17504 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
17506 switch(IntrData->Type) {
17507 case INTR_TYPE_1OP:
17508 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
17509 case INTR_TYPE_2OP:
17510 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17512 case INTR_TYPE_3OP:
17513 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17514 Op.getOperand(2), Op.getOperand(3));
17515 case INTR_TYPE_1OP_MASK_RM: {
17516 SDValue Src = Op.getOperand(1);
17517 SDValue Src0 = Op.getOperand(2);
17518 SDValue Mask = Op.getOperand(3);
17519 SDValue RoundingMode = Op.getOperand(4);
17520 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
17522 Mask, Src0, Subtarget, DAG);
17524 case INTR_TYPE_SCALAR_MASK_RM: {
17525 SDValue Src1 = Op.getOperand(1);
17526 SDValue Src2 = Op.getOperand(2);
17527 SDValue Src0 = Op.getOperand(3);
17528 SDValue Mask = Op.getOperand(4);
17529 // There are 2 kinds of intrinsics in this group:
17530 // (1) With supress-all-exceptions (sae) - 6 operands
17531 // (2) With rounding mode and sae - 7 operands.
17532 if (Op.getNumOperands() == 6) {
17533 SDValue Sae = Op.getOperand(5);
17534 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
17536 Mask, Src0, Subtarget, DAG);
17538 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
17539 SDValue RoundingMode = Op.getOperand(5);
17540 SDValue Sae = Op.getOperand(6);
17541 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
17542 RoundingMode, Sae),
17543 Mask, Src0, Subtarget, DAG);
17545 case INTR_TYPE_2OP_MASK: {
17546 SDValue Src1 = Op.getOperand(1);
17547 SDValue Src2 = Op.getOperand(2);
17548 SDValue PassThru = Op.getOperand(3);
17549 SDValue Mask = Op.getOperand(4);
17550 // We specify 2 possible opcodes for intrinsics with rounding modes.
17551 // First, we check if the intrinsic may have non-default rounding mode,
17552 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
17553 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
17554 if (IntrWithRoundingModeOpcode != 0) {
17555 SDValue Rnd = Op.getOperand(5);
17556 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
17557 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
17558 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
17559 dl, Op.getValueType(),
17561 Mask, PassThru, Subtarget, DAG);
17564 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
17566 Mask, PassThru, Subtarget, DAG);
17568 case FMA_OP_MASK: {
17569 SDValue Src1 = Op.getOperand(1);
17570 SDValue Src2 = Op.getOperand(2);
17571 SDValue Src3 = Op.getOperand(3);
17572 SDValue Mask = Op.getOperand(4);
17573 // We specify 2 possible opcodes for intrinsics with rounding modes.
17574 // First, we check if the intrinsic may have non-default rounding mode,
17575 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
17576 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
17577 if (IntrWithRoundingModeOpcode != 0) {
17578 SDValue Rnd = Op.getOperand(5);
17579 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
17580 X86::STATIC_ROUNDING::CUR_DIRECTION)
17581 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
17582 dl, Op.getValueType(),
17583 Src1, Src2, Src3, Rnd),
17584 Mask, Src1, Subtarget, DAG);
17586 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
17587 dl, Op.getValueType(),
17589 Mask, Src1, Subtarget, DAG);
17592 case CMP_MASK_CC: {
17593 // Comparison intrinsics with masks.
17594 // Example of transformation:
17595 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
17596 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
17598 // (v8i1 (insert_subvector undef,
17599 // (v2i1 (and (PCMPEQM %a, %b),
17600 // (extract_subvector
17601 // (v8i1 (bitcast %mask)), 0))), 0))))
17602 EVT VT = Op.getOperand(1).getValueType();
17603 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17604 VT.getVectorNumElements());
17605 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
17606 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17607 Mask.getValueType().getSizeInBits());
17609 if (IntrData->Type == CMP_MASK_CC) {
17610 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17611 Op.getOperand(2), Op.getOperand(3));
17613 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
17614 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17617 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
17618 DAG.getTargetConstant(0, MaskVT),
17620 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
17621 DAG.getUNDEF(BitcastVT), CmpMask,
17622 DAG.getIntPtrConstant(0));
17623 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
17625 case COMI: { // Comparison intrinsics
17626 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
17627 SDValue LHS = Op.getOperand(1);
17628 SDValue RHS = Op.getOperand(2);
17629 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
17630 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
17631 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
17632 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17633 DAG.getConstant(X86CC, MVT::i8), Cond);
17634 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17637 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
17638 Op.getOperand(1), Op.getOperand(2), DAG);
17640 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
17641 Op.getSimpleValueType(),
17643 Op.getOperand(2), DAG),
17644 Op.getOperand(4), Op.getOperand(3), Subtarget,
17646 case COMPRESS_EXPAND_IN_REG: {
17647 SDValue Mask = Op.getOperand(3);
17648 SDValue DataToCompress = Op.getOperand(1);
17649 SDValue PassThru = Op.getOperand(2);
17650 if (isAllOnes(Mask)) // return data as is
17651 return Op.getOperand(1);
17652 EVT VT = Op.getValueType();
17653 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17654 VT.getVectorNumElements());
17655 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17656 Mask.getValueType().getSizeInBits());
17658 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17659 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17660 DAG.getIntPtrConstant(0));
17662 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
17666 SDValue Mask = Op.getOperand(3);
17667 EVT VT = Op.getValueType();
17668 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17669 VT.getVectorNumElements());
17670 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17671 Mask.getValueType().getSizeInBits());
17673 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17674 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17675 DAG.getIntPtrConstant(0));
17676 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
17685 default: return SDValue(); // Don't custom lower most intrinsics.
17687 case Intrinsic::x86_avx512_mask_valign_q_512:
17688 case Intrinsic::x86_avx512_mask_valign_d_512:
17689 // Vector source operands are swapped.
17690 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
17691 Op.getValueType(), Op.getOperand(2),
17694 Op.getOperand(5), Op.getOperand(4),
17697 // ptest and testp intrinsics. The intrinsic these come from are designed to
17698 // return an integer value, not just an instruction so lower it to the ptest
17699 // or testp pattern and a setcc for the result.
17700 case Intrinsic::x86_sse41_ptestz:
17701 case Intrinsic::x86_sse41_ptestc:
17702 case Intrinsic::x86_sse41_ptestnzc:
17703 case Intrinsic::x86_avx_ptestz_256:
17704 case Intrinsic::x86_avx_ptestc_256:
17705 case Intrinsic::x86_avx_ptestnzc_256:
17706 case Intrinsic::x86_avx_vtestz_ps:
17707 case Intrinsic::x86_avx_vtestc_ps:
17708 case Intrinsic::x86_avx_vtestnzc_ps:
17709 case Intrinsic::x86_avx_vtestz_pd:
17710 case Intrinsic::x86_avx_vtestc_pd:
17711 case Intrinsic::x86_avx_vtestnzc_pd:
17712 case Intrinsic::x86_avx_vtestz_ps_256:
17713 case Intrinsic::x86_avx_vtestc_ps_256:
17714 case Intrinsic::x86_avx_vtestnzc_ps_256:
17715 case Intrinsic::x86_avx_vtestz_pd_256:
17716 case Intrinsic::x86_avx_vtestc_pd_256:
17717 case Intrinsic::x86_avx_vtestnzc_pd_256: {
17718 bool IsTestPacked = false;
17721 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
17722 case Intrinsic::x86_avx_vtestz_ps:
17723 case Intrinsic::x86_avx_vtestz_pd:
17724 case Intrinsic::x86_avx_vtestz_ps_256:
17725 case Intrinsic::x86_avx_vtestz_pd_256:
17726 IsTestPacked = true; // Fallthrough
17727 case Intrinsic::x86_sse41_ptestz:
17728 case Intrinsic::x86_avx_ptestz_256:
17730 X86CC = X86::COND_E;
17732 case Intrinsic::x86_avx_vtestc_ps:
17733 case Intrinsic::x86_avx_vtestc_pd:
17734 case Intrinsic::x86_avx_vtestc_ps_256:
17735 case Intrinsic::x86_avx_vtestc_pd_256:
17736 IsTestPacked = true; // Fallthrough
17737 case Intrinsic::x86_sse41_ptestc:
17738 case Intrinsic::x86_avx_ptestc_256:
17740 X86CC = X86::COND_B;
17742 case Intrinsic::x86_avx_vtestnzc_ps:
17743 case Intrinsic::x86_avx_vtestnzc_pd:
17744 case Intrinsic::x86_avx_vtestnzc_ps_256:
17745 case Intrinsic::x86_avx_vtestnzc_pd_256:
17746 IsTestPacked = true; // Fallthrough
17747 case Intrinsic::x86_sse41_ptestnzc:
17748 case Intrinsic::x86_avx_ptestnzc_256:
17750 X86CC = X86::COND_A;
17754 SDValue LHS = Op.getOperand(1);
17755 SDValue RHS = Op.getOperand(2);
17756 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
17757 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
17758 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17759 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
17760 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17762 case Intrinsic::x86_avx512_kortestz_w:
17763 case Intrinsic::x86_avx512_kortestc_w: {
17764 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
17765 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
17766 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
17767 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17768 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
17769 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17770 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17773 case Intrinsic::x86_sse42_pcmpistria128:
17774 case Intrinsic::x86_sse42_pcmpestria128:
17775 case Intrinsic::x86_sse42_pcmpistric128:
17776 case Intrinsic::x86_sse42_pcmpestric128:
17777 case Intrinsic::x86_sse42_pcmpistrio128:
17778 case Intrinsic::x86_sse42_pcmpestrio128:
17779 case Intrinsic::x86_sse42_pcmpistris128:
17780 case Intrinsic::x86_sse42_pcmpestris128:
17781 case Intrinsic::x86_sse42_pcmpistriz128:
17782 case Intrinsic::x86_sse42_pcmpestriz128: {
17786 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17787 case Intrinsic::x86_sse42_pcmpistria128:
17788 Opcode = X86ISD::PCMPISTRI;
17789 X86CC = X86::COND_A;
17791 case Intrinsic::x86_sse42_pcmpestria128:
17792 Opcode = X86ISD::PCMPESTRI;
17793 X86CC = X86::COND_A;
17795 case Intrinsic::x86_sse42_pcmpistric128:
17796 Opcode = X86ISD::PCMPISTRI;
17797 X86CC = X86::COND_B;
17799 case Intrinsic::x86_sse42_pcmpestric128:
17800 Opcode = X86ISD::PCMPESTRI;
17801 X86CC = X86::COND_B;
17803 case Intrinsic::x86_sse42_pcmpistrio128:
17804 Opcode = X86ISD::PCMPISTRI;
17805 X86CC = X86::COND_O;
17807 case Intrinsic::x86_sse42_pcmpestrio128:
17808 Opcode = X86ISD::PCMPESTRI;
17809 X86CC = X86::COND_O;
17811 case Intrinsic::x86_sse42_pcmpistris128:
17812 Opcode = X86ISD::PCMPISTRI;
17813 X86CC = X86::COND_S;
17815 case Intrinsic::x86_sse42_pcmpestris128:
17816 Opcode = X86ISD::PCMPESTRI;
17817 X86CC = X86::COND_S;
17819 case Intrinsic::x86_sse42_pcmpistriz128:
17820 Opcode = X86ISD::PCMPISTRI;
17821 X86CC = X86::COND_E;
17823 case Intrinsic::x86_sse42_pcmpestriz128:
17824 Opcode = X86ISD::PCMPESTRI;
17825 X86CC = X86::COND_E;
17828 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17829 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17830 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17831 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17832 DAG.getConstant(X86CC, MVT::i8),
17833 SDValue(PCMP.getNode(), 1));
17834 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17837 case Intrinsic::x86_sse42_pcmpistri128:
17838 case Intrinsic::x86_sse42_pcmpestri128: {
17840 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17841 Opcode = X86ISD::PCMPISTRI;
17843 Opcode = X86ISD::PCMPESTRI;
17845 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17846 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17847 return DAG.getNode(Opcode, dl, VTs, NewOps);
17852 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17853 SDValue Src, SDValue Mask, SDValue Base,
17854 SDValue Index, SDValue ScaleOp, SDValue Chain,
17855 const X86Subtarget * Subtarget) {
17857 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17858 assert(C && "Invalid scale type");
17859 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17860 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17861 Index.getSimpleValueType().getVectorNumElements());
17863 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17865 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17867 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17868 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17869 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17870 SDValue Segment = DAG.getRegister(0, MVT::i32);
17871 if (Src.getOpcode() == ISD::UNDEF)
17872 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
17873 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17874 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17875 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17876 return DAG.getMergeValues(RetOps, dl);
17879 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17880 SDValue Src, SDValue Mask, SDValue Base,
17881 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17883 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17884 assert(C && "Invalid scale type");
17885 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17886 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17887 SDValue Segment = DAG.getRegister(0, MVT::i32);
17888 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17889 Index.getSimpleValueType().getVectorNumElements());
17891 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17893 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17895 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17896 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17897 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17898 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17899 return SDValue(Res, 1);
17902 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17903 SDValue Mask, SDValue Base, SDValue Index,
17904 SDValue ScaleOp, SDValue Chain) {
17906 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17907 assert(C && "Invalid scale type");
17908 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17909 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17910 SDValue Segment = DAG.getRegister(0, MVT::i32);
17912 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17914 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17916 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17918 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17919 //SDVTList VTs = DAG.getVTList(MVT::Other);
17920 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17921 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17922 return SDValue(Res, 0);
17925 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17926 // read performance monitor counters (x86_rdpmc).
17927 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17928 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17929 SmallVectorImpl<SDValue> &Results) {
17930 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17931 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17934 // The ECX register is used to select the index of the performance counter
17936 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17938 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17940 // Reads the content of a 64-bit performance counter and returns it in the
17941 // registers EDX:EAX.
17942 if (Subtarget->is64Bit()) {
17943 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17944 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17947 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17948 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17951 Chain = HI.getValue(1);
17953 if (Subtarget->is64Bit()) {
17954 // The EAX register is loaded with the low-order 32 bits. The EDX register
17955 // is loaded with the supported high-order bits of the counter.
17956 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17957 DAG.getConstant(32, MVT::i8));
17958 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17959 Results.push_back(Chain);
17963 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17964 SDValue Ops[] = { LO, HI };
17965 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17966 Results.push_back(Pair);
17967 Results.push_back(Chain);
17970 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17971 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17972 // also used to custom lower READCYCLECOUNTER nodes.
17973 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17974 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17975 SmallVectorImpl<SDValue> &Results) {
17976 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17977 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17980 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17981 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17982 // and the EAX register is loaded with the low-order 32 bits.
17983 if (Subtarget->is64Bit()) {
17984 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17985 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17988 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17989 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17992 SDValue Chain = HI.getValue(1);
17994 if (Opcode == X86ISD::RDTSCP_DAG) {
17995 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17997 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17998 // the ECX register. Add 'ecx' explicitly to the chain.
17999 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
18001 // Explicitly store the content of ECX at the location passed in input
18002 // to the 'rdtscp' intrinsic.
18003 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
18004 MachinePointerInfo(), false, false, 0);
18007 if (Subtarget->is64Bit()) {
18008 // The EDX register is loaded with the high-order 32 bits of the MSR, and
18009 // the EAX register is loaded with the low-order 32 bits.
18010 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
18011 DAG.getConstant(32, MVT::i8));
18012 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
18013 Results.push_back(Chain);
18017 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
18018 SDValue Ops[] = { LO, HI };
18019 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
18020 Results.push_back(Pair);
18021 Results.push_back(Chain);
18024 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
18025 SelectionDAG &DAG) {
18026 SmallVector<SDValue, 2> Results;
18028 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
18030 return DAG.getMergeValues(Results, DL);
18034 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
18035 SelectionDAG &DAG) {
18036 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
18038 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
18043 switch(IntrData->Type) {
18045 llvm_unreachable("Unknown Intrinsic Type");
18049 // Emit the node with the right value type.
18050 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
18051 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
18053 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
18054 // Otherwise return the value from Rand, which is always 0, casted to i32.
18055 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
18056 DAG.getConstant(1, Op->getValueType(1)),
18057 DAG.getConstant(X86::COND_B, MVT::i32),
18058 SDValue(Result.getNode(), 1) };
18059 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
18060 DAG.getVTList(Op->getValueType(1), MVT::Glue),
18063 // Return { result, isValid, chain }.
18064 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
18065 SDValue(Result.getNode(), 2));
18068 //gather(v1, mask, index, base, scale);
18069 SDValue Chain = Op.getOperand(0);
18070 SDValue Src = Op.getOperand(2);
18071 SDValue Base = Op.getOperand(3);
18072 SDValue Index = Op.getOperand(4);
18073 SDValue Mask = Op.getOperand(5);
18074 SDValue Scale = Op.getOperand(6);
18075 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
18079 //scatter(base, mask, index, v1, scale);
18080 SDValue Chain = Op.getOperand(0);
18081 SDValue Base = Op.getOperand(2);
18082 SDValue Mask = Op.getOperand(3);
18083 SDValue Index = Op.getOperand(4);
18084 SDValue Src = Op.getOperand(5);
18085 SDValue Scale = Op.getOperand(6);
18086 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
18089 SDValue Hint = Op.getOperand(6);
18091 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
18092 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
18093 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
18094 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
18095 SDValue Chain = Op.getOperand(0);
18096 SDValue Mask = Op.getOperand(2);
18097 SDValue Index = Op.getOperand(3);
18098 SDValue Base = Op.getOperand(4);
18099 SDValue Scale = Op.getOperand(5);
18100 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
18102 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
18104 SmallVector<SDValue, 2> Results;
18105 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
18106 return DAG.getMergeValues(Results, dl);
18108 // Read Performance Monitoring Counters.
18110 SmallVector<SDValue, 2> Results;
18111 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
18112 return DAG.getMergeValues(Results, dl);
18114 // XTEST intrinsics.
18116 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
18117 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
18118 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18119 DAG.getConstant(X86::COND_NE, MVT::i8),
18121 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
18122 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
18123 Ret, SDValue(InTrans.getNode(), 1));
18127 SmallVector<SDValue, 2> Results;
18128 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
18129 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
18130 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
18131 DAG.getConstant(-1, MVT::i8));
18132 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
18133 Op.getOperand(4), GenCF.getValue(1));
18134 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
18135 Op.getOperand(5), MachinePointerInfo(),
18137 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18138 DAG.getConstant(X86::COND_B, MVT::i8),
18140 Results.push_back(SetCC);
18141 Results.push_back(Store);
18142 return DAG.getMergeValues(Results, dl);
18144 case COMPRESS_TO_MEM: {
18146 SDValue Mask = Op.getOperand(4);
18147 SDValue DataToCompress = Op.getOperand(3);
18148 SDValue Addr = Op.getOperand(2);
18149 SDValue Chain = Op.getOperand(0);
18151 if (isAllOnes(Mask)) // return just a store
18152 return DAG.getStore(Chain, dl, DataToCompress, Addr,
18153 MachinePointerInfo(), false, false, 0);
18155 EVT VT = DataToCompress.getValueType();
18156 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
18157 VT.getVectorNumElements());
18158 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
18159 Mask.getValueType().getSizeInBits());
18160 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
18161 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
18162 DAG.getIntPtrConstant(0));
18164 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
18165 DataToCompress, DAG.getUNDEF(VT));
18166 return DAG.getStore(Chain, dl, Compressed, Addr,
18167 MachinePointerInfo(), false, false, 0);
18169 case EXPAND_FROM_MEM: {
18171 SDValue Mask = Op.getOperand(4);
18172 SDValue PathThru = Op.getOperand(3);
18173 SDValue Addr = Op.getOperand(2);
18174 SDValue Chain = Op.getOperand(0);
18175 EVT VT = Op.getValueType();
18177 if (isAllOnes(Mask)) // return just a load
18178 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
18180 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
18181 VT.getVectorNumElements());
18182 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
18183 Mask.getValueType().getSizeInBits());
18184 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
18185 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
18186 DAG.getIntPtrConstant(0));
18188 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
18189 false, false, false, 0);
18191 SmallVector<SDValue, 2> Results;
18192 Results.push_back(DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand,
18194 Results.push_back(Chain);
18195 return DAG.getMergeValues(Results, dl);
18200 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
18201 SelectionDAG &DAG) const {
18202 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
18203 MFI->setReturnAddressIsTaken(true);
18205 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
18208 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18210 EVT PtrVT = getPointerTy();
18213 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
18214 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18215 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
18216 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
18217 DAG.getNode(ISD::ADD, dl, PtrVT,
18218 FrameAddr, Offset),
18219 MachinePointerInfo(), false, false, false, 0);
18222 // Just load the return address.
18223 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
18224 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
18225 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
18228 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
18229 MachineFunction &MF = DAG.getMachineFunction();
18230 MachineFrameInfo *MFI = MF.getFrameInfo();
18231 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
18232 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18233 EVT VT = Op.getValueType();
18235 MFI->setFrameAddressIsTaken(true);
18237 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
18238 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
18239 // is not possible to crawl up the stack without looking at the unwind codes
18241 int FrameAddrIndex = FuncInfo->getFAIndex();
18242 if (!FrameAddrIndex) {
18243 // Set up a frame object for the return address.
18244 unsigned SlotSize = RegInfo->getSlotSize();
18245 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
18246 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
18247 FuncInfo->setFAIndex(FrameAddrIndex);
18249 return DAG.getFrameIndex(FrameAddrIndex, VT);
18252 unsigned FrameReg =
18253 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
18254 SDLoc dl(Op); // FIXME probably not meaningful
18255 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18256 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
18257 (FrameReg == X86::EBP && VT == MVT::i32)) &&
18258 "Invalid Frame Register!");
18259 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
18261 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
18262 MachinePointerInfo(),
18263 false, false, false, 0);
18267 // FIXME? Maybe this could be a TableGen attribute on some registers and
18268 // this table could be generated automatically from RegInfo.
18269 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
18271 unsigned Reg = StringSwitch<unsigned>(RegName)
18272 .Case("esp", X86::ESP)
18273 .Case("rsp", X86::RSP)
18277 report_fatal_error("Invalid register name global variable");
18280 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
18281 SelectionDAG &DAG) const {
18282 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18283 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
18286 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
18287 SDValue Chain = Op.getOperand(0);
18288 SDValue Offset = Op.getOperand(1);
18289 SDValue Handler = Op.getOperand(2);
18292 EVT PtrVT = getPointerTy();
18293 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18294 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
18295 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
18296 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
18297 "Invalid Frame Register!");
18298 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
18299 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
18301 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
18302 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
18303 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
18304 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
18306 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
18308 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
18309 DAG.getRegister(StoreAddrReg, PtrVT));
18312 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
18313 SelectionDAG &DAG) const {
18315 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
18316 DAG.getVTList(MVT::i32, MVT::Other),
18317 Op.getOperand(0), Op.getOperand(1));
18320 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
18321 SelectionDAG &DAG) const {
18323 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
18324 Op.getOperand(0), Op.getOperand(1));
18327 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
18328 return Op.getOperand(0);
18331 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
18332 SelectionDAG &DAG) const {
18333 SDValue Root = Op.getOperand(0);
18334 SDValue Trmp = Op.getOperand(1); // trampoline
18335 SDValue FPtr = Op.getOperand(2); // nested function
18336 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
18339 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
18340 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18342 if (Subtarget->is64Bit()) {
18343 SDValue OutChains[6];
18345 // Large code-model.
18346 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
18347 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
18349 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
18350 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
18352 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
18354 // Load the pointer to the nested function into R11.
18355 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
18356 SDValue Addr = Trmp;
18357 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18358 Addr, MachinePointerInfo(TrmpAddr),
18361 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18362 DAG.getConstant(2, MVT::i64));
18363 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
18364 MachinePointerInfo(TrmpAddr, 2),
18367 // Load the 'nest' parameter value into R10.
18368 // R10 is specified in X86CallingConv.td
18369 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
18370 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18371 DAG.getConstant(10, MVT::i64));
18372 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18373 Addr, MachinePointerInfo(TrmpAddr, 10),
18376 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18377 DAG.getConstant(12, MVT::i64));
18378 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
18379 MachinePointerInfo(TrmpAddr, 12),
18382 // Jump to the nested function.
18383 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
18384 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18385 DAG.getConstant(20, MVT::i64));
18386 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18387 Addr, MachinePointerInfo(TrmpAddr, 20),
18390 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
18391 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18392 DAG.getConstant(22, MVT::i64));
18393 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
18394 MachinePointerInfo(TrmpAddr, 22),
18397 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
18399 const Function *Func =
18400 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
18401 CallingConv::ID CC = Func->getCallingConv();
18406 llvm_unreachable("Unsupported calling convention");
18407 case CallingConv::C:
18408 case CallingConv::X86_StdCall: {
18409 // Pass 'nest' parameter in ECX.
18410 // Must be kept in sync with X86CallingConv.td
18411 NestReg = X86::ECX;
18413 // Check that ECX wasn't needed by an 'inreg' parameter.
18414 FunctionType *FTy = Func->getFunctionType();
18415 const AttributeSet &Attrs = Func->getAttributes();
18417 if (!Attrs.isEmpty() && !Func->isVarArg()) {
18418 unsigned InRegCount = 0;
18421 for (FunctionType::param_iterator I = FTy->param_begin(),
18422 E = FTy->param_end(); I != E; ++I, ++Idx)
18423 if (Attrs.hasAttribute(Idx, Attribute::InReg))
18424 // FIXME: should only count parameters that are lowered to integers.
18425 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
18427 if (InRegCount > 2) {
18428 report_fatal_error("Nest register in use - reduce number of inreg"
18434 case CallingConv::X86_FastCall:
18435 case CallingConv::X86_ThisCall:
18436 case CallingConv::Fast:
18437 // Pass 'nest' parameter in EAX.
18438 // Must be kept in sync with X86CallingConv.td
18439 NestReg = X86::EAX;
18443 SDValue OutChains[4];
18444 SDValue Addr, Disp;
18446 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18447 DAG.getConstant(10, MVT::i32));
18448 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
18450 // This is storing the opcode for MOV32ri.
18451 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
18452 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
18453 OutChains[0] = DAG.getStore(Root, dl,
18454 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
18455 Trmp, MachinePointerInfo(TrmpAddr),
18458 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18459 DAG.getConstant(1, MVT::i32));
18460 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
18461 MachinePointerInfo(TrmpAddr, 1),
18464 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
18465 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18466 DAG.getConstant(5, MVT::i32));
18467 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
18468 MachinePointerInfo(TrmpAddr, 5),
18471 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18472 DAG.getConstant(6, MVT::i32));
18473 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
18474 MachinePointerInfo(TrmpAddr, 6),
18477 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
18481 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
18482 SelectionDAG &DAG) const {
18484 The rounding mode is in bits 11:10 of FPSR, and has the following
18486 00 Round to nearest
18491 FLT_ROUNDS, on the other hand, expects the following:
18498 To perform the conversion, we do:
18499 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
18502 MachineFunction &MF = DAG.getMachineFunction();
18503 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
18504 unsigned StackAlignment = TFI.getStackAlignment();
18505 MVT VT = Op.getSimpleValueType();
18508 // Save FP Control Word to stack slot
18509 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
18510 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
18512 MachineMemOperand *MMO =
18513 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
18514 MachineMemOperand::MOStore, 2, 2);
18516 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
18517 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
18518 DAG.getVTList(MVT::Other),
18519 Ops, MVT::i16, MMO);
18521 // Load FP Control Word from stack slot
18522 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
18523 MachinePointerInfo(), false, false, false, 0);
18525 // Transform as necessary
18527 DAG.getNode(ISD::SRL, DL, MVT::i16,
18528 DAG.getNode(ISD::AND, DL, MVT::i16,
18529 CWD, DAG.getConstant(0x800, MVT::i16)),
18530 DAG.getConstant(11, MVT::i8));
18532 DAG.getNode(ISD::SRL, DL, MVT::i16,
18533 DAG.getNode(ISD::AND, DL, MVT::i16,
18534 CWD, DAG.getConstant(0x400, MVT::i16)),
18535 DAG.getConstant(9, MVT::i8));
18538 DAG.getNode(ISD::AND, DL, MVT::i16,
18539 DAG.getNode(ISD::ADD, DL, MVT::i16,
18540 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
18541 DAG.getConstant(1, MVT::i16)),
18542 DAG.getConstant(3, MVT::i16));
18544 return DAG.getNode((VT.getSizeInBits() < 16 ?
18545 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
18548 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
18549 MVT VT = Op.getSimpleValueType();
18551 unsigned NumBits = VT.getSizeInBits();
18554 Op = Op.getOperand(0);
18555 if (VT == MVT::i8) {
18556 // Zero extend to i32 since there is not an i8 bsr.
18558 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18561 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
18562 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18563 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18565 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18568 DAG.getConstant(NumBits+NumBits-1, OpVT),
18569 DAG.getConstant(X86::COND_E, MVT::i8),
18572 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18574 // Finally xor with NumBits-1.
18575 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18578 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18582 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
18583 MVT VT = Op.getSimpleValueType();
18585 unsigned NumBits = VT.getSizeInBits();
18588 Op = Op.getOperand(0);
18589 if (VT == MVT::i8) {
18590 // Zero extend to i32 since there is not an i8 bsr.
18592 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18595 // Issue a bsr (scan bits in reverse).
18596 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18597 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18599 // And xor with NumBits-1.
18600 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18603 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18607 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18608 MVT VT = Op.getSimpleValueType();
18609 unsigned NumBits = VT.getSizeInBits();
18611 Op = Op.getOperand(0);
18613 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18614 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18615 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
18617 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18620 DAG.getConstant(NumBits, VT),
18621 DAG.getConstant(X86::COND_E, MVT::i8),
18624 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18627 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18628 // ones, and then concatenate the result back.
18629 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18630 MVT VT = Op.getSimpleValueType();
18632 assert(VT.is256BitVector() && VT.isInteger() &&
18633 "Unsupported value type for operation");
18635 unsigned NumElems = VT.getVectorNumElements();
18638 // Extract the LHS vectors
18639 SDValue LHS = Op.getOperand(0);
18640 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18641 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18643 // Extract the RHS vectors
18644 SDValue RHS = Op.getOperand(1);
18645 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18646 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18648 MVT EltVT = VT.getVectorElementType();
18649 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18651 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18652 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18653 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18656 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18657 assert(Op.getSimpleValueType().is256BitVector() &&
18658 Op.getSimpleValueType().isInteger() &&
18659 "Only handle AVX 256-bit vector integer operation");
18660 return Lower256IntArith(Op, DAG);
18663 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18664 assert(Op.getSimpleValueType().is256BitVector() &&
18665 Op.getSimpleValueType().isInteger() &&
18666 "Only handle AVX 256-bit vector integer operation");
18667 return Lower256IntArith(Op, DAG);
18670 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18671 SelectionDAG &DAG) {
18673 MVT VT = Op.getSimpleValueType();
18675 // Decompose 256-bit ops into smaller 128-bit ops.
18676 if (VT.is256BitVector() && !Subtarget->hasInt256())
18677 return Lower256IntArith(Op, DAG);
18679 SDValue A = Op.getOperand(0);
18680 SDValue B = Op.getOperand(1);
18682 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18683 if (VT == MVT::v4i32) {
18684 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18685 "Should not custom lower when pmuldq is available!");
18687 // Extract the odd parts.
18688 static const int UnpackMask[] = { 1, -1, 3, -1 };
18689 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18690 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18692 // Multiply the even parts.
18693 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18694 // Now multiply odd parts.
18695 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18697 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
18698 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
18700 // Merge the two vectors back together with a shuffle. This expands into 2
18702 static const int ShufMask[] = { 0, 4, 2, 6 };
18703 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18706 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18707 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18709 // Ahi = psrlqi(a, 32);
18710 // Bhi = psrlqi(b, 32);
18712 // AloBlo = pmuludq(a, b);
18713 // AloBhi = pmuludq(a, Bhi);
18714 // AhiBlo = pmuludq(Ahi, b);
18716 // AloBhi = psllqi(AloBhi, 32);
18717 // AhiBlo = psllqi(AhiBlo, 32);
18718 // return AloBlo + AloBhi + AhiBlo;
18720 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18721 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18723 // Bit cast to 32-bit vectors for MULUDQ
18724 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18725 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18726 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
18727 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
18728 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
18729 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
18731 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18732 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18733 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18735 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18736 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18738 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18739 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18742 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18743 assert(Subtarget->isTargetWin64() && "Unexpected target");
18744 EVT VT = Op.getValueType();
18745 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18746 "Unexpected return type for lowering");
18750 switch (Op->getOpcode()) {
18751 default: llvm_unreachable("Unexpected request for libcall!");
18752 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18753 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18754 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18755 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18756 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18757 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18761 SDValue InChain = DAG.getEntryNode();
18763 TargetLowering::ArgListTy Args;
18764 TargetLowering::ArgListEntry Entry;
18765 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18766 EVT ArgVT = Op->getOperand(i).getValueType();
18767 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18768 "Unexpected argument type for lowering");
18769 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18770 Entry.Node = StackPtr;
18771 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18773 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18774 Entry.Ty = PointerType::get(ArgTy,0);
18775 Entry.isSExt = false;
18776 Entry.isZExt = false;
18777 Args.push_back(Entry);
18780 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18783 TargetLowering::CallLoweringInfo CLI(DAG);
18784 CLI.setDebugLoc(dl).setChain(InChain)
18785 .setCallee(getLibcallCallingConv(LC),
18786 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18787 Callee, std::move(Args), 0)
18788 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18790 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18791 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
18794 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18795 SelectionDAG &DAG) {
18796 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18797 EVT VT = Op0.getValueType();
18800 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18801 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18803 // PMULxD operations multiply each even value (starting at 0) of LHS with
18804 // the related value of RHS and produce a widen result.
18805 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18806 // => <2 x i64> <ae|cg>
18808 // In other word, to have all the results, we need to perform two PMULxD:
18809 // 1. one with the even values.
18810 // 2. one with the odd values.
18811 // To achieve #2, with need to place the odd values at an even position.
18813 // Place the odd value at an even position (basically, shift all values 1
18814 // step to the left):
18815 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18816 // <a|b|c|d> => <b|undef|d|undef>
18817 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18818 // <e|f|g|h> => <f|undef|h|undef>
18819 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18821 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18823 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18824 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18826 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18827 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18828 // => <2 x i64> <ae|cg>
18829 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
18830 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18831 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18832 // => <2 x i64> <bf|dh>
18833 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
18834 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18836 // Shuffle it back into the right order.
18837 SDValue Highs, Lows;
18838 if (VT == MVT::v8i32) {
18839 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18840 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18841 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18842 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18844 const int HighMask[] = {1, 5, 3, 7};
18845 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18846 const int LowMask[] = {0, 4, 2, 6};
18847 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18850 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18851 // unsigned multiply.
18852 if (IsSigned && !Subtarget->hasSSE41()) {
18854 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
18855 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18856 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18857 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18858 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18860 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18861 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18864 // The first result of MUL_LOHI is actually the low value, followed by the
18866 SDValue Ops[] = {Lows, Highs};
18867 return DAG.getMergeValues(Ops, dl);
18870 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18871 const X86Subtarget *Subtarget) {
18872 MVT VT = Op.getSimpleValueType();
18874 SDValue R = Op.getOperand(0);
18875 SDValue Amt = Op.getOperand(1);
18877 // Optimize shl/srl/sra with constant shift amount.
18878 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18879 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18880 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18882 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
18883 (Subtarget->hasInt256() &&
18884 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
18885 (Subtarget->hasAVX512() &&
18886 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
18887 if (Op.getOpcode() == ISD::SHL)
18888 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
18890 if (Op.getOpcode() == ISD::SRL)
18891 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
18893 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
18894 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
18898 if (VT == MVT::v16i8) {
18899 if (Op.getOpcode() == ISD::SHL) {
18900 // Make a large shift.
18901 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18902 MVT::v8i16, R, ShiftAmt,
18904 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18905 // Zero out the rightmost bits.
18906 SmallVector<SDValue, 16> V(16,
18907 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18909 return DAG.getNode(ISD::AND, dl, VT, SHL,
18910 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18912 if (Op.getOpcode() == ISD::SRL) {
18913 // Make a large shift.
18914 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18915 MVT::v8i16, R, ShiftAmt,
18917 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18918 // Zero out the leftmost bits.
18919 SmallVector<SDValue, 16> V(16,
18920 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18922 return DAG.getNode(ISD::AND, dl, VT, SRL,
18923 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18925 if (Op.getOpcode() == ISD::SRA) {
18926 if (ShiftAmt == 7) {
18927 // R s>> 7 === R s< 0
18928 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18929 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18932 // R s>> a === ((R u>> a) ^ m) - m
18933 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18934 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
18936 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18937 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18938 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18941 llvm_unreachable("Unknown shift opcode.");
18944 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
18945 if (Op.getOpcode() == ISD::SHL) {
18946 // Make a large shift.
18947 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18948 MVT::v16i16, R, ShiftAmt,
18950 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18951 // Zero out the rightmost bits.
18952 SmallVector<SDValue, 32> V(32,
18953 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18955 return DAG.getNode(ISD::AND, dl, VT, SHL,
18956 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18958 if (Op.getOpcode() == ISD::SRL) {
18959 // Make a large shift.
18960 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18961 MVT::v16i16, R, ShiftAmt,
18963 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18964 // Zero out the leftmost bits.
18965 SmallVector<SDValue, 32> V(32,
18966 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18968 return DAG.getNode(ISD::AND, dl, VT, SRL,
18969 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18971 if (Op.getOpcode() == ISD::SRA) {
18972 if (ShiftAmt == 7) {
18973 // R s>> 7 === R s< 0
18974 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18975 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18978 // R s>> a === ((R u>> a) ^ m) - m
18979 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18980 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
18982 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18983 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18984 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18987 llvm_unreachable("Unknown shift opcode.");
18992 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18993 if (!Subtarget->is64Bit() &&
18994 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18995 Amt.getOpcode() == ISD::BITCAST &&
18996 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18997 Amt = Amt.getOperand(0);
18998 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18999 VT.getVectorNumElements();
19000 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
19001 uint64_t ShiftAmt = 0;
19002 for (unsigned i = 0; i != Ratio; ++i) {
19003 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
19007 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
19009 // Check remaining shift amounts.
19010 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
19011 uint64_t ShAmt = 0;
19012 for (unsigned j = 0; j != Ratio; ++j) {
19013 ConstantSDNode *C =
19014 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
19018 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
19020 if (ShAmt != ShiftAmt)
19023 switch (Op.getOpcode()) {
19025 llvm_unreachable("Unknown shift opcode!");
19027 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
19030 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
19033 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
19041 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
19042 const X86Subtarget* Subtarget) {
19043 MVT VT = Op.getSimpleValueType();
19045 SDValue R = Op.getOperand(0);
19046 SDValue Amt = Op.getOperand(1);
19048 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
19049 VT == MVT::v4i32 || VT == MVT::v8i16 ||
19050 (Subtarget->hasInt256() &&
19051 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
19052 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
19053 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
19055 EVT EltVT = VT.getVectorElementType();
19057 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
19058 // Check if this build_vector node is doing a splat.
19059 // If so, then set BaseShAmt equal to the splat value.
19060 BaseShAmt = BV->getSplatValue();
19061 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
19062 BaseShAmt = SDValue();
19064 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
19065 Amt = Amt.getOperand(0);
19067 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
19068 if (SVN && SVN->isSplat()) {
19069 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
19070 SDValue InVec = Amt.getOperand(0);
19071 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
19072 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
19073 "Unexpected shuffle index found!");
19074 BaseShAmt = InVec.getOperand(SplatIdx);
19075 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
19076 if (ConstantSDNode *C =
19077 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
19078 if (C->getZExtValue() == SplatIdx)
19079 BaseShAmt = InVec.getOperand(1);
19084 // Avoid introducing an extract element from a shuffle.
19085 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
19086 DAG.getIntPtrConstant(SplatIdx));
19090 if (BaseShAmt.getNode()) {
19091 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
19092 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
19093 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
19094 else if (EltVT.bitsLT(MVT::i32))
19095 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
19097 switch (Op.getOpcode()) {
19099 llvm_unreachable("Unknown shift opcode!");
19101 switch (VT.SimpleTy) {
19102 default: return SDValue();
19111 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
19114 switch (VT.SimpleTy) {
19115 default: return SDValue();
19122 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
19125 switch (VT.SimpleTy) {
19126 default: return SDValue();
19135 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
19141 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
19142 if (!Subtarget->is64Bit() &&
19143 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
19144 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
19145 Amt.getOpcode() == ISD::BITCAST &&
19146 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
19147 Amt = Amt.getOperand(0);
19148 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
19149 VT.getVectorNumElements();
19150 std::vector<SDValue> Vals(Ratio);
19151 for (unsigned i = 0; i != Ratio; ++i)
19152 Vals[i] = Amt.getOperand(i);
19153 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
19154 for (unsigned j = 0; j != Ratio; ++j)
19155 if (Vals[j] != Amt.getOperand(i + j))
19158 switch (Op.getOpcode()) {
19160 llvm_unreachable("Unknown shift opcode!");
19162 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
19164 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
19166 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
19173 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
19174 SelectionDAG &DAG) {
19175 MVT VT = Op.getSimpleValueType();
19177 SDValue R = Op.getOperand(0);
19178 SDValue Amt = Op.getOperand(1);
19181 assert(VT.isVector() && "Custom lowering only for vector shifts!");
19182 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
19184 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
19188 V = LowerScalarVariableShift(Op, DAG, Subtarget);
19192 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
19194 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
19195 if (Subtarget->hasInt256()) {
19196 if (Op.getOpcode() == ISD::SRL &&
19197 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
19198 VT == MVT::v4i64 || VT == MVT::v8i32))
19200 if (Op.getOpcode() == ISD::SHL &&
19201 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
19202 VT == MVT::v4i64 || VT == MVT::v8i32))
19204 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
19208 // If possible, lower this packed shift into a vector multiply instead of
19209 // expanding it into a sequence of scalar shifts.
19210 // Do this only if the vector shift count is a constant build_vector.
19211 if (Op.getOpcode() == ISD::SHL &&
19212 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
19213 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
19214 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
19215 SmallVector<SDValue, 8> Elts;
19216 EVT SVT = VT.getScalarType();
19217 unsigned SVTBits = SVT.getSizeInBits();
19218 const APInt &One = APInt(SVTBits, 1);
19219 unsigned NumElems = VT.getVectorNumElements();
19221 for (unsigned i=0; i !=NumElems; ++i) {
19222 SDValue Op = Amt->getOperand(i);
19223 if (Op->getOpcode() == ISD::UNDEF) {
19224 Elts.push_back(Op);
19228 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
19229 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
19230 uint64_t ShAmt = C.getZExtValue();
19231 if (ShAmt >= SVTBits) {
19232 Elts.push_back(DAG.getUNDEF(SVT));
19235 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
19237 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
19238 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
19241 // Lower SHL with variable shift amount.
19242 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
19243 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
19245 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
19246 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
19247 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
19248 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
19251 // If possible, lower this shift as a sequence of two shifts by
19252 // constant plus a MOVSS/MOVSD instead of scalarizing it.
19254 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
19256 // Could be rewritten as:
19257 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
19259 // The advantage is that the two shifts from the example would be
19260 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
19261 // the vector shift into four scalar shifts plus four pairs of vector
19263 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
19264 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
19265 unsigned TargetOpcode = X86ISD::MOVSS;
19266 bool CanBeSimplified;
19267 // The splat value for the first packed shift (the 'X' from the example).
19268 SDValue Amt1 = Amt->getOperand(0);
19269 // The splat value for the second packed shift (the 'Y' from the example).
19270 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
19271 Amt->getOperand(2);
19273 // See if it is possible to replace this node with a sequence of
19274 // two shifts followed by a MOVSS/MOVSD
19275 if (VT == MVT::v4i32) {
19276 // Check if it is legal to use a MOVSS.
19277 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
19278 Amt2 == Amt->getOperand(3);
19279 if (!CanBeSimplified) {
19280 // Otherwise, check if we can still simplify this node using a MOVSD.
19281 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
19282 Amt->getOperand(2) == Amt->getOperand(3);
19283 TargetOpcode = X86ISD::MOVSD;
19284 Amt2 = Amt->getOperand(2);
19287 // Do similar checks for the case where the machine value type
19289 CanBeSimplified = Amt1 == Amt->getOperand(1);
19290 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
19291 CanBeSimplified = Amt2 == Amt->getOperand(i);
19293 if (!CanBeSimplified) {
19294 TargetOpcode = X86ISD::MOVSD;
19295 CanBeSimplified = true;
19296 Amt2 = Amt->getOperand(4);
19297 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
19298 CanBeSimplified = Amt1 == Amt->getOperand(i);
19299 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
19300 CanBeSimplified = Amt2 == Amt->getOperand(j);
19304 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
19305 isa<ConstantSDNode>(Amt2)) {
19306 // Replace this node with two shifts followed by a MOVSS/MOVSD.
19307 EVT CastVT = MVT::v4i32;
19309 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
19310 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
19312 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
19313 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
19314 if (TargetOpcode == X86ISD::MOVSD)
19315 CastVT = MVT::v2i64;
19316 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
19317 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
19318 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
19320 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
19324 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
19325 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
19328 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
19329 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
19331 // Turn 'a' into a mask suitable for VSELECT
19332 SDValue VSelM = DAG.getConstant(0x80, VT);
19333 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19334 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19336 SDValue CM1 = DAG.getConstant(0x0f, VT);
19337 SDValue CM2 = DAG.getConstant(0x3f, VT);
19339 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
19340 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
19341 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
19342 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
19343 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
19346 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
19347 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19348 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19350 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
19351 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
19352 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
19353 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
19354 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
19357 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
19358 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19359 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19361 // return VSELECT(r, r+r, a);
19362 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
19363 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
19367 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
19368 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
19369 // solution better.
19370 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
19371 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
19373 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
19374 R = DAG.getNode(ExtOpc, dl, NewVT, R);
19375 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
19376 return DAG.getNode(ISD::TRUNCATE, dl, VT,
19377 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
19380 // Decompose 256-bit shifts into smaller 128-bit shifts.
19381 if (VT.is256BitVector()) {
19382 unsigned NumElems = VT.getVectorNumElements();
19383 MVT EltVT = VT.getVectorElementType();
19384 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19386 // Extract the two vectors
19387 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
19388 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
19390 // Recreate the shift amount vectors
19391 SDValue Amt1, Amt2;
19392 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
19393 // Constant shift amount
19394 SmallVector<SDValue, 4> Amt1Csts;
19395 SmallVector<SDValue, 4> Amt2Csts;
19396 for (unsigned i = 0; i != NumElems/2; ++i)
19397 Amt1Csts.push_back(Amt->getOperand(i));
19398 for (unsigned i = NumElems/2; i != NumElems; ++i)
19399 Amt2Csts.push_back(Amt->getOperand(i));
19401 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
19402 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19404 // Variable shift amount
19405 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19406 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19409 // Issue new vector shifts for the smaller types
19410 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19411 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19413 // Concatenate the result back
19414 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19420 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19421 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19422 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19423 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19424 // has only one use.
19425 SDNode *N = Op.getNode();
19426 SDValue LHS = N->getOperand(0);
19427 SDValue RHS = N->getOperand(1);
19428 unsigned BaseOp = 0;
19431 switch (Op.getOpcode()) {
19432 default: llvm_unreachable("Unknown ovf instruction!");
19434 // A subtract of one will be selected as a INC. Note that INC doesn't
19435 // set CF, so we can't do this for UADDO.
19436 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
19438 BaseOp = X86ISD::INC;
19439 Cond = X86::COND_O;
19442 BaseOp = X86ISD::ADD;
19443 Cond = X86::COND_O;
19446 BaseOp = X86ISD::ADD;
19447 Cond = X86::COND_B;
19450 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19451 // set CF, so we can't do this for USUBO.
19452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
19454 BaseOp = X86ISD::DEC;
19455 Cond = X86::COND_O;
19458 BaseOp = X86ISD::SUB;
19459 Cond = X86::COND_O;
19462 BaseOp = X86ISD::SUB;
19463 Cond = X86::COND_B;
19466 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19467 Cond = X86::COND_O;
19469 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19470 if (N->getValueType(0) == MVT::i8) {
19471 BaseOp = X86ISD::UMUL8;
19472 Cond = X86::COND_O;
19475 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19477 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19480 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19481 DAG.getConstant(X86::COND_O, MVT::i32),
19482 SDValue(Sum.getNode(), 2));
19484 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19488 // Also sets EFLAGS.
19489 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19490 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19493 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19494 DAG.getConstant(Cond, MVT::i32),
19495 SDValue(Sum.getNode(), 1));
19497 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19500 // Sign extension of the low part of vector elements. This may be used either
19501 // when sign extend instructions are not available or if the vector element
19502 // sizes already match the sign-extended size. If the vector elements are in
19503 // their pre-extended size and sign extend instructions are available, that will
19504 // be handled by LowerSIGN_EXTEND.
19505 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
19506 SelectionDAG &DAG) const {
19508 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
19509 MVT VT = Op.getSimpleValueType();
19511 if (!Subtarget->hasSSE2() || !VT.isVector())
19514 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
19515 ExtraVT.getScalarType().getSizeInBits();
19517 switch (VT.SimpleTy) {
19518 default: return SDValue();
19521 if (!Subtarget->hasFp256())
19523 if (!Subtarget->hasInt256()) {
19524 // needs to be split
19525 unsigned NumElems = VT.getVectorNumElements();
19527 // Extract the LHS vectors
19528 SDValue LHS = Op.getOperand(0);
19529 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
19530 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
19532 MVT EltVT = VT.getVectorElementType();
19533 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19535 EVT ExtraEltVT = ExtraVT.getVectorElementType();
19536 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
19537 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
19539 SDValue Extra = DAG.getValueType(ExtraVT);
19541 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
19542 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
19544 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
19549 SDValue Op0 = Op.getOperand(0);
19551 // This is a sign extension of some low part of vector elements without
19552 // changing the size of the vector elements themselves:
19553 // Shift-Left + Shift-Right-Algebraic.
19554 SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
19556 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
19562 /// Returns true if the operand type is exactly twice the native width, and
19563 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19564 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19565 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19566 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
19567 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19570 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19571 else if (OpWidth == 128)
19572 return Subtarget->hasCmpxchg16b();
19577 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19578 return needsCmpXchgNb(SI->getValueOperand()->getType());
19581 // Note: this turns large loads into lock cmpxchg8b/16b.
19582 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19583 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19584 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19585 return needsCmpXchgNb(PTy->getElementType());
19588 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19589 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19590 const Type *MemType = AI->getType();
19592 // If the operand is too big, we must see if cmpxchg8/16b is available
19593 // and default to library calls otherwise.
19594 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19595 return needsCmpXchgNb(MemType);
19597 AtomicRMWInst::BinOp Op = AI->getOperation();
19600 llvm_unreachable("Unknown atomic operation");
19601 case AtomicRMWInst::Xchg:
19602 case AtomicRMWInst::Add:
19603 case AtomicRMWInst::Sub:
19604 // It's better to use xadd, xsub or xchg for these in all cases.
19606 case AtomicRMWInst::Or:
19607 case AtomicRMWInst::And:
19608 case AtomicRMWInst::Xor:
19609 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19610 // prefix to a normal instruction for these operations.
19611 return !AI->use_empty();
19612 case AtomicRMWInst::Nand:
19613 case AtomicRMWInst::Max:
19614 case AtomicRMWInst::Min:
19615 case AtomicRMWInst::UMax:
19616 case AtomicRMWInst::UMin:
19617 // These always require a non-trivial set of data operations on x86. We must
19618 // use a cmpxchg loop.
19623 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19624 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19625 // no-sse2). There isn't any reason to disable it if the target processor
19627 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19631 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19632 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19633 const Type *MemType = AI->getType();
19634 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19635 // there is no benefit in turning such RMWs into loads, and it is actually
19636 // harmful as it introduces a mfence.
19637 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19640 auto Builder = IRBuilder<>(AI);
19641 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19642 auto SynchScope = AI->getSynchScope();
19643 // We must restrict the ordering to avoid generating loads with Release or
19644 // ReleaseAcquire orderings.
19645 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19646 auto Ptr = AI->getPointerOperand();
19648 // Before the load we need a fence. Here is an example lifted from
19649 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19652 // x.store(1, relaxed);
19653 // r1 = y.fetch_add(0, release);
19655 // y.fetch_add(42, acquire);
19656 // r2 = x.load(relaxed);
19657 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19658 // lowered to just a load without a fence. A mfence flushes the store buffer,
19659 // making the optimization clearly correct.
19660 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19661 // otherwise, we might be able to be more agressive on relaxed idempotent
19662 // rmw. In practice, they do not look useful, so we don't try to be
19663 // especially clever.
19664 if (SynchScope == SingleThread) {
19665 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19666 // the IR level, so we must wrap it in an intrinsic.
19668 } else if (hasMFENCE(*Subtarget)) {
19669 Function *MFence = llvm::Intrinsic::getDeclaration(M,
19670 Intrinsic::x86_sse2_mfence);
19671 Builder.CreateCall(MFence);
19673 // FIXME: it might make sense to use a locked operation here but on a
19674 // different cache-line to prevent cache-line bouncing. In practice it
19675 // is probably a small win, and x86 processors without mfence are rare
19676 // enough that we do not bother.
19680 // Finally we can emit the atomic load.
19681 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19682 AI->getType()->getPrimitiveSizeInBits());
19683 Loaded->setAtomic(Order, SynchScope);
19684 AI->replaceAllUsesWith(Loaded);
19685 AI->eraseFromParent();
19689 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19690 SelectionDAG &DAG) {
19692 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19693 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19694 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19695 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19697 // The only fence that needs an instruction is a sequentially-consistent
19698 // cross-thread fence.
19699 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19700 if (hasMFENCE(*Subtarget))
19701 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19703 SDValue Chain = Op.getOperand(0);
19704 SDValue Zero = DAG.getConstant(0, MVT::i32);
19706 DAG.getRegister(X86::ESP, MVT::i32), // Base
19707 DAG.getTargetConstant(1, MVT::i8), // Scale
19708 DAG.getRegister(0, MVT::i32), // Index
19709 DAG.getTargetConstant(0, MVT::i32), // Disp
19710 DAG.getRegister(0, MVT::i32), // Segment.
19714 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19715 return SDValue(Res, 0);
19718 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19719 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19722 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19723 SelectionDAG &DAG) {
19724 MVT T = Op.getSimpleValueType();
19728 switch(T.SimpleTy) {
19729 default: llvm_unreachable("Invalid value type!");
19730 case MVT::i8: Reg = X86::AL; size = 1; break;
19731 case MVT::i16: Reg = X86::AX; size = 2; break;
19732 case MVT::i32: Reg = X86::EAX; size = 4; break;
19734 assert(Subtarget->is64Bit() && "Node not type legal!");
19735 Reg = X86::RAX; size = 8;
19738 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19739 Op.getOperand(2), SDValue());
19740 SDValue Ops[] = { cpIn.getValue(0),
19743 DAG.getTargetConstant(size, MVT::i8),
19744 cpIn.getValue(1) };
19745 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19746 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19747 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19751 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19752 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19753 MVT::i32, cpOut.getValue(2));
19754 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19755 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
19757 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19758 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19759 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19763 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19764 SelectionDAG &DAG) {
19765 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19766 MVT DstVT = Op.getSimpleValueType();
19768 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19769 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19770 if (DstVT != MVT::f64)
19771 // This conversion needs to be expanded.
19774 SDValue InVec = Op->getOperand(0);
19776 unsigned NumElts = SrcVT.getVectorNumElements();
19777 EVT SVT = SrcVT.getVectorElementType();
19779 // Widen the vector in input in the case of MVT::v2i32.
19780 // Example: from MVT::v2i32 to MVT::v4i32.
19781 SmallVector<SDValue, 16> Elts;
19782 for (unsigned i = 0, e = NumElts; i != e; ++i)
19783 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19784 DAG.getIntPtrConstant(i)));
19786 // Explicitly mark the extra elements as Undef.
19787 Elts.append(NumElts, DAG.getUNDEF(SVT));
19789 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19790 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19791 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
19792 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19793 DAG.getIntPtrConstant(0));
19796 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19797 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19798 assert((DstVT == MVT::i64 ||
19799 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19800 "Unexpected custom BITCAST");
19801 // i64 <=> MMX conversions are Legal.
19802 if (SrcVT==MVT::i64 && DstVT.isVector())
19804 if (DstVT==MVT::i64 && SrcVT.isVector())
19806 // MMX <=> MMX conversions are Legal.
19807 if (SrcVT.isVector() && DstVT.isVector())
19809 // All other conversions need to be expanded.
19813 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19814 SelectionDAG &DAG) {
19815 SDNode *Node = Op.getNode();
19818 Op = Op.getOperand(0);
19819 EVT VT = Op.getValueType();
19820 assert((VT.is128BitVector() || VT.is256BitVector()) &&
19821 "CTPOP lowering only implemented for 128/256-bit wide vector types");
19823 unsigned NumElts = VT.getVectorNumElements();
19824 EVT EltVT = VT.getVectorElementType();
19825 unsigned Len = EltVT.getSizeInBits();
19827 // This is the vectorized version of the "best" algorithm from
19828 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19829 // with a minor tweak to use a series of adds + shifts instead of vector
19830 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
19832 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
19833 // v8i32 => Always profitable
19835 // FIXME: There a couple of possible improvements:
19837 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
19838 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
19840 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
19841 "CTPOP not implemented for this vector element type.");
19843 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
19844 // extra legalization.
19845 bool NeedsBitcast = EltVT == MVT::i32;
19846 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
19848 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
19849 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
19850 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
19852 // v = v - ((v >> 1) & 0x55555555...)
19853 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
19854 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
19855 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
19857 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19859 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
19860 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
19862 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
19864 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
19865 if (VT != And.getValueType())
19866 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19867 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
19869 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19870 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
19871 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
19872 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
19873 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
19875 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
19876 if (NeedsBitcast) {
19877 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19878 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
19879 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
19882 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
19883 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
19884 if (VT != AndRHS.getValueType()) {
19885 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
19886 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
19888 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
19890 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19891 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
19892 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
19893 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
19894 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19896 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
19897 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
19898 if (NeedsBitcast) {
19899 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19900 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
19902 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
19903 if (VT != And.getValueType())
19904 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19906 // The algorithm mentioned above uses:
19907 // v = (v * 0x01010101...) >> (Len - 8)
19909 // Change it to use vector adds + vector shifts which yield faster results on
19910 // Haswell than using vector integer multiplication.
19912 // For i32 elements:
19913 // v = v + (v >> 8)
19914 // v = v + (v >> 16)
19916 // For i64 elements:
19917 // v = v + (v >> 8)
19918 // v = v + (v >> 16)
19919 // v = v + (v >> 32)
19922 SmallVector<SDValue, 8> Csts;
19923 for (unsigned i = 8; i <= Len/2; i *= 2) {
19924 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
19925 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
19926 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
19927 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19931 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
19932 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
19933 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
19934 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
19935 if (NeedsBitcast) {
19936 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19937 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
19939 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
19940 if (VT != And.getValueType())
19941 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19946 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19947 SDNode *Node = Op.getNode();
19949 EVT T = Node->getValueType(0);
19950 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19951 DAG.getConstant(0, T), Node->getOperand(2));
19952 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19953 cast<AtomicSDNode>(Node)->getMemoryVT(),
19954 Node->getOperand(0),
19955 Node->getOperand(1), negOp,
19956 cast<AtomicSDNode>(Node)->getMemOperand(),
19957 cast<AtomicSDNode>(Node)->getOrdering(),
19958 cast<AtomicSDNode>(Node)->getSynchScope());
19961 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19962 SDNode *Node = Op.getNode();
19964 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19966 // Convert seq_cst store -> xchg
19967 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19968 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19969 // (The only way to get a 16-byte store is cmpxchg16b)
19970 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19971 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19972 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19973 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19974 cast<AtomicSDNode>(Node)->getMemoryVT(),
19975 Node->getOperand(0),
19976 Node->getOperand(1), Node->getOperand(2),
19977 cast<AtomicSDNode>(Node)->getMemOperand(),
19978 cast<AtomicSDNode>(Node)->getOrdering(),
19979 cast<AtomicSDNode>(Node)->getSynchScope());
19980 return Swap.getValue(1);
19982 // Other atomic stores have a simple pattern.
19986 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19987 EVT VT = Op.getNode()->getSimpleValueType(0);
19989 // Let legalize expand this if it isn't a legal type yet.
19990 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19993 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19996 bool ExtraOp = false;
19997 switch (Op.getOpcode()) {
19998 default: llvm_unreachable("Invalid code");
19999 case ISD::ADDC: Opc = X86ISD::ADD; break;
20000 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
20001 case ISD::SUBC: Opc = X86ISD::SUB; break;
20002 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
20006 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
20008 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
20009 Op.getOperand(1), Op.getOperand(2));
20012 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
20013 SelectionDAG &DAG) {
20014 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
20016 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
20017 // which returns the values as { float, float } (in XMM0) or
20018 // { double, double } (which is returned in XMM0, XMM1).
20020 SDValue Arg = Op.getOperand(0);
20021 EVT ArgVT = Arg.getValueType();
20022 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
20024 TargetLowering::ArgListTy Args;
20025 TargetLowering::ArgListEntry Entry;
20029 Entry.isSExt = false;
20030 Entry.isZExt = false;
20031 Args.push_back(Entry);
20033 bool isF64 = ArgVT == MVT::f64;
20034 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
20035 // the small struct {f32, f32} is returned in (eax, edx). For f64,
20036 // the results are returned via SRet in memory.
20037 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
20038 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20039 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
20041 Type *RetTy = isF64
20042 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
20043 : (Type*)VectorType::get(ArgTy, 4);
20045 TargetLowering::CallLoweringInfo CLI(DAG);
20046 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
20047 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
20049 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
20052 // Returned in xmm0 and xmm1.
20053 return CallResult.first;
20055 // Returned in bits 0:31 and 32:64 xmm0.
20056 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
20057 CallResult.first, DAG.getIntPtrConstant(0));
20058 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
20059 CallResult.first, DAG.getIntPtrConstant(1));
20060 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
20061 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
20064 /// LowerOperation - Provide custom lowering hooks for some operations.
20066 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20067 switch (Op.getOpcode()) {
20068 default: llvm_unreachable("Should not custom lower this!");
20069 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
20070 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20071 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20072 return LowerCMP_SWAP(Op, Subtarget, DAG);
20073 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20074 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20075 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20076 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20077 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
20078 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
20079 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20080 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20081 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20082 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20083 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20084 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20085 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20086 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20087 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20088 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20089 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20090 case ISD::SHL_PARTS:
20091 case ISD::SRA_PARTS:
20092 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20093 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20094 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20095 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20096 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20097 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20098 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20099 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20100 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20101 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20102 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20104 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20105 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20106 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20107 case ISD::SETCC: return LowerSETCC(Op, DAG);
20108 case ISD::SELECT: return LowerSELECT(Op, DAG);
20109 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20110 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20111 case ISD::VASTART: return LowerVASTART(Op, DAG);
20112 case ISD::VAARG: return LowerVAARG(Op, DAG);
20113 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20114 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20115 case ISD::INTRINSIC_VOID:
20116 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20117 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20118 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20119 case ISD::FRAME_TO_ARGS_OFFSET:
20120 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20121 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20122 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20123 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20124 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20125 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20126 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20127 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20128 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
20129 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
20130 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
20131 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20132 case ISD::UMUL_LOHI:
20133 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20136 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20142 case ISD::UMULO: return LowerXALUO(Op, DAG);
20143 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20144 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20148 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20149 case ISD::ADD: return LowerADD(Op, DAG);
20150 case ISD::SUB: return LowerSUB(Op, DAG);
20151 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20155 /// ReplaceNodeResults - Replace a node with an illegal result type
20156 /// with a new node built out of custom code.
20157 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20158 SmallVectorImpl<SDValue>&Results,
20159 SelectionDAG &DAG) const {
20161 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20162 switch (N->getOpcode()) {
20164 llvm_unreachable("Do not know how to custom type legalize this operation!");
20165 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20166 case X86ISD::FMINC:
20168 case X86ISD::FMAXC:
20169 case X86ISD::FMAX: {
20170 EVT VT = N->getValueType(0);
20171 if (VT != MVT::v2f32)
20172 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
20173 SDValue UNDEF = DAG.getUNDEF(VT);
20174 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20175 N->getOperand(0), UNDEF);
20176 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20177 N->getOperand(1), UNDEF);
20178 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20181 case ISD::SIGN_EXTEND_INREG:
20186 // We don't want to expand or promote these.
20193 case ISD::UDIVREM: {
20194 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20195 Results.push_back(V);
20198 case ISD::FP_TO_SINT:
20199 case ISD::FP_TO_UINT: {
20200 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20202 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
20205 std::pair<SDValue,SDValue> Vals =
20206 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20207 SDValue FIST = Vals.first, StackSlot = Vals.second;
20208 if (FIST.getNode()) {
20209 EVT VT = N->getValueType(0);
20210 // Return a load from the stack slot.
20211 if (StackSlot.getNode())
20212 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20213 MachinePointerInfo(),
20214 false, false, false, 0));
20216 Results.push_back(FIST);
20220 case ISD::UINT_TO_FP: {
20221 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20222 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20223 N->getValueType(0) != MVT::v2f32)
20225 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20227 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
20229 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20230 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20231 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
20232 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
20233 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20234 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20237 case ISD::FP_ROUND: {
20238 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20240 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20241 Results.push_back(V);
20244 case ISD::INTRINSIC_W_CHAIN: {
20245 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20247 default : llvm_unreachable("Do not know how to custom type "
20248 "legalize this intrinsic operation!");
20249 case Intrinsic::x86_rdtsc:
20250 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20252 case Intrinsic::x86_rdtscp:
20253 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20255 case Intrinsic::x86_rdpmc:
20256 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20259 case ISD::READCYCLECOUNTER: {
20260 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20263 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20264 EVT T = N->getValueType(0);
20265 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20266 bool Regs64bit = T == MVT::i128;
20267 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20268 SDValue cpInL, cpInH;
20269 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20270 DAG.getConstant(0, HalfT));
20271 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20272 DAG.getConstant(1, HalfT));
20273 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20274 Regs64bit ? X86::RAX : X86::EAX,
20276 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20277 Regs64bit ? X86::RDX : X86::EDX,
20278 cpInH, cpInL.getValue(1));
20279 SDValue swapInL, swapInH;
20280 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20281 DAG.getConstant(0, HalfT));
20282 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20283 DAG.getConstant(1, HalfT));
20284 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20285 Regs64bit ? X86::RBX : X86::EBX,
20286 swapInL, cpInH.getValue(1));
20287 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20288 Regs64bit ? X86::RCX : X86::ECX,
20289 swapInH, swapInL.getValue(1));
20290 SDValue Ops[] = { swapInH.getValue(0),
20292 swapInH.getValue(1) };
20293 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20294 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20295 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20296 X86ISD::LCMPXCHG8_DAG;
20297 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20298 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20299 Regs64bit ? X86::RAX : X86::EAX,
20300 HalfT, Result.getValue(1));
20301 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20302 Regs64bit ? X86::RDX : X86::EDX,
20303 HalfT, cpOutL.getValue(2));
20304 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20306 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20307 MVT::i32, cpOutH.getValue(2));
20309 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20310 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
20311 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20313 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20314 Results.push_back(Success);
20315 Results.push_back(EFLAGS.getValue(1));
20318 case ISD::ATOMIC_SWAP:
20319 case ISD::ATOMIC_LOAD_ADD:
20320 case ISD::ATOMIC_LOAD_SUB:
20321 case ISD::ATOMIC_LOAD_AND:
20322 case ISD::ATOMIC_LOAD_OR:
20323 case ISD::ATOMIC_LOAD_XOR:
20324 case ISD::ATOMIC_LOAD_NAND:
20325 case ISD::ATOMIC_LOAD_MIN:
20326 case ISD::ATOMIC_LOAD_MAX:
20327 case ISD::ATOMIC_LOAD_UMIN:
20328 case ISD::ATOMIC_LOAD_UMAX:
20329 case ISD::ATOMIC_LOAD: {
20330 // Delegate to generic TypeLegalization. Situations we can really handle
20331 // should have already been dealt with by AtomicExpandPass.cpp.
20334 case ISD::BITCAST: {
20335 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20336 EVT DstVT = N->getValueType(0);
20337 EVT SrcVT = N->getOperand(0)->getValueType(0);
20339 if (SrcVT != MVT::f64 ||
20340 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20343 unsigned NumElts = DstVT.getVectorNumElements();
20344 EVT SVT = DstVT.getVectorElementType();
20345 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20346 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20347 MVT::v2f64, N->getOperand(0));
20348 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
20350 if (ExperimentalVectorWideningLegalization) {
20351 // If we are legalizing vectors by widening, we already have the desired
20352 // legal vector type, just return it.
20353 Results.push_back(ToVecInt);
20357 SmallVector<SDValue, 8> Elts;
20358 for (unsigned i = 0, e = NumElts; i != e; ++i)
20359 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20360 ToVecInt, DAG.getIntPtrConstant(i)));
20362 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20367 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20369 default: return nullptr;
20370 case X86ISD::BSF: return "X86ISD::BSF";
20371 case X86ISD::BSR: return "X86ISD::BSR";
20372 case X86ISD::SHLD: return "X86ISD::SHLD";
20373 case X86ISD::SHRD: return "X86ISD::SHRD";
20374 case X86ISD::FAND: return "X86ISD::FAND";
20375 case X86ISD::FANDN: return "X86ISD::FANDN";
20376 case X86ISD::FOR: return "X86ISD::FOR";
20377 case X86ISD::FXOR: return "X86ISD::FXOR";
20378 case X86ISD::FSRL: return "X86ISD::FSRL";
20379 case X86ISD::FILD: return "X86ISD::FILD";
20380 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20381 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20382 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20383 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20384 case X86ISD::FLD: return "X86ISD::FLD";
20385 case X86ISD::FST: return "X86ISD::FST";
20386 case X86ISD::CALL: return "X86ISD::CALL";
20387 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20388 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20389 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20390 case X86ISD::BT: return "X86ISD::BT";
20391 case X86ISD::CMP: return "X86ISD::CMP";
20392 case X86ISD::COMI: return "X86ISD::COMI";
20393 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20394 case X86ISD::CMPM: return "X86ISD::CMPM";
20395 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20396 case X86ISD::SETCC: return "X86ISD::SETCC";
20397 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20398 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20399 case X86ISD::CMOV: return "X86ISD::CMOV";
20400 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20401 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20402 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20403 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20404 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20405 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20406 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20407 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20408 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20409 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20410 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20411 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20412 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20413 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20414 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20415 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20416 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20417 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20418 case X86ISD::HADD: return "X86ISD::HADD";
20419 case X86ISD::HSUB: return "X86ISD::HSUB";
20420 case X86ISD::FHADD: return "X86ISD::FHADD";
20421 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20422 case X86ISD::UMAX: return "X86ISD::UMAX";
20423 case X86ISD::UMIN: return "X86ISD::UMIN";
20424 case X86ISD::SMAX: return "X86ISD::SMAX";
20425 case X86ISD::SMIN: return "X86ISD::SMIN";
20426 case X86ISD::FMAX: return "X86ISD::FMAX";
20427 case X86ISD::FMIN: return "X86ISD::FMIN";
20428 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20429 case X86ISD::FMINC: return "X86ISD::FMINC";
20430 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20431 case X86ISD::FRCP: return "X86ISD::FRCP";
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::VTRUNCM: return "X86ISD::VTRUNCM";
20450 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20451 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20452 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20453 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20454 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20455 case X86ISD::VSHL: return "X86ISD::VSHL";
20456 case X86ISD::VSRL: return "X86ISD::VSRL";
20457 case X86ISD::VSRA: return "X86ISD::VSRA";
20458 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20459 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20460 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20461 case X86ISD::CMPP: return "X86ISD::CMPP";
20462 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20463 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20464 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20465 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20466 case X86ISD::ADD: return "X86ISD::ADD";
20467 case X86ISD::SUB: return "X86ISD::SUB";
20468 case X86ISD::ADC: return "X86ISD::ADC";
20469 case X86ISD::SBB: return "X86ISD::SBB";
20470 case X86ISD::SMUL: return "X86ISD::SMUL";
20471 case X86ISD::UMUL: return "X86ISD::UMUL";
20472 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20473 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20474 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20475 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20476 case X86ISD::INC: return "X86ISD::INC";
20477 case X86ISD::DEC: return "X86ISD::DEC";
20478 case X86ISD::OR: return "X86ISD::OR";
20479 case X86ISD::XOR: return "X86ISD::XOR";
20480 case X86ISD::AND: return "X86ISD::AND";
20481 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20482 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20483 case X86ISD::PTEST: return "X86ISD::PTEST";
20484 case X86ISD::TESTP: return "X86ISD::TESTP";
20485 case X86ISD::TESTM: return "X86ISD::TESTM";
20486 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20487 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20488 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20489 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20490 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20491 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20492 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20493 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20494 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20495 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20496 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20497 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20498 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20499 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20500 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20501 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20502 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20503 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20504 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20505 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20506 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20507 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20508 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20509 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20510 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20511 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20512 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20513 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20514 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20515 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20516 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20517 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20518 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20519 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20520 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20521 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20522 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20523 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20524 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
20525 case X86ISD::SAHF: return "X86ISD::SAHF";
20526 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20527 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20528 case X86ISD::FMADD: return "X86ISD::FMADD";
20529 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20530 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20531 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20532 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20533 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20534 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20535 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20536 case X86ISD::XTEST: return "X86ISD::XTEST";
20537 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20538 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20539 case X86ISD::SELECT: return "X86ISD::SELECT";
20540 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20541 case X86ISD::RCP28: return "X86ISD::RCP28";
20542 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20543 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20544 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20545 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20546 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20550 // isLegalAddressingMode - Return true if the addressing mode represented
20551 // by AM is legal for this target, for a load/store of the specified type.
20552 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
20554 // X86 supports extremely general addressing modes.
20555 CodeModel::Model M = getTargetMachine().getCodeModel();
20556 Reloc::Model R = getTargetMachine().getRelocationModel();
20558 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20559 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20564 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20566 // If a reference to this global requires an extra load, we can't fold it.
20567 if (isGlobalStubReference(GVFlags))
20570 // If BaseGV requires a register for the PIC base, we cannot also have a
20571 // BaseReg specified.
20572 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20575 // If lower 4G is not available, then we must use rip-relative addressing.
20576 if ((M != CodeModel::Small || R != Reloc::Static) &&
20577 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20581 switch (AM.Scale) {
20587 // These scales always work.
20592 // These scales are formed with basereg+scalereg. Only accept if there is
20597 default: // Other stuff never works.
20604 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20605 unsigned Bits = Ty->getScalarSizeInBits();
20607 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20608 // particularly cheaper than those without.
20612 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20613 // variable shifts just as cheap as scalar ones.
20614 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20617 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20618 // fully general vector.
20622 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20623 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20625 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20626 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20627 return NumBits1 > NumBits2;
20630 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20631 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20634 if (!isTypeLegal(EVT::getEVT(Ty1)))
20637 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20639 // Assuming the caller doesn't have a zeroext or signext return parameter,
20640 // truncation all the way down to i1 is valid.
20644 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20645 return isInt<32>(Imm);
20648 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20649 // Can also use sub to handle negated immediates.
20650 return isInt<32>(Imm);
20653 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20654 if (!VT1.isInteger() || !VT2.isInteger())
20656 unsigned NumBits1 = VT1.getSizeInBits();
20657 unsigned NumBits2 = VT2.getSizeInBits();
20658 return NumBits1 > NumBits2;
20661 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20662 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20663 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20666 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20667 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20668 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20671 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20672 EVT VT1 = Val.getValueType();
20673 if (isZExtFree(VT1, VT2))
20676 if (Val.getOpcode() != ISD::LOAD)
20679 if (!VT1.isSimple() || !VT1.isInteger() ||
20680 !VT2.isSimple() || !VT2.isInteger())
20683 switch (VT1.getSimpleVT().SimpleTy) {
20688 // X86 has 8, 16, and 32-bit zero-extending loads.
20695 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20698 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20699 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
20702 VT = VT.getScalarType();
20704 if (!VT.isSimple())
20707 switch (VT.getSimpleVT().SimpleTy) {
20718 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20719 // i16 instructions are longer (0x66 prefix) and potentially slower.
20720 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20723 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20724 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20725 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20726 /// are assumed to be legal.
20728 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20730 if (!VT.isSimple())
20733 MVT SVT = VT.getSimpleVT();
20735 // Very little shuffling can be done for 64-bit vectors right now.
20736 if (VT.getSizeInBits() == 64)
20739 // This is an experimental legality test that is tailored to match the
20740 // legality test of the experimental lowering more closely. They are gated
20741 // separately to ease testing of performance differences.
20742 if (ExperimentalVectorShuffleLegality)
20743 // We only care that the types being shuffled are legal. The lowering can
20744 // handle any possible shuffle mask that results.
20745 return isTypeLegal(SVT);
20747 // If this is a single-input shuffle with no 128 bit lane crossings we can
20748 // lower it into pshufb.
20749 if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
20750 (SVT.is256BitVector() && Subtarget->hasInt256())) {
20751 bool isLegal = true;
20752 for (unsigned I = 0, E = M.size(); I != E; ++I) {
20753 if (M[I] >= (int)SVT.getVectorNumElements() ||
20754 ShuffleCrosses128bitLane(SVT, I, M[I])) {
20763 // FIXME: blends, shifts.
20764 return (SVT.getVectorNumElements() == 2 ||
20765 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
20766 isMOVLMask(M, SVT) ||
20767 isCommutedMOVLMask(M, SVT) ||
20768 isMOVHLPSMask(M, SVT) ||
20769 isSHUFPMask(M, SVT) ||
20770 isSHUFPMask(M, SVT, /* Commuted */ true) ||
20771 isPSHUFDMask(M, SVT) ||
20772 isPSHUFDMask(M, SVT, /* SecondOperand */ true) ||
20773 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
20774 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
20775 isPALIGNRMask(M, SVT, Subtarget) ||
20776 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
20777 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
20778 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20779 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20780 isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()) ||
20781 (Subtarget->hasSSE41() && isINSERTPSMask(M, SVT)));
20785 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20787 if (!VT.isSimple())
20790 MVT SVT = VT.getSimpleVT();
20792 // This is an experimental legality test that is tailored to match the
20793 // legality test of the experimental lowering more closely. They are gated
20794 // separately to ease testing of performance differences.
20795 if (ExperimentalVectorShuffleLegality)
20796 // The new vector shuffle lowering is very good at managing zero-inputs.
20797 return isShuffleMaskLegal(Mask, VT);
20799 unsigned NumElts = SVT.getVectorNumElements();
20800 // FIXME: This collection of masks seems suspect.
20803 if (NumElts == 4 && SVT.is128BitVector()) {
20804 return (isMOVLMask(Mask, SVT) ||
20805 isCommutedMOVLMask(Mask, SVT, true) ||
20806 isSHUFPMask(Mask, SVT) ||
20807 isSHUFPMask(Mask, SVT, /* Commuted */ true) ||
20808 isBlendMask(Mask, SVT, Subtarget->hasSSE41(),
20809 Subtarget->hasInt256()));
20814 //===----------------------------------------------------------------------===//
20815 // X86 Scheduler Hooks
20816 //===----------------------------------------------------------------------===//
20818 /// Utility function to emit xbegin specifying the start of an RTM region.
20819 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20820 const TargetInstrInfo *TII) {
20821 DebugLoc DL = MI->getDebugLoc();
20823 const BasicBlock *BB = MBB->getBasicBlock();
20824 MachineFunction::iterator I = MBB;
20827 // For the v = xbegin(), we generate
20838 MachineBasicBlock *thisMBB = MBB;
20839 MachineFunction *MF = MBB->getParent();
20840 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20841 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20842 MF->insert(I, mainMBB);
20843 MF->insert(I, sinkMBB);
20845 // Transfer the remainder of BB and its successor edges to sinkMBB.
20846 sinkMBB->splice(sinkMBB->begin(), MBB,
20847 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20848 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20852 // # fallthrough to mainMBB
20853 // # abortion to sinkMBB
20854 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20855 thisMBB->addSuccessor(mainMBB);
20856 thisMBB->addSuccessor(sinkMBB);
20860 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20861 mainMBB->addSuccessor(sinkMBB);
20864 // EAX is live into the sinkMBB
20865 sinkMBB->addLiveIn(X86::EAX);
20866 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20867 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20870 MI->eraseFromParent();
20874 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20875 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20876 // in the .td file.
20877 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20878 const TargetInstrInfo *TII) {
20880 switch (MI->getOpcode()) {
20881 default: llvm_unreachable("illegal opcode!");
20882 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20883 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20884 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20885 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20886 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20887 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20888 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20889 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20892 DebugLoc dl = MI->getDebugLoc();
20893 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20895 unsigned NumArgs = MI->getNumOperands();
20896 for (unsigned i = 1; i < NumArgs; ++i) {
20897 MachineOperand &Op = MI->getOperand(i);
20898 if (!(Op.isReg() && Op.isImplicit()))
20899 MIB.addOperand(Op);
20901 if (MI->hasOneMemOperand())
20902 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20904 BuildMI(*BB, MI, dl,
20905 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20906 .addReg(X86::XMM0);
20908 MI->eraseFromParent();
20912 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20913 // defs in an instruction pattern
20914 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20915 const TargetInstrInfo *TII) {
20917 switch (MI->getOpcode()) {
20918 default: llvm_unreachable("illegal opcode!");
20919 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20920 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20921 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20922 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20923 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20924 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20925 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20926 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20929 DebugLoc dl = MI->getDebugLoc();
20930 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20932 unsigned NumArgs = MI->getNumOperands(); // remove the results
20933 for (unsigned i = 1; i < NumArgs; ++i) {
20934 MachineOperand &Op = MI->getOperand(i);
20935 if (!(Op.isReg() && Op.isImplicit()))
20936 MIB.addOperand(Op);
20938 if (MI->hasOneMemOperand())
20939 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20941 BuildMI(*BB, MI, dl,
20942 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20945 MI->eraseFromParent();
20949 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20950 const X86Subtarget *Subtarget) {
20951 DebugLoc dl = MI->getDebugLoc();
20952 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20953 // Address into RAX/EAX, other two args into ECX, EDX.
20954 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20955 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20956 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20957 for (int i = 0; i < X86::AddrNumOperands; ++i)
20958 MIB.addOperand(MI->getOperand(i));
20960 unsigned ValOps = X86::AddrNumOperands;
20961 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20962 .addReg(MI->getOperand(ValOps).getReg());
20963 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20964 .addReg(MI->getOperand(ValOps+1).getReg());
20966 // The instruction doesn't actually take any operands though.
20967 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20969 MI->eraseFromParent(); // The pseudo is gone now.
20973 MachineBasicBlock *
20974 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20975 MachineBasicBlock *MBB) const {
20976 // Emit va_arg instruction on X86-64.
20978 // Operands to this pseudo-instruction:
20979 // 0 ) Output : destination address (reg)
20980 // 1-5) Input : va_list address (addr, i64mem)
20981 // 6 ) ArgSize : Size (in bytes) of vararg type
20982 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20983 // 8 ) Align : Alignment of type
20984 // 9 ) EFLAGS (implicit-def)
20986 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20987 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
20989 unsigned DestReg = MI->getOperand(0).getReg();
20990 MachineOperand &Base = MI->getOperand(1);
20991 MachineOperand &Scale = MI->getOperand(2);
20992 MachineOperand &Index = MI->getOperand(3);
20993 MachineOperand &Disp = MI->getOperand(4);
20994 MachineOperand &Segment = MI->getOperand(5);
20995 unsigned ArgSize = MI->getOperand(6).getImm();
20996 unsigned ArgMode = MI->getOperand(7).getImm();
20997 unsigned Align = MI->getOperand(8).getImm();
20999 // Memory Reference
21000 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
21001 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21002 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21004 // Machine Information
21005 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21006 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
21007 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
21008 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
21009 DebugLoc DL = MI->getDebugLoc();
21011 // struct va_list {
21014 // i64 overflow_area (address)
21015 // i64 reg_save_area (address)
21017 // sizeof(va_list) = 24
21018 // alignment(va_list) = 8
21020 unsigned TotalNumIntRegs = 6;
21021 unsigned TotalNumXMMRegs = 8;
21022 bool UseGPOffset = (ArgMode == 1);
21023 bool UseFPOffset = (ArgMode == 2);
21024 unsigned MaxOffset = TotalNumIntRegs * 8 +
21025 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21027 /* Align ArgSize to a multiple of 8 */
21028 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21029 bool NeedsAlign = (Align > 8);
21031 MachineBasicBlock *thisMBB = MBB;
21032 MachineBasicBlock *overflowMBB;
21033 MachineBasicBlock *offsetMBB;
21034 MachineBasicBlock *endMBB;
21036 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21037 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21038 unsigned OffsetReg = 0;
21040 if (!UseGPOffset && !UseFPOffset) {
21041 // If we only pull from the overflow region, we don't create a branch.
21042 // We don't need to alter control flow.
21043 OffsetDestReg = 0; // unused
21044 OverflowDestReg = DestReg;
21046 offsetMBB = nullptr;
21047 overflowMBB = thisMBB;
21050 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21051 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21052 // If not, pull from overflow_area. (branch to overflowMBB)
21057 // offsetMBB overflowMBB
21062 // Registers for the PHI in endMBB
21063 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21064 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21066 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21067 MachineFunction *MF = MBB->getParent();
21068 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21069 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21070 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21072 MachineFunction::iterator MBBIter = MBB;
21075 // Insert the new basic blocks
21076 MF->insert(MBBIter, offsetMBB);
21077 MF->insert(MBBIter, overflowMBB);
21078 MF->insert(MBBIter, endMBB);
21080 // Transfer the remainder of MBB and its successor edges to endMBB.
21081 endMBB->splice(endMBB->begin(), thisMBB,
21082 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21083 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21085 // Make offsetMBB and overflowMBB successors of thisMBB
21086 thisMBB->addSuccessor(offsetMBB);
21087 thisMBB->addSuccessor(overflowMBB);
21089 // endMBB is a successor of both offsetMBB and overflowMBB
21090 offsetMBB->addSuccessor(endMBB);
21091 overflowMBB->addSuccessor(endMBB);
21093 // Load the offset value into a register
21094 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21095 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21099 .addDisp(Disp, UseFPOffset ? 4 : 0)
21100 .addOperand(Segment)
21101 .setMemRefs(MMOBegin, MMOEnd);
21103 // Check if there is enough room left to pull this argument.
21104 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21106 .addImm(MaxOffset + 8 - ArgSizeA8);
21108 // Branch to "overflowMBB" if offset >= max
21109 // Fall through to "offsetMBB" otherwise
21110 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21111 .addMBB(overflowMBB);
21114 // In offsetMBB, emit code to use the reg_save_area.
21116 assert(OffsetReg != 0);
21118 // Read the reg_save_area address.
21119 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21120 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21125 .addOperand(Segment)
21126 .setMemRefs(MMOBegin, MMOEnd);
21128 // Zero-extend the offset
21129 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21130 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21133 .addImm(X86::sub_32bit);
21135 // Add the offset to the reg_save_area to get the final address.
21136 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21137 .addReg(OffsetReg64)
21138 .addReg(RegSaveReg);
21140 // Compute the offset for the next argument
21141 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21142 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21144 .addImm(UseFPOffset ? 16 : 8);
21146 // Store it back into the va_list.
21147 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21151 .addDisp(Disp, UseFPOffset ? 4 : 0)
21152 .addOperand(Segment)
21153 .addReg(NextOffsetReg)
21154 .setMemRefs(MMOBegin, MMOEnd);
21157 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21162 // Emit code to use overflow area
21165 // Load the overflow_area address into a register.
21166 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21167 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21172 .addOperand(Segment)
21173 .setMemRefs(MMOBegin, MMOEnd);
21175 // If we need to align it, do so. Otherwise, just copy the address
21176 // to OverflowDestReg.
21178 // Align the overflow address
21179 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21180 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21182 // aligned_addr = (addr + (align-1)) & ~(align-1)
21183 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21184 .addReg(OverflowAddrReg)
21187 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21189 .addImm(~(uint64_t)(Align-1));
21191 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21192 .addReg(OverflowAddrReg);
21195 // Compute the next overflow address after this argument.
21196 // (the overflow address should be kept 8-byte aligned)
21197 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21198 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21199 .addReg(OverflowDestReg)
21200 .addImm(ArgSizeA8);
21202 // Store the new overflow address.
21203 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21208 .addOperand(Segment)
21209 .addReg(NextAddrReg)
21210 .setMemRefs(MMOBegin, MMOEnd);
21212 // If we branched, emit the PHI to the front of endMBB.
21214 BuildMI(*endMBB, endMBB->begin(), DL,
21215 TII->get(X86::PHI), DestReg)
21216 .addReg(OffsetDestReg).addMBB(offsetMBB)
21217 .addReg(OverflowDestReg).addMBB(overflowMBB);
21220 // Erase the pseudo instruction
21221 MI->eraseFromParent();
21226 MachineBasicBlock *
21227 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21229 MachineBasicBlock *MBB) const {
21230 // Emit code to save XMM registers to the stack. The ABI says that the
21231 // number of registers to save is given in %al, so it's theoretically
21232 // possible to do an indirect jump trick to avoid saving all of them,
21233 // however this code takes a simpler approach and just executes all
21234 // of the stores if %al is non-zero. It's less code, and it's probably
21235 // easier on the hardware branch predictor, and stores aren't all that
21236 // expensive anyway.
21238 // Create the new basic blocks. One block contains all the XMM stores,
21239 // and one block is the final destination regardless of whether any
21240 // stores were performed.
21241 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21242 MachineFunction *F = MBB->getParent();
21243 MachineFunction::iterator MBBIter = MBB;
21245 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21246 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21247 F->insert(MBBIter, XMMSaveMBB);
21248 F->insert(MBBIter, EndMBB);
21250 // Transfer the remainder of MBB and its successor edges to EndMBB.
21251 EndMBB->splice(EndMBB->begin(), MBB,
21252 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21253 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21255 // The original block will now fall through to the XMM save block.
21256 MBB->addSuccessor(XMMSaveMBB);
21257 // The XMMSaveMBB will fall through to the end block.
21258 XMMSaveMBB->addSuccessor(EndMBB);
21260 // Now add the instructions.
21261 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21262 DebugLoc DL = MI->getDebugLoc();
21264 unsigned CountReg = MI->getOperand(0).getReg();
21265 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21266 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21268 if (!Subtarget->isTargetWin64()) {
21269 // If %al is 0, branch around the XMM save block.
21270 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21271 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21272 MBB->addSuccessor(EndMBB);
21275 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21276 // that was just emitted, but clearly shouldn't be "saved".
21277 assert((MI->getNumOperands() <= 3 ||
21278 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21279 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21280 && "Expected last argument to be EFLAGS");
21281 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21282 // In the XMM save block, save all the XMM argument registers.
21283 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21284 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21285 MachineMemOperand *MMO =
21286 F->getMachineMemOperand(
21287 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
21288 MachineMemOperand::MOStore,
21289 /*Size=*/16, /*Align=*/16);
21290 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21291 .addFrameIndex(RegSaveFrameIndex)
21292 .addImm(/*Scale=*/1)
21293 .addReg(/*IndexReg=*/0)
21294 .addImm(/*Disp=*/Offset)
21295 .addReg(/*Segment=*/0)
21296 .addReg(MI->getOperand(i).getReg())
21297 .addMemOperand(MMO);
21300 MI->eraseFromParent(); // The pseudo instruction is gone now.
21305 // The EFLAGS operand of SelectItr might be missing a kill marker
21306 // because there were multiple uses of EFLAGS, and ISel didn't know
21307 // which to mark. Figure out whether SelectItr should have had a
21308 // kill marker, and set it if it should. Returns the correct kill
21310 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21311 MachineBasicBlock* BB,
21312 const TargetRegisterInfo* TRI) {
21313 // Scan forward through BB for a use/def of EFLAGS.
21314 MachineBasicBlock::iterator miI(std::next(SelectItr));
21315 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21316 const MachineInstr& mi = *miI;
21317 if (mi.readsRegister(X86::EFLAGS))
21319 if (mi.definesRegister(X86::EFLAGS))
21320 break; // Should have kill-flag - update below.
21323 // If we hit the end of the block, check whether EFLAGS is live into a
21325 if (miI == BB->end()) {
21326 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21327 sEnd = BB->succ_end();
21328 sItr != sEnd; ++sItr) {
21329 MachineBasicBlock* succ = *sItr;
21330 if (succ->isLiveIn(X86::EFLAGS))
21335 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21336 // out. SelectMI should have a kill flag on EFLAGS.
21337 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21341 MachineBasicBlock *
21342 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21343 MachineBasicBlock *BB) const {
21344 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21345 DebugLoc DL = MI->getDebugLoc();
21347 // To "insert" a SELECT_CC instruction, we actually have to insert the
21348 // diamond control-flow pattern. The incoming instruction knows the
21349 // destination vreg to set, the condition code register to branch on, the
21350 // true/false values to select between, and a branch opcode to use.
21351 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21352 MachineFunction::iterator It = BB;
21358 // cmpTY ccX, r1, r2
21360 // fallthrough --> copy0MBB
21361 MachineBasicBlock *thisMBB = BB;
21362 MachineFunction *F = BB->getParent();
21363 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21364 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21365 F->insert(It, copy0MBB);
21366 F->insert(It, sinkMBB);
21368 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21369 // live into the sink and copy blocks.
21370 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21371 if (!MI->killsRegister(X86::EFLAGS) &&
21372 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
21373 copy0MBB->addLiveIn(X86::EFLAGS);
21374 sinkMBB->addLiveIn(X86::EFLAGS);
21377 // Transfer the remainder of BB and its successor edges to sinkMBB.
21378 sinkMBB->splice(sinkMBB->begin(), BB,
21379 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21380 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21382 // Add the true and fallthrough blocks as its successors.
21383 BB->addSuccessor(copy0MBB);
21384 BB->addSuccessor(sinkMBB);
21386 // Create the conditional branch instruction.
21388 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
21389 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21392 // %FalseValue = ...
21393 // # fallthrough to sinkMBB
21394 copy0MBB->addSuccessor(sinkMBB);
21397 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21399 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21400 TII->get(X86::PHI), MI->getOperand(0).getReg())
21401 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
21402 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
21404 MI->eraseFromParent(); // The pseudo instruction is gone now.
21408 MachineBasicBlock *
21409 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21410 MachineBasicBlock *BB) const {
21411 MachineFunction *MF = BB->getParent();
21412 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21413 DebugLoc DL = MI->getDebugLoc();
21414 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21416 assert(MF->shouldSplitStack());
21418 const bool Is64Bit = Subtarget->is64Bit();
21419 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21421 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21422 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21425 // ... [Till the alloca]
21426 // If stacklet is not large enough, jump to mallocMBB
21429 // Allocate by subtracting from RSP
21430 // Jump to continueMBB
21433 // Allocate by call to runtime
21437 // [rest of original BB]
21440 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21441 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21442 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21444 MachineRegisterInfo &MRI = MF->getRegInfo();
21445 const TargetRegisterClass *AddrRegClass =
21446 getRegClassFor(getPointerTy());
21448 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21449 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21450 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21451 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21452 sizeVReg = MI->getOperand(1).getReg(),
21453 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21455 MachineFunction::iterator MBBIter = BB;
21458 MF->insert(MBBIter, bumpMBB);
21459 MF->insert(MBBIter, mallocMBB);
21460 MF->insert(MBBIter, continueMBB);
21462 continueMBB->splice(continueMBB->begin(), BB,
21463 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21464 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21466 // Add code to the main basic block to check if the stack limit has been hit,
21467 // and if so, jump to mallocMBB otherwise to bumpMBB.
21468 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21469 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21470 .addReg(tmpSPVReg).addReg(sizeVReg);
21471 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21472 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21473 .addReg(SPLimitVReg);
21474 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21476 // bumpMBB simply decreases the stack pointer, since we know the current
21477 // stacklet has enough space.
21478 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21479 .addReg(SPLimitVReg);
21480 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21481 .addReg(SPLimitVReg);
21482 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21484 // Calls into a routine in libgcc to allocate more space from the heap.
21485 const uint32_t *RegMask =
21486 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
21488 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21490 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21491 .addExternalSymbol("__morestack_allocate_stack_space")
21492 .addRegMask(RegMask)
21493 .addReg(X86::RDI, RegState::Implicit)
21494 .addReg(X86::RAX, RegState::ImplicitDefine);
21495 } else if (Is64Bit) {
21496 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21498 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21499 .addExternalSymbol("__morestack_allocate_stack_space")
21500 .addRegMask(RegMask)
21501 .addReg(X86::EDI, RegState::Implicit)
21502 .addReg(X86::EAX, RegState::ImplicitDefine);
21504 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21506 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21507 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21508 .addExternalSymbol("__morestack_allocate_stack_space")
21509 .addRegMask(RegMask)
21510 .addReg(X86::EAX, RegState::ImplicitDefine);
21514 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21517 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21518 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21519 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21521 // Set up the CFG correctly.
21522 BB->addSuccessor(bumpMBB);
21523 BB->addSuccessor(mallocMBB);
21524 mallocMBB->addSuccessor(continueMBB);
21525 bumpMBB->addSuccessor(continueMBB);
21527 // Take care of the PHI nodes.
21528 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21529 MI->getOperand(0).getReg())
21530 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21531 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21533 // Delete the original pseudo instruction.
21534 MI->eraseFromParent();
21537 return continueMBB;
21540 MachineBasicBlock *
21541 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21542 MachineBasicBlock *BB) const {
21543 DebugLoc DL = MI->getDebugLoc();
21545 assert(!Subtarget->isTargetMachO());
21547 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
21549 MI->eraseFromParent(); // The pseudo instruction is gone now.
21553 MachineBasicBlock *
21554 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21555 MachineBasicBlock *BB) const {
21556 // This is pretty easy. We're taking the value that we received from
21557 // our load from the relocation, sticking it in either RDI (x86-64)
21558 // or EAX and doing an indirect call. The return value will then
21559 // be in the normal return register.
21560 MachineFunction *F = BB->getParent();
21561 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21562 DebugLoc DL = MI->getDebugLoc();
21564 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21565 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21567 // Get a register mask for the lowered call.
21568 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21569 // proper register mask.
21570 const uint32_t *RegMask =
21571 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
21572 if (Subtarget->is64Bit()) {
21573 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21574 TII->get(X86::MOV64rm), X86::RDI)
21576 .addImm(0).addReg(0)
21577 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21578 MI->getOperand(3).getTargetFlags())
21580 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21581 addDirectMem(MIB, X86::RDI);
21582 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21583 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21584 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21585 TII->get(X86::MOV32rm), X86::EAX)
21587 .addImm(0).addReg(0)
21588 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21589 MI->getOperand(3).getTargetFlags())
21591 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21592 addDirectMem(MIB, X86::EAX);
21593 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21595 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21596 TII->get(X86::MOV32rm), X86::EAX)
21597 .addReg(TII->getGlobalBaseReg(F))
21598 .addImm(0).addReg(0)
21599 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21600 MI->getOperand(3).getTargetFlags())
21602 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21603 addDirectMem(MIB, X86::EAX);
21604 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21607 MI->eraseFromParent(); // The pseudo instruction is gone now.
21611 MachineBasicBlock *
21612 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21613 MachineBasicBlock *MBB) const {
21614 DebugLoc DL = MI->getDebugLoc();
21615 MachineFunction *MF = MBB->getParent();
21616 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21617 MachineRegisterInfo &MRI = MF->getRegInfo();
21619 const BasicBlock *BB = MBB->getBasicBlock();
21620 MachineFunction::iterator I = MBB;
21623 // Memory Reference
21624 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21625 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21628 unsigned MemOpndSlot = 0;
21630 unsigned CurOp = 0;
21632 DstReg = MI->getOperand(CurOp++).getReg();
21633 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21634 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21635 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21636 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21638 MemOpndSlot = CurOp;
21640 MVT PVT = getPointerTy();
21641 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21642 "Invalid Pointer Size!");
21644 // For v = setjmp(buf), we generate
21647 // buf[LabelOffset] = restoreMBB
21648 // SjLjSetup restoreMBB
21654 // v = phi(main, restore)
21657 // if base pointer being used, load it from frame
21660 MachineBasicBlock *thisMBB = MBB;
21661 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21662 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21663 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21664 MF->insert(I, mainMBB);
21665 MF->insert(I, sinkMBB);
21666 MF->push_back(restoreMBB);
21668 MachineInstrBuilder MIB;
21670 // Transfer the remainder of BB and its successor edges to sinkMBB.
21671 sinkMBB->splice(sinkMBB->begin(), MBB,
21672 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21673 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21676 unsigned PtrStoreOpc = 0;
21677 unsigned LabelReg = 0;
21678 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21679 Reloc::Model RM = MF->getTarget().getRelocationModel();
21680 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21681 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21683 // Prepare IP either in reg or imm.
21684 if (!UseImmLabel) {
21685 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21686 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21687 LabelReg = MRI.createVirtualRegister(PtrRC);
21688 if (Subtarget->is64Bit()) {
21689 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21693 .addMBB(restoreMBB)
21696 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21697 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21698 .addReg(XII->getGlobalBaseReg(MF))
21701 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21705 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21707 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21708 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21709 if (i == X86::AddrDisp)
21710 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21712 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21715 MIB.addReg(LabelReg);
21717 MIB.addMBB(restoreMBB);
21718 MIB.setMemRefs(MMOBegin, MMOEnd);
21720 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21721 .addMBB(restoreMBB);
21723 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21724 MIB.addRegMask(RegInfo->getNoPreservedMask());
21725 thisMBB->addSuccessor(mainMBB);
21726 thisMBB->addSuccessor(restoreMBB);
21730 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21731 mainMBB->addSuccessor(sinkMBB);
21734 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21735 TII->get(X86::PHI), DstReg)
21736 .addReg(mainDstReg).addMBB(mainMBB)
21737 .addReg(restoreDstReg).addMBB(restoreMBB);
21740 if (RegInfo->hasBasePointer(*MF)) {
21741 const bool Uses64BitFramePtr =
21742 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21743 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21744 X86FI->setRestoreBasePointer(MF);
21745 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21746 unsigned BasePtr = RegInfo->getBaseRegister();
21747 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21748 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21749 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21750 .setMIFlag(MachineInstr::FrameSetup);
21752 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21753 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21754 restoreMBB->addSuccessor(sinkMBB);
21756 MI->eraseFromParent();
21760 MachineBasicBlock *
21761 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21762 MachineBasicBlock *MBB) const {
21763 DebugLoc DL = MI->getDebugLoc();
21764 MachineFunction *MF = MBB->getParent();
21765 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21766 MachineRegisterInfo &MRI = MF->getRegInfo();
21768 // Memory Reference
21769 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21770 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21772 MVT PVT = getPointerTy();
21773 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21774 "Invalid Pointer Size!");
21776 const TargetRegisterClass *RC =
21777 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21778 unsigned Tmp = MRI.createVirtualRegister(RC);
21779 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21780 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21781 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21782 unsigned SP = RegInfo->getStackRegister();
21784 MachineInstrBuilder MIB;
21786 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21787 const int64_t SPOffset = 2 * PVT.getStoreSize();
21789 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21790 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21793 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21794 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21795 MIB.addOperand(MI->getOperand(i));
21796 MIB.setMemRefs(MMOBegin, MMOEnd);
21798 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21799 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21800 if (i == X86::AddrDisp)
21801 MIB.addDisp(MI->getOperand(i), LabelOffset);
21803 MIB.addOperand(MI->getOperand(i));
21805 MIB.setMemRefs(MMOBegin, MMOEnd);
21807 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21808 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21809 if (i == X86::AddrDisp)
21810 MIB.addDisp(MI->getOperand(i), SPOffset);
21812 MIB.addOperand(MI->getOperand(i));
21814 MIB.setMemRefs(MMOBegin, MMOEnd);
21816 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21818 MI->eraseFromParent();
21822 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21823 // accumulator loops. Writing back to the accumulator allows the coalescer
21824 // to remove extra copies in the loop.
21825 MachineBasicBlock *
21826 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21827 MachineBasicBlock *MBB) const {
21828 MachineOperand &AddendOp = MI->getOperand(3);
21830 // Bail out early if the addend isn't a register - we can't switch these.
21831 if (!AddendOp.isReg())
21834 MachineFunction &MF = *MBB->getParent();
21835 MachineRegisterInfo &MRI = MF.getRegInfo();
21837 // Check whether the addend is defined by a PHI:
21838 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21839 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21840 if (!AddendDef.isPHI())
21843 // Look for the following pattern:
21845 // %addend = phi [%entry, 0], [%loop, %result]
21847 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21851 // %addend = phi [%entry, 0], [%loop, %result]
21853 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21855 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21856 assert(AddendDef.getOperand(i).isReg());
21857 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21858 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21859 if (&PHISrcInst == MI) {
21860 // Found a matching instruction.
21861 unsigned NewFMAOpc = 0;
21862 switch (MI->getOpcode()) {
21863 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21864 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21865 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21866 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21867 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21868 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21869 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21870 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21871 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21872 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21873 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21874 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21875 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21876 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21877 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21878 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21879 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21880 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21881 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21882 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21884 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21885 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21886 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21887 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21888 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21889 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21890 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21891 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21892 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21893 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21894 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21895 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21896 default: llvm_unreachable("Unrecognized FMA variant.");
21899 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21900 MachineInstrBuilder MIB =
21901 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21902 .addOperand(MI->getOperand(0))
21903 .addOperand(MI->getOperand(3))
21904 .addOperand(MI->getOperand(2))
21905 .addOperand(MI->getOperand(1));
21906 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21907 MI->eraseFromParent();
21914 MachineBasicBlock *
21915 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21916 MachineBasicBlock *BB) const {
21917 switch (MI->getOpcode()) {
21918 default: llvm_unreachable("Unexpected instr type to insert");
21919 case X86::TAILJMPd64:
21920 case X86::TAILJMPr64:
21921 case X86::TAILJMPm64:
21922 case X86::TAILJMPd64_REX:
21923 case X86::TAILJMPr64_REX:
21924 case X86::TAILJMPm64_REX:
21925 llvm_unreachable("TAILJMP64 would not be touched here.");
21926 case X86::TCRETURNdi64:
21927 case X86::TCRETURNri64:
21928 case X86::TCRETURNmi64:
21930 case X86::WIN_ALLOCA:
21931 return EmitLoweredWinAlloca(MI, BB);
21932 case X86::SEG_ALLOCA_32:
21933 case X86::SEG_ALLOCA_64:
21934 return EmitLoweredSegAlloca(MI, BB);
21935 case X86::TLSCall_32:
21936 case X86::TLSCall_64:
21937 return EmitLoweredTLSCall(MI, BB);
21938 case X86::CMOV_GR8:
21939 case X86::CMOV_FR32:
21940 case X86::CMOV_FR64:
21941 case X86::CMOV_V4F32:
21942 case X86::CMOV_V2F64:
21943 case X86::CMOV_V2I64:
21944 case X86::CMOV_V8F32:
21945 case X86::CMOV_V4F64:
21946 case X86::CMOV_V4I64:
21947 case X86::CMOV_V16F32:
21948 case X86::CMOV_V8F64:
21949 case X86::CMOV_V8I64:
21950 case X86::CMOV_GR16:
21951 case X86::CMOV_GR32:
21952 case X86::CMOV_RFP32:
21953 case X86::CMOV_RFP64:
21954 case X86::CMOV_RFP80:
21955 return EmitLoweredSelect(MI, BB);
21957 case X86::FP32_TO_INT16_IN_MEM:
21958 case X86::FP32_TO_INT32_IN_MEM:
21959 case X86::FP32_TO_INT64_IN_MEM:
21960 case X86::FP64_TO_INT16_IN_MEM:
21961 case X86::FP64_TO_INT32_IN_MEM:
21962 case X86::FP64_TO_INT64_IN_MEM:
21963 case X86::FP80_TO_INT16_IN_MEM:
21964 case X86::FP80_TO_INT32_IN_MEM:
21965 case X86::FP80_TO_INT64_IN_MEM: {
21966 MachineFunction *F = BB->getParent();
21967 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21968 DebugLoc DL = MI->getDebugLoc();
21970 // Change the floating point control register to use "round towards zero"
21971 // mode when truncating to an integer value.
21972 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21973 addFrameReference(BuildMI(*BB, MI, DL,
21974 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21976 // Load the old value of the high byte of the control word...
21978 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21979 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21982 // Set the high part to be round to zero...
21983 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21986 // Reload the modified control word now...
21987 addFrameReference(BuildMI(*BB, MI, DL,
21988 TII->get(X86::FLDCW16m)), CWFrameIdx);
21990 // Restore the memory image of control word to original value
21991 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21994 // Get the X86 opcode to use.
21996 switch (MI->getOpcode()) {
21997 default: llvm_unreachable("illegal opcode!");
21998 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21999 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22000 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22001 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22002 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22003 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22004 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22005 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22006 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22010 MachineOperand &Op = MI->getOperand(0);
22012 AM.BaseType = X86AddressMode::RegBase;
22013 AM.Base.Reg = Op.getReg();
22015 AM.BaseType = X86AddressMode::FrameIndexBase;
22016 AM.Base.FrameIndex = Op.getIndex();
22018 Op = MI->getOperand(1);
22020 AM.Scale = Op.getImm();
22021 Op = MI->getOperand(2);
22023 AM.IndexReg = Op.getImm();
22024 Op = MI->getOperand(3);
22025 if (Op.isGlobal()) {
22026 AM.GV = Op.getGlobal();
22028 AM.Disp = Op.getImm();
22030 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22031 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22033 // Reload the original control word now.
22034 addFrameReference(BuildMI(*BB, MI, DL,
22035 TII->get(X86::FLDCW16m)), CWFrameIdx);
22037 MI->eraseFromParent(); // The pseudo instruction is gone now.
22040 // String/text processing lowering.
22041 case X86::PCMPISTRM128REG:
22042 case X86::VPCMPISTRM128REG:
22043 case X86::PCMPISTRM128MEM:
22044 case X86::VPCMPISTRM128MEM:
22045 case X86::PCMPESTRM128REG:
22046 case X86::VPCMPESTRM128REG:
22047 case X86::PCMPESTRM128MEM:
22048 case X86::VPCMPESTRM128MEM:
22049 assert(Subtarget->hasSSE42() &&
22050 "Target must have SSE4.2 or AVX features enabled");
22051 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22053 // String/text processing lowering.
22054 case X86::PCMPISTRIREG:
22055 case X86::VPCMPISTRIREG:
22056 case X86::PCMPISTRIMEM:
22057 case X86::VPCMPISTRIMEM:
22058 case X86::PCMPESTRIREG:
22059 case X86::VPCMPESTRIREG:
22060 case X86::PCMPESTRIMEM:
22061 case X86::VPCMPESTRIMEM:
22062 assert(Subtarget->hasSSE42() &&
22063 "Target must have SSE4.2 or AVX features enabled");
22064 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22066 // Thread synchronization.
22068 return EmitMonitor(MI, BB, Subtarget);
22072 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22074 case X86::VASTART_SAVE_XMM_REGS:
22075 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22077 case X86::VAARG_64:
22078 return EmitVAARG64WithCustomInserter(MI, BB);
22080 case X86::EH_SjLj_SetJmp32:
22081 case X86::EH_SjLj_SetJmp64:
22082 return emitEHSjLjSetJmp(MI, BB);
22084 case X86::EH_SjLj_LongJmp32:
22085 case X86::EH_SjLj_LongJmp64:
22086 return emitEHSjLjLongJmp(MI, BB);
22088 case TargetOpcode::STATEPOINT:
22089 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22090 // this point in the process. We diverge later.
22091 return emitPatchPoint(MI, BB);
22093 case TargetOpcode::STACKMAP:
22094 case TargetOpcode::PATCHPOINT:
22095 return emitPatchPoint(MI, BB);
22097 case X86::VFMADDPDr213r:
22098 case X86::VFMADDPSr213r:
22099 case X86::VFMADDSDr213r:
22100 case X86::VFMADDSSr213r:
22101 case X86::VFMSUBPDr213r:
22102 case X86::VFMSUBPSr213r:
22103 case X86::VFMSUBSDr213r:
22104 case X86::VFMSUBSSr213r:
22105 case X86::VFNMADDPDr213r:
22106 case X86::VFNMADDPSr213r:
22107 case X86::VFNMADDSDr213r:
22108 case X86::VFNMADDSSr213r:
22109 case X86::VFNMSUBPDr213r:
22110 case X86::VFNMSUBPSr213r:
22111 case X86::VFNMSUBSDr213r:
22112 case X86::VFNMSUBSSr213r:
22113 case X86::VFMADDSUBPDr213r:
22114 case X86::VFMADDSUBPSr213r:
22115 case X86::VFMSUBADDPDr213r:
22116 case X86::VFMSUBADDPSr213r:
22117 case X86::VFMADDPDr213rY:
22118 case X86::VFMADDPSr213rY:
22119 case X86::VFMSUBPDr213rY:
22120 case X86::VFMSUBPSr213rY:
22121 case X86::VFNMADDPDr213rY:
22122 case X86::VFNMADDPSr213rY:
22123 case X86::VFNMSUBPDr213rY:
22124 case X86::VFNMSUBPSr213rY:
22125 case X86::VFMADDSUBPDr213rY:
22126 case X86::VFMADDSUBPSr213rY:
22127 case X86::VFMSUBADDPDr213rY:
22128 case X86::VFMSUBADDPSr213rY:
22129 return emitFMA3Instr(MI, BB);
22133 //===----------------------------------------------------------------------===//
22134 // X86 Optimization Hooks
22135 //===----------------------------------------------------------------------===//
22137 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22140 const SelectionDAG &DAG,
22141 unsigned Depth) const {
22142 unsigned BitWidth = KnownZero.getBitWidth();
22143 unsigned Opc = Op.getOpcode();
22144 assert((Opc >= ISD::BUILTIN_OP_END ||
22145 Opc == ISD::INTRINSIC_WO_CHAIN ||
22146 Opc == ISD::INTRINSIC_W_CHAIN ||
22147 Opc == ISD::INTRINSIC_VOID) &&
22148 "Should use MaskedValueIsZero if you don't know whether Op"
22149 " is a target node!");
22151 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22165 // These nodes' second result is a boolean.
22166 if (Op.getResNo() == 0)
22169 case X86ISD::SETCC:
22170 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22172 case ISD::INTRINSIC_WO_CHAIN: {
22173 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22174 unsigned NumLoBits = 0;
22177 case Intrinsic::x86_sse_movmsk_ps:
22178 case Intrinsic::x86_avx_movmsk_ps_256:
22179 case Intrinsic::x86_sse2_movmsk_pd:
22180 case Intrinsic::x86_avx_movmsk_pd_256:
22181 case Intrinsic::x86_mmx_pmovmskb:
22182 case Intrinsic::x86_sse2_pmovmskb_128:
22183 case Intrinsic::x86_avx2_pmovmskb: {
22184 // High bits of movmskp{s|d}, pmovmskb are known zero.
22186 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22187 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22188 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22189 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22190 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22191 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22192 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22193 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22195 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22204 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22206 const SelectionDAG &,
22207 unsigned Depth) const {
22208 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22209 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22210 return Op.getValueType().getScalarType().getSizeInBits();
22216 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22217 /// node is a GlobalAddress + offset.
22218 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22219 const GlobalValue* &GA,
22220 int64_t &Offset) const {
22221 if (N->getOpcode() == X86ISD::Wrapper) {
22222 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22223 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22224 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22228 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22231 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22232 /// same as extracting the high 128-bit part of 256-bit vector and then
22233 /// inserting the result into the low part of a new 256-bit vector
22234 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22235 EVT VT = SVOp->getValueType(0);
22236 unsigned NumElems = VT.getVectorNumElements();
22238 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22239 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22240 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22241 SVOp->getMaskElt(j) >= 0)
22247 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22248 /// same as extracting the low 128-bit part of 256-bit vector and then
22249 /// inserting the result into the high part of a new 256-bit vector
22250 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22251 EVT VT = SVOp->getValueType(0);
22252 unsigned NumElems = VT.getVectorNumElements();
22254 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22255 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22256 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22257 SVOp->getMaskElt(j) >= 0)
22263 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22264 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22265 TargetLowering::DAGCombinerInfo &DCI,
22266 const X86Subtarget* Subtarget) {
22268 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22269 SDValue V1 = SVOp->getOperand(0);
22270 SDValue V2 = SVOp->getOperand(1);
22271 EVT VT = SVOp->getValueType(0);
22272 unsigned NumElems = VT.getVectorNumElements();
22274 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22275 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22279 // V UNDEF BUILD_VECTOR UNDEF
22281 // CONCAT_VECTOR CONCAT_VECTOR
22284 // RESULT: V + zero extended
22286 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22287 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22288 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22291 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22294 // To match the shuffle mask, the first half of the mask should
22295 // be exactly the first vector, and all the rest a splat with the
22296 // first element of the second one.
22297 for (unsigned i = 0; i != NumElems/2; ++i)
22298 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22299 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22302 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22303 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22304 if (Ld->hasNUsesOfValue(1, 0)) {
22305 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22306 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22308 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22310 Ld->getPointerInfo(),
22311 Ld->getAlignment(),
22312 false/*isVolatile*/, true/*ReadMem*/,
22313 false/*WriteMem*/);
22315 // Make sure the newly-created LOAD is in the same position as Ld in
22316 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22317 // and update uses of Ld's output chain to use the TokenFactor.
22318 if (Ld->hasAnyUseOfValue(1)) {
22319 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22320 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22321 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22322 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22323 SDValue(ResNode.getNode(), 1));
22326 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
22330 // Emit a zeroed vector and insert the desired subvector on its
22332 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22333 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22334 return DCI.CombineTo(N, InsV);
22337 //===--------------------------------------------------------------------===//
22338 // Combine some shuffles into subvector extracts and inserts:
22341 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22342 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22343 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22344 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22345 return DCI.CombineTo(N, InsV);
22348 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22349 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22350 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22351 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22352 return DCI.CombineTo(N, InsV);
22358 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22361 /// This is the leaf of the recursive combinine below. When we have found some
22362 /// chain of single-use x86 shuffle instructions and accumulated the combined
22363 /// shuffle mask represented by them, this will try to pattern match that mask
22364 /// into either a single instruction if there is a special purpose instruction
22365 /// for this operation, or into a PSHUFB instruction which is a fully general
22366 /// instruction but should only be used to replace chains over a certain depth.
22367 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22368 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22369 TargetLowering::DAGCombinerInfo &DCI,
22370 const X86Subtarget *Subtarget) {
22371 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22373 // Find the operand that enters the chain. Note that multiple uses are OK
22374 // here, we're not going to remove the operand we find.
22375 SDValue Input = Op.getOperand(0);
22376 while (Input.getOpcode() == ISD::BITCAST)
22377 Input = Input.getOperand(0);
22379 MVT VT = Input.getSimpleValueType();
22380 MVT RootVT = Root.getSimpleValueType();
22383 // Just remove no-op shuffle masks.
22384 if (Mask.size() == 1) {
22385 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
22390 // Use the float domain if the operand type is a floating point type.
22391 bool FloatDomain = VT.isFloatingPoint();
22393 // For floating point shuffles, we don't have free copies in the shuffle
22394 // instructions or the ability to load as part of the instruction, so
22395 // canonicalize their shuffles to UNPCK or MOV variants.
22397 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22398 // vectors because it can have a load folded into it that UNPCK cannot. This
22399 // doesn't preclude something switching to the shorter encoding post-RA.
22401 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
22402 bool Lo = Mask.equals(0, 0);
22405 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22406 // is no slower than UNPCKLPD but has the option to fold the input operand
22407 // into even an unaligned memory load.
22408 if (Lo && Subtarget->hasSSE3()) {
22409 Shuffle = X86ISD::MOVDDUP;
22410 ShuffleVT = MVT::v2f64;
22412 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22413 // than the UNPCK variants.
22414 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22415 ShuffleVT = MVT::v4f32;
22417 if (Depth == 1 && Root->getOpcode() == Shuffle)
22418 return false; // Nothing to do!
22419 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22420 DCI.AddToWorklist(Op.getNode());
22421 if (Shuffle == X86ISD::MOVDDUP)
22422 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22424 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22425 DCI.AddToWorklist(Op.getNode());
22426 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22430 if (Subtarget->hasSSE3() &&
22431 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
22432 bool Lo = Mask.equals(0, 0, 2, 2);
22433 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22434 MVT ShuffleVT = MVT::v4f32;
22435 if (Depth == 1 && Root->getOpcode() == Shuffle)
22436 return false; // Nothing to do!
22437 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22438 DCI.AddToWorklist(Op.getNode());
22439 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22440 DCI.AddToWorklist(Op.getNode());
22441 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22445 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
22446 bool Lo = Mask.equals(0, 0, 1, 1);
22447 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22448 MVT ShuffleVT = MVT::v4f32;
22449 if (Depth == 1 && Root->getOpcode() == Shuffle)
22450 return false; // Nothing to do!
22451 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22452 DCI.AddToWorklist(Op.getNode());
22453 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22454 DCI.AddToWorklist(Op.getNode());
22455 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22461 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22462 // variants as none of these have single-instruction variants that are
22463 // superior to the UNPCK formulation.
22464 if (!FloatDomain &&
22465 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
22466 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
22467 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
22468 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
22470 bool Lo = Mask[0] == 0;
22471 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22472 if (Depth == 1 && Root->getOpcode() == Shuffle)
22473 return false; // Nothing to do!
22475 switch (Mask.size()) {
22477 ShuffleVT = MVT::v8i16;
22480 ShuffleVT = MVT::v16i8;
22483 llvm_unreachable("Impossible mask size!");
22485 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22486 DCI.AddToWorklist(Op.getNode());
22487 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22488 DCI.AddToWorklist(Op.getNode());
22489 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22494 // Don't try to re-form single instruction chains under any circumstances now
22495 // that we've done encoding canonicalization for them.
22499 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22500 // can replace them with a single PSHUFB instruction profitably. Intel's
22501 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22502 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22503 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22504 SmallVector<SDValue, 16> PSHUFBMask;
22505 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
22506 int Ratio = 16 / Mask.size();
22507 for (unsigned i = 0; i < 16; ++i) {
22508 if (Mask[i / Ratio] == SM_SentinelUndef) {
22509 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22512 int M = Mask[i / Ratio] != SM_SentinelZero
22513 ? Ratio * Mask[i / Ratio] + i % Ratio
22515 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
22517 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
22518 DCI.AddToWorklist(Op.getNode());
22519 SDValue PSHUFBMaskOp =
22520 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
22521 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22522 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
22523 DCI.AddToWorklist(Op.getNode());
22524 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22529 // Failed to find any combines.
22533 /// \brief Fully generic combining of x86 shuffle instructions.
22535 /// This should be the last combine run over the x86 shuffle instructions. Once
22536 /// they have been fully optimized, this will recursively consider all chains
22537 /// of single-use shuffle instructions, build a generic model of the cumulative
22538 /// shuffle operation, and check for simpler instructions which implement this
22539 /// operation. We use this primarily for two purposes:
22541 /// 1) Collapse generic shuffles to specialized single instructions when
22542 /// equivalent. In most cases, this is just an encoding size win, but
22543 /// sometimes we will collapse multiple generic shuffles into a single
22544 /// special-purpose shuffle.
22545 /// 2) Look for sequences of shuffle instructions with 3 or more total
22546 /// instructions, and replace them with the slightly more expensive SSSE3
22547 /// PSHUFB instruction if available. We do this as the last combining step
22548 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22549 /// a suitable short sequence of other instructions. The PHUFB will either
22550 /// use a register or have to read from memory and so is slightly (but only
22551 /// slightly) more expensive than the other shuffle instructions.
22553 /// Because this is inherently a quadratic operation (for each shuffle in
22554 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22555 /// This should never be an issue in practice as the shuffle lowering doesn't
22556 /// produce sequences of more than 8 instructions.
22558 /// FIXME: We will currently miss some cases where the redundant shuffling
22559 /// would simplify under the threshold for PSHUFB formation because of
22560 /// combine-ordering. To fix this, we should do the redundant instruction
22561 /// combining in this recursive walk.
22562 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22563 ArrayRef<int> RootMask,
22564 int Depth, bool HasPSHUFB,
22566 TargetLowering::DAGCombinerInfo &DCI,
22567 const X86Subtarget *Subtarget) {
22568 // Bound the depth of our recursive combine because this is ultimately
22569 // quadratic in nature.
22573 // Directly rip through bitcasts to find the underlying operand.
22574 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22575 Op = Op.getOperand(0);
22577 MVT VT = Op.getSimpleValueType();
22578 if (!VT.isVector())
22579 return false; // Bail if we hit a non-vector.
22580 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
22581 // version should be added.
22582 if (VT.getSizeInBits() != 128)
22585 assert(Root.getSimpleValueType().isVector() &&
22586 "Shuffles operate on vector types!");
22587 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22588 "Can only combine shuffles of the same vector register size.");
22590 if (!isTargetShuffle(Op.getOpcode()))
22592 SmallVector<int, 16> OpMask;
22594 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22595 // We only can combine unary shuffles which we can decode the mask for.
22596 if (!HaveMask || !IsUnary)
22599 assert(VT.getVectorNumElements() == OpMask.size() &&
22600 "Different mask size from vector size!");
22601 assert(((RootMask.size() > OpMask.size() &&
22602 RootMask.size() % OpMask.size() == 0) ||
22603 (OpMask.size() > RootMask.size() &&
22604 OpMask.size() % RootMask.size() == 0) ||
22605 OpMask.size() == RootMask.size()) &&
22606 "The smaller number of elements must divide the larger.");
22607 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22608 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22609 assert(((RootRatio == 1 && OpRatio == 1) ||
22610 (RootRatio == 1) != (OpRatio == 1)) &&
22611 "Must not have a ratio for both incoming and op masks!");
22613 SmallVector<int, 16> Mask;
22614 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22616 // Merge this shuffle operation's mask into our accumulated mask. Note that
22617 // this shuffle's mask will be the first applied to the input, followed by the
22618 // root mask to get us all the way to the root value arrangement. The reason
22619 // for this order is that we are recursing up the operation chain.
22620 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22621 int RootIdx = i / RootRatio;
22622 if (RootMask[RootIdx] < 0) {
22623 // This is a zero or undef lane, we're done.
22624 Mask.push_back(RootMask[RootIdx]);
22628 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22629 int OpIdx = RootMaskedIdx / OpRatio;
22630 if (OpMask[OpIdx] < 0) {
22631 // The incoming lanes are zero or undef, it doesn't matter which ones we
22633 Mask.push_back(OpMask[OpIdx]);
22637 // Ok, we have non-zero lanes, map them through.
22638 Mask.push_back(OpMask[OpIdx] * OpRatio +
22639 RootMaskedIdx % OpRatio);
22642 // See if we can recurse into the operand to combine more things.
22643 switch (Op.getOpcode()) {
22644 case X86ISD::PSHUFB:
22646 case X86ISD::PSHUFD:
22647 case X86ISD::PSHUFHW:
22648 case X86ISD::PSHUFLW:
22649 if (Op.getOperand(0).hasOneUse() &&
22650 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22651 HasPSHUFB, DAG, DCI, Subtarget))
22655 case X86ISD::UNPCKL:
22656 case X86ISD::UNPCKH:
22657 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
22658 // We can't check for single use, we have to check that this shuffle is the only user.
22659 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22660 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22661 HasPSHUFB, DAG, DCI, Subtarget))
22666 // Minor canonicalization of the accumulated shuffle mask to make it easier
22667 // to match below. All this does is detect masks with squential pairs of
22668 // elements, and shrink them to the half-width mask. It does this in a loop
22669 // so it will reduce the size of the mask to the minimal width mask which
22670 // performs an equivalent shuffle.
22671 SmallVector<int, 16> WidenedMask;
22672 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22673 Mask = std::move(WidenedMask);
22674 WidenedMask.clear();
22677 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22681 /// \brief Get the PSHUF-style mask from PSHUF node.
22683 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22684 /// PSHUF-style masks that can be reused with such instructions.
22685 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22686 SmallVector<int, 4> Mask;
22688 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
22692 switch (N.getOpcode()) {
22693 case X86ISD::PSHUFD:
22695 case X86ISD::PSHUFLW:
22698 case X86ISD::PSHUFHW:
22699 Mask.erase(Mask.begin(), Mask.begin() + 4);
22700 for (int &M : Mask)
22704 llvm_unreachable("No valid shuffle instruction found!");
22708 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22710 /// We walk up the chain and look for a combinable shuffle, skipping over
22711 /// shuffles that we could hoist this shuffle's transformation past without
22712 /// altering anything.
22714 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22716 TargetLowering::DAGCombinerInfo &DCI) {
22717 assert(N.getOpcode() == X86ISD::PSHUFD &&
22718 "Called with something other than an x86 128-bit half shuffle!");
22721 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22722 // of the shuffles in the chain so that we can form a fresh chain to replace
22724 SmallVector<SDValue, 8> Chain;
22725 SDValue V = N.getOperand(0);
22726 for (; V.hasOneUse(); V = V.getOperand(0)) {
22727 switch (V.getOpcode()) {
22729 return SDValue(); // Nothing combined!
22732 // Skip bitcasts as we always know the type for the target specific
22736 case X86ISD::PSHUFD:
22737 // Found another dword shuffle.
22740 case X86ISD::PSHUFLW:
22741 // Check that the low words (being shuffled) are the identity in the
22742 // dword shuffle, and the high words are self-contained.
22743 if (Mask[0] != 0 || Mask[1] != 1 ||
22744 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22747 Chain.push_back(V);
22750 case X86ISD::PSHUFHW:
22751 // Check that the high words (being shuffled) are the identity in the
22752 // dword shuffle, and the low words are self-contained.
22753 if (Mask[2] != 2 || Mask[3] != 3 ||
22754 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22757 Chain.push_back(V);
22760 case X86ISD::UNPCKL:
22761 case X86ISD::UNPCKH:
22762 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22763 // shuffle into a preceding word shuffle.
22764 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
22767 // Search for a half-shuffle which we can combine with.
22768 unsigned CombineOp =
22769 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22770 if (V.getOperand(0) != V.getOperand(1) ||
22771 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22773 Chain.push_back(V);
22774 V = V.getOperand(0);
22776 switch (V.getOpcode()) {
22778 return SDValue(); // Nothing to combine.
22780 case X86ISD::PSHUFLW:
22781 case X86ISD::PSHUFHW:
22782 if (V.getOpcode() == CombineOp)
22785 Chain.push_back(V);
22789 V = V.getOperand(0);
22793 } while (V.hasOneUse());
22796 // Break out of the loop if we break out of the switch.
22800 if (!V.hasOneUse())
22801 // We fell out of the loop without finding a viable combining instruction.
22804 // Merge this node's mask and our incoming mask.
22805 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22806 for (int &M : Mask)
22808 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22809 getV4X86ShuffleImm8ForMask(Mask, DAG));
22811 // Rebuild the chain around this new shuffle.
22812 while (!Chain.empty()) {
22813 SDValue W = Chain.pop_back_val();
22815 if (V.getValueType() != W.getOperand(0).getValueType())
22816 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
22818 switch (W.getOpcode()) {
22820 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22822 case X86ISD::UNPCKL:
22823 case X86ISD::UNPCKH:
22824 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22827 case X86ISD::PSHUFD:
22828 case X86ISD::PSHUFLW:
22829 case X86ISD::PSHUFHW:
22830 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22834 if (V.getValueType() != N.getValueType())
22835 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
22837 // Return the new chain to replace N.
22841 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
22843 /// We walk up the chain, skipping shuffles of the other half and looking
22844 /// through shuffles which switch halves trying to find a shuffle of the same
22845 /// pair of dwords.
22846 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22848 TargetLowering::DAGCombinerInfo &DCI) {
22850 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22851 "Called with something other than an x86 128-bit half shuffle!");
22853 unsigned CombineOpcode = N.getOpcode();
22855 // Walk up a single-use chain looking for a combinable shuffle.
22856 SDValue V = N.getOperand(0);
22857 for (; V.hasOneUse(); V = V.getOperand(0)) {
22858 switch (V.getOpcode()) {
22860 return false; // Nothing combined!
22863 // Skip bitcasts as we always know the type for the target specific
22867 case X86ISD::PSHUFLW:
22868 case X86ISD::PSHUFHW:
22869 if (V.getOpcode() == CombineOpcode)
22872 // Other-half shuffles are no-ops.
22875 // Break out of the loop if we break out of the switch.
22879 if (!V.hasOneUse())
22880 // We fell out of the loop without finding a viable combining instruction.
22883 // Combine away the bottom node as its shuffle will be accumulated into
22884 // a preceding shuffle.
22885 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22887 // Record the old value.
22890 // Merge this node's mask and our incoming mask (adjusted to account for all
22891 // the pshufd instructions encountered).
22892 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22893 for (int &M : Mask)
22895 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22896 getV4X86ShuffleImm8ForMask(Mask, DAG));
22898 // Check that the shuffles didn't cancel each other out. If not, we need to
22899 // combine to the new one.
22901 // Replace the combinable shuffle with the combined one, updating all users
22902 // so that we re-evaluate the chain here.
22903 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22908 /// \brief Try to combine x86 target specific shuffles.
22909 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22910 TargetLowering::DAGCombinerInfo &DCI,
22911 const X86Subtarget *Subtarget) {
22913 MVT VT = N.getSimpleValueType();
22914 SmallVector<int, 4> Mask;
22916 switch (N.getOpcode()) {
22917 case X86ISD::PSHUFD:
22918 case X86ISD::PSHUFLW:
22919 case X86ISD::PSHUFHW:
22920 Mask = getPSHUFShuffleMask(N);
22921 assert(Mask.size() == 4);
22927 // Nuke no-op shuffles that show up after combining.
22928 if (isNoopShuffleMask(Mask))
22929 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22931 // Look for simplifications involving one or two shuffle instructions.
22932 SDValue V = N.getOperand(0);
22933 switch (N.getOpcode()) {
22936 case X86ISD::PSHUFLW:
22937 case X86ISD::PSHUFHW:
22938 assert(VT == MVT::v8i16);
22941 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22942 return SDValue(); // We combined away this shuffle, so we're done.
22944 // See if this reduces to a PSHUFD which is no more expensive and can
22945 // combine with more operations. Note that it has to at least flip the
22946 // dwords as otherwise it would have been removed as a no-op.
22947 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
22948 int DMask[] = {0, 1, 2, 3};
22949 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22950 DMask[DOffset + 0] = DOffset + 1;
22951 DMask[DOffset + 1] = DOffset + 0;
22952 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
22953 DCI.AddToWorklist(V.getNode());
22954 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
22955 getV4X86ShuffleImm8ForMask(DMask, DAG));
22956 DCI.AddToWorklist(V.getNode());
22957 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
22960 // Look for shuffle patterns which can be implemented as a single unpack.
22961 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22962 // only works when we have a PSHUFD followed by two half-shuffles.
22963 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22964 (V.getOpcode() == X86ISD::PSHUFLW ||
22965 V.getOpcode() == X86ISD::PSHUFHW) &&
22966 V.getOpcode() != N.getOpcode() &&
22968 SDValue D = V.getOperand(0);
22969 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22970 D = D.getOperand(0);
22971 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22972 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22973 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22974 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22975 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22977 for (int i = 0; i < 4; ++i) {
22978 WordMask[i + NOffset] = Mask[i] + NOffset;
22979 WordMask[i + VOffset] = VMask[i] + VOffset;
22981 // Map the word mask through the DWord mask.
22983 for (int i = 0; i < 8; ++i)
22984 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22985 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
22986 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
22987 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
22988 std::begin(UnpackLoMask)) ||
22989 std::equal(std::begin(MappedMask), std::end(MappedMask),
22990 std::begin(UnpackHiMask))) {
22991 // We can replace all three shuffles with an unpack.
22992 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
22993 DCI.AddToWorklist(V.getNode());
22994 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22996 DL, MVT::v8i16, V, V);
23003 case X86ISD::PSHUFD:
23004 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23013 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23015 /// We combine this directly on the abstract vector shuffle nodes so it is
23016 /// easier to generically match. We also insert dummy vector shuffle nodes for
23017 /// the operands which explicitly discard the lanes which are unused by this
23018 /// operation to try to flow through the rest of the combiner the fact that
23019 /// they're unused.
23020 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23022 EVT VT = N->getValueType(0);
23024 // We only handle target-independent shuffles.
23025 // FIXME: It would be easy and harmless to use the target shuffle mask
23026 // extraction tool to support more.
23027 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23030 auto *SVN = cast<ShuffleVectorSDNode>(N);
23031 ArrayRef<int> Mask = SVN->getMask();
23032 SDValue V1 = N->getOperand(0);
23033 SDValue V2 = N->getOperand(1);
23035 // We require the first shuffle operand to be the SUB node, and the second to
23036 // be the ADD node.
23037 // FIXME: We should support the commuted patterns.
23038 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
23041 // If there are other uses of these operations we can't fold them.
23042 if (!V1->hasOneUse() || !V2->hasOneUse())
23045 // Ensure that both operations have the same operands. Note that we can
23046 // commute the FADD operands.
23047 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23048 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23049 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23052 // We're looking for blends between FADD and FSUB nodes. We insist on these
23053 // nodes being lined up in a specific expected pattern.
23054 if (!(isShuffleEquivalent(V1, V2, Mask, 0, 3) ||
23055 isShuffleEquivalent(V1, V2, Mask, 0, 5, 2, 7) ||
23056 isShuffleEquivalent(V1, V2, Mask, 0, 9, 2, 11, 4, 13, 6, 15)))
23059 // Only specific types are legal at this point, assert so we notice if and
23060 // when these change.
23061 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23062 VT == MVT::v4f64) &&
23063 "Unknown vector type encountered!");
23065 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23068 /// PerformShuffleCombine - Performs several different shuffle combines.
23069 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23070 TargetLowering::DAGCombinerInfo &DCI,
23071 const X86Subtarget *Subtarget) {
23073 SDValue N0 = N->getOperand(0);
23074 SDValue N1 = N->getOperand(1);
23075 EVT VT = N->getValueType(0);
23077 // Don't create instructions with illegal types after legalize types has run.
23078 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23079 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23082 // If we have legalized the vector types, look for blends of FADD and FSUB
23083 // nodes that we can fuse into an ADDSUB node.
23084 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23085 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23088 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23089 if (Subtarget->hasFp256() && VT.is256BitVector() &&
23090 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23091 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23093 // During Type Legalization, when promoting illegal vector types,
23094 // the backend might introduce new shuffle dag nodes and bitcasts.
23096 // This code performs the following transformation:
23097 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23098 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23100 // We do this only if both the bitcast and the BINOP dag nodes have
23101 // one use. Also, perform this transformation only if the new binary
23102 // operation is legal. This is to avoid introducing dag nodes that
23103 // potentially need to be further expanded (or custom lowered) into a
23104 // less optimal sequence of dag nodes.
23105 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23106 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23107 N0.getOpcode() == ISD::BITCAST) {
23108 SDValue BC0 = N0.getOperand(0);
23109 EVT SVT = BC0.getValueType();
23110 unsigned Opcode = BC0.getOpcode();
23111 unsigned NumElts = VT.getVectorNumElements();
23113 if (BC0.hasOneUse() && SVT.isVector() &&
23114 SVT.getVectorNumElements() * 2 == NumElts &&
23115 TLI.isOperationLegal(Opcode, VT)) {
23116 bool CanFold = false;
23128 unsigned SVTNumElts = SVT.getVectorNumElements();
23129 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23130 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23131 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23132 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23133 CanFold = SVOp->getMaskElt(i) < 0;
23136 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
23137 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
23138 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23139 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23144 // Only handle 128 wide vector from here on.
23145 if (!VT.is128BitVector())
23148 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23149 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23150 // consecutive, non-overlapping, and in the right order.
23151 SmallVector<SDValue, 16> Elts;
23152 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23153 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23155 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
23159 if (isTargetShuffle(N->getOpcode())) {
23161 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23162 if (Shuffle.getNode())
23165 // Try recursively combining arbitrary sequences of x86 shuffle
23166 // instructions into higher-order shuffles. We do this after combining
23167 // specific PSHUF instruction sequences into their minimal form so that we
23168 // can evaluate how many specialized shuffle instructions are involved in
23169 // a particular chain.
23170 SmallVector<int, 1> NonceMask; // Just a placeholder.
23171 NonceMask.push_back(0);
23172 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23173 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23175 return SDValue(); // This routine will use CombineTo to replace N.
23181 /// PerformTruncateCombine - Converts truncate operation to
23182 /// a sequence of vector shuffle operations.
23183 /// It is possible when we truncate 256-bit vector to 128-bit vector
23184 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
23185 TargetLowering::DAGCombinerInfo &DCI,
23186 const X86Subtarget *Subtarget) {
23190 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23191 /// specific shuffle of a load can be folded into a single element load.
23192 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23193 /// shuffles have been custom lowered so we need to handle those here.
23194 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23195 TargetLowering::DAGCombinerInfo &DCI) {
23196 if (DCI.isBeforeLegalizeOps())
23199 SDValue InVec = N->getOperand(0);
23200 SDValue EltNo = N->getOperand(1);
23202 if (!isa<ConstantSDNode>(EltNo))
23205 EVT OriginalVT = InVec.getValueType();
23207 if (InVec.getOpcode() == ISD::BITCAST) {
23208 // Don't duplicate a load with other uses.
23209 if (!InVec.hasOneUse())
23211 EVT BCVT = InVec.getOperand(0).getValueType();
23212 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23214 InVec = InVec.getOperand(0);
23217 EVT CurrentVT = InVec.getValueType();
23219 if (!isTargetShuffle(InVec.getOpcode()))
23222 // Don't duplicate a load with other uses.
23223 if (!InVec.hasOneUse())
23226 SmallVector<int, 16> ShuffleMask;
23228 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23229 ShuffleMask, UnaryShuffle))
23232 // Select the input vector, guarding against out of range extract vector.
23233 unsigned NumElems = CurrentVT.getVectorNumElements();
23234 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23235 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23236 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23237 : InVec.getOperand(1);
23239 // If inputs to shuffle are the same for both ops, then allow 2 uses
23240 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23241 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23243 if (LdNode.getOpcode() == ISD::BITCAST) {
23244 // Don't duplicate a load with other uses.
23245 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23248 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23249 LdNode = LdNode.getOperand(0);
23252 if (!ISD::isNormalLoad(LdNode.getNode()))
23255 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23257 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23260 EVT EltVT = N->getValueType(0);
23261 // If there's a bitcast before the shuffle, check if the load type and
23262 // alignment is valid.
23263 unsigned Align = LN0->getAlignment();
23264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23265 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
23266 EltVT.getTypeForEVT(*DAG.getContext()));
23268 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23271 // All checks match so transform back to vector_shuffle so that DAG combiner
23272 // can finish the job
23275 // Create shuffle node taking into account the case that its a unary shuffle
23276 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23277 : InVec.getOperand(1);
23278 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23279 InVec.getOperand(0), Shuffle,
23281 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
23282 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23286 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23287 /// special and don't usually play with other vector types, it's better to
23288 /// handle them early to be sure we emit efficient code by avoiding
23289 /// store-load conversions.
23290 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
23291 if (N->getValueType(0) != MVT::x86mmx ||
23292 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
23293 N->getOperand(0)->getValueType(0) != MVT::v2i32)
23296 SDValue V = N->getOperand(0);
23297 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
23298 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
23299 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
23300 N->getValueType(0), V.getOperand(0));
23305 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23306 /// generation and convert it from being a bunch of shuffles and extracts
23307 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23308 /// storing the value and loading scalars back, while for x64 we should
23309 /// use 64-bit extracts and shifts.
23310 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23311 TargetLowering::DAGCombinerInfo &DCI) {
23312 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
23313 if (NewOp.getNode())
23316 SDValue InputVector = N->getOperand(0);
23318 // Detect mmx to i32 conversion through a v2i32 elt extract.
23319 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23320 N->getValueType(0) == MVT::i32 &&
23321 InputVector.getValueType() == MVT::v2i32) {
23323 // The bitcast source is a direct mmx result.
23324 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23325 if (MMXSrc.getValueType() == MVT::x86mmx)
23326 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23327 N->getValueType(0),
23328 InputVector.getNode()->getOperand(0));
23330 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23331 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23332 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23333 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
23334 MMXSrcOp.getOpcode() == ISD::BITCAST &&
23335 MMXSrcOp.getValueType() == MVT::v1i64 &&
23336 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23337 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23338 N->getValueType(0),
23339 MMXSrcOp.getOperand(0));
23342 // Only operate on vectors of 4 elements, where the alternative shuffling
23343 // gets to be more expensive.
23344 if (InputVector.getValueType() != MVT::v4i32)
23347 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23348 // single use which is a sign-extend or zero-extend, and all elements are
23350 SmallVector<SDNode *, 4> Uses;
23351 unsigned ExtractedElements = 0;
23352 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23353 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23354 if (UI.getUse().getResNo() != InputVector.getResNo())
23357 SDNode *Extract = *UI;
23358 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23361 if (Extract->getValueType(0) != MVT::i32)
23363 if (!Extract->hasOneUse())
23365 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23366 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23368 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23371 // Record which element was extracted.
23372 ExtractedElements |=
23373 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23375 Uses.push_back(Extract);
23378 // If not all the elements were used, this may not be worthwhile.
23379 if (ExtractedElements != 15)
23382 // Ok, we've now decided to do the transformation.
23383 // If 64-bit shifts are legal, use the extract-shift sequence,
23384 // otherwise bounce the vector off the cache.
23385 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23387 SDLoc dl(InputVector);
23389 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23390 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
23391 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
23392 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23393 DAG.getConstant(0, VecIdxTy));
23394 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23395 DAG.getConstant(1, VecIdxTy));
23397 SDValue ShAmt = DAG.getConstant(32,
23398 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
23399 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23400 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23401 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23402 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23403 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23404 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23406 // Store the value to a temporary stack slot.
23407 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23408 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23409 MachinePointerInfo(), false, false, 0);
23411 EVT ElementType = InputVector.getValueType().getVectorElementType();
23412 unsigned EltSize = ElementType.getSizeInBits() / 8;
23414 // Replace each use (extract) with a load of the appropriate element.
23415 for (unsigned i = 0; i < 4; ++i) {
23416 uint64_t Offset = EltSize * i;
23417 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
23419 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
23420 StackPtr, OffsetVal);
23422 // Load the scalar.
23423 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23424 ScalarAddr, MachinePointerInfo(),
23425 false, false, false, 0);
23430 // Replace the extracts
23431 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23432 UE = Uses.end(); UI != UE; ++UI) {
23433 SDNode *Extract = *UI;
23435 SDValue Idx = Extract->getOperand(1);
23436 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23437 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23440 // The replacement was made in place; don't return anything.
23444 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
23445 static std::pair<unsigned, bool>
23446 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
23447 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
23448 if (!VT.isVector())
23449 return std::make_pair(0, false);
23451 bool NeedSplit = false;
23452 switch (VT.getSimpleVT().SimpleTy) {
23453 default: return std::make_pair(0, false);
23456 if (!Subtarget->hasVLX())
23457 return std::make_pair(0, false);
23461 if (!Subtarget->hasBWI())
23462 return std::make_pair(0, false);
23466 if (!Subtarget->hasAVX512())
23467 return std::make_pair(0, false);
23472 if (!Subtarget->hasAVX2())
23474 if (!Subtarget->hasAVX())
23475 return std::make_pair(0, false);
23480 if (!Subtarget->hasSSE2())
23481 return std::make_pair(0, false);
23484 // SSE2 has only a small subset of the operations.
23485 bool hasUnsigned = Subtarget->hasSSE41() ||
23486 (Subtarget->hasSSE2() && VT == MVT::v16i8);
23487 bool hasSigned = Subtarget->hasSSE41() ||
23488 (Subtarget->hasSSE2() && VT == MVT::v8i16);
23490 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23493 // Check for x CC y ? x : y.
23494 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23495 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23500 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
23503 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
23506 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23509 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23511 // Check for x CC y ? y : x -- a min/max with reversed arms.
23512 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23513 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23518 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
23521 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
23524 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23527 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23531 return std::make_pair(Opc, NeedSplit);
23535 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23536 const X86Subtarget *Subtarget) {
23538 SDValue Cond = N->getOperand(0);
23539 SDValue LHS = N->getOperand(1);
23540 SDValue RHS = N->getOperand(2);
23542 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23543 SDValue CondSrc = Cond->getOperand(0);
23544 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23545 Cond = CondSrc->getOperand(0);
23548 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23551 // A vselect where all conditions and data are constants can be optimized into
23552 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23553 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23554 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23557 unsigned MaskValue = 0;
23558 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23561 MVT VT = N->getSimpleValueType(0);
23562 unsigned NumElems = VT.getVectorNumElements();
23563 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23564 for (unsigned i = 0; i < NumElems; ++i) {
23565 // Be sure we emit undef where we can.
23566 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23567 ShuffleMask[i] = -1;
23569 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23572 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23573 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23575 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23578 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23580 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23581 TargetLowering::DAGCombinerInfo &DCI,
23582 const X86Subtarget *Subtarget) {
23584 SDValue Cond = N->getOperand(0);
23585 // Get the LHS/RHS of the select.
23586 SDValue LHS = N->getOperand(1);
23587 SDValue RHS = N->getOperand(2);
23588 EVT VT = LHS.getValueType();
23589 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23591 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23592 // instructions match the semantics of the common C idiom x<y?x:y but not
23593 // x<=y?x:y, because of how they handle negative zero (which can be
23594 // ignored in unsafe-math mode).
23595 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23596 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23597 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23598 (Subtarget->hasSSE2() ||
23599 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23600 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23602 unsigned Opcode = 0;
23603 // Check for x CC y ? x : y.
23604 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23605 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23609 // Converting this to a min would handle NaNs incorrectly, and swapping
23610 // the operands would cause it to handle comparisons between positive
23611 // and negative zero incorrectly.
23612 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23613 if (!DAG.getTarget().Options.UnsafeFPMath &&
23614 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23616 std::swap(LHS, RHS);
23618 Opcode = X86ISD::FMIN;
23621 // Converting this to a min would handle comparisons between positive
23622 // and negative zero incorrectly.
23623 if (!DAG.getTarget().Options.UnsafeFPMath &&
23624 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23626 Opcode = X86ISD::FMIN;
23629 // Converting this to a min would handle both negative zeros and NaNs
23630 // incorrectly, but we can swap the operands to fix both.
23631 std::swap(LHS, RHS);
23635 Opcode = X86ISD::FMIN;
23639 // Converting this to a max would handle comparisons between positive
23640 // and negative zero incorrectly.
23641 if (!DAG.getTarget().Options.UnsafeFPMath &&
23642 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23644 Opcode = X86ISD::FMAX;
23647 // Converting this to a max would handle NaNs incorrectly, and swapping
23648 // the operands would cause it to handle comparisons between positive
23649 // and negative zero incorrectly.
23650 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23651 if (!DAG.getTarget().Options.UnsafeFPMath &&
23652 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23654 std::swap(LHS, RHS);
23656 Opcode = X86ISD::FMAX;
23659 // Converting this to a max would handle both negative zeros and NaNs
23660 // incorrectly, but we can swap the operands to fix both.
23661 std::swap(LHS, RHS);
23665 Opcode = X86ISD::FMAX;
23668 // Check for x CC y ? y : x -- a min/max with reversed arms.
23669 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23670 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23674 // Converting this to a min would handle comparisons between positive
23675 // and negative zero incorrectly, and swapping the operands would
23676 // cause it to handle NaNs incorrectly.
23677 if (!DAG.getTarget().Options.UnsafeFPMath &&
23678 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23679 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23681 std::swap(LHS, RHS);
23683 Opcode = X86ISD::FMIN;
23686 // Converting this to a min would handle NaNs incorrectly.
23687 if (!DAG.getTarget().Options.UnsafeFPMath &&
23688 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23690 Opcode = X86ISD::FMIN;
23693 // Converting this to a min would handle both negative zeros and NaNs
23694 // incorrectly, but we can swap the operands to fix both.
23695 std::swap(LHS, RHS);
23699 Opcode = X86ISD::FMIN;
23703 // Converting this to a max would handle NaNs incorrectly.
23704 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23706 Opcode = X86ISD::FMAX;
23709 // Converting this to a max would handle comparisons between positive
23710 // and negative zero incorrectly, and swapping the operands would
23711 // cause it to handle NaNs incorrectly.
23712 if (!DAG.getTarget().Options.UnsafeFPMath &&
23713 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23714 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23716 std::swap(LHS, RHS);
23718 Opcode = X86ISD::FMAX;
23721 // Converting this to a max would handle both negative zeros and NaNs
23722 // incorrectly, but we can swap the operands to fix both.
23723 std::swap(LHS, RHS);
23727 Opcode = X86ISD::FMAX;
23733 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23736 EVT CondVT = Cond.getValueType();
23737 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23738 CondVT.getVectorElementType() == MVT::i1) {
23739 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23740 // lowering on KNL. In this case we convert it to
23741 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23742 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23743 // Since SKX these selects have a proper lowering.
23744 EVT OpVT = LHS.getValueType();
23745 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23746 (OpVT.getVectorElementType() == MVT::i8 ||
23747 OpVT.getVectorElementType() == MVT::i16) &&
23748 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23749 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23750 DCI.AddToWorklist(Cond.getNode());
23751 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23754 // If this is a select between two integer constants, try to do some
23756 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23757 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23758 // Don't do this for crazy integer types.
23759 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23760 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23761 // so that TrueC (the true value) is larger than FalseC.
23762 bool NeedsCondInvert = false;
23764 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23765 // Efficiently invertible.
23766 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23767 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23768 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23769 NeedsCondInvert = true;
23770 std::swap(TrueC, FalseC);
23773 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23774 if (FalseC->getAPIntValue() == 0 &&
23775 TrueC->getAPIntValue().isPowerOf2()) {
23776 if (NeedsCondInvert) // Invert the condition if needed.
23777 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23778 DAG.getConstant(1, Cond.getValueType()));
23780 // Zero extend the condition if needed.
23781 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23783 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23784 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23785 DAG.getConstant(ShAmt, MVT::i8));
23788 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23789 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23790 if (NeedsCondInvert) // Invert the condition if needed.
23791 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23792 DAG.getConstant(1, Cond.getValueType()));
23794 // Zero extend the condition if needed.
23795 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23796 FalseC->getValueType(0), Cond);
23797 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23798 SDValue(FalseC, 0));
23801 // Optimize cases that will turn into an LEA instruction. This requires
23802 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23803 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23804 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23805 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23807 bool isFastMultiplier = false;
23809 switch ((unsigned char)Diff) {
23811 case 1: // result = add base, cond
23812 case 2: // result = lea base( , cond*2)
23813 case 3: // result = lea base(cond, cond*2)
23814 case 4: // result = lea base( , cond*4)
23815 case 5: // result = lea base(cond, cond*4)
23816 case 8: // result = lea base( , cond*8)
23817 case 9: // result = lea base(cond, cond*8)
23818 isFastMultiplier = true;
23823 if (isFastMultiplier) {
23824 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23825 if (NeedsCondInvert) // Invert the condition if needed.
23826 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23827 DAG.getConstant(1, Cond.getValueType()));
23829 // Zero extend the condition if needed.
23830 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23832 // Scale the condition by the difference.
23834 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23835 DAG.getConstant(Diff, Cond.getValueType()));
23837 // Add the base if non-zero.
23838 if (FalseC->getAPIntValue() != 0)
23839 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23840 SDValue(FalseC, 0));
23847 // Canonicalize max and min:
23848 // (x > y) ? x : y -> (x >= y) ? x : y
23849 // (x < y) ? x : y -> (x <= y) ? x : y
23850 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23851 // the need for an extra compare
23852 // against zero. e.g.
23853 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23855 // testl %edi, %edi
23857 // cmovgl %edi, %eax
23861 // cmovsl %eax, %edi
23862 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23863 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23864 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23865 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23870 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23871 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23872 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23873 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23878 // Early exit check
23879 if (!TLI.isTypeLegal(VT))
23882 // Match VSELECTs into subs with unsigned saturation.
23883 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23884 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23885 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23886 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23887 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23889 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23890 // left side invert the predicate to simplify logic below.
23892 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23894 CC = ISD::getSetCCInverse(CC, true);
23895 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23899 if (Other.getNode() && Other->getNumOperands() == 2 &&
23900 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23901 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23902 SDValue CondRHS = Cond->getOperand(1);
23904 // Look for a general sub with unsigned saturation first.
23905 // x >= y ? x-y : 0 --> subus x, y
23906 // x > y ? x-y : 0 --> subus x, y
23907 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23908 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23909 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23911 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23912 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23913 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23914 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23915 // If the RHS is a constant we have to reverse the const
23916 // canonicalization.
23917 // x > C-1 ? x+-C : 0 --> subus x, C
23918 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23919 CondRHSConst->getAPIntValue() ==
23920 (-OpRHSConst->getAPIntValue() - 1))
23921 return DAG.getNode(
23922 X86ISD::SUBUS, DL, VT, OpLHS,
23923 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
23925 // Another special case: If C was a sign bit, the sub has been
23926 // canonicalized into a xor.
23927 // FIXME: Would it be better to use computeKnownBits to determine
23928 // whether it's safe to decanonicalize the xor?
23929 // x s< 0 ? x^C : 0 --> subus x, C
23930 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23931 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23932 OpRHSConst->getAPIntValue().isSignBit())
23933 // Note that we have to rebuild the RHS constant here to ensure we
23934 // don't rely on particular values of undef lanes.
23935 return DAG.getNode(
23936 X86ISD::SUBUS, DL, VT, OpLHS,
23937 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
23942 // Try to match a min/max vector operation.
23943 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
23944 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
23945 unsigned Opc = ret.first;
23946 bool NeedSplit = ret.second;
23948 if (Opc && NeedSplit) {
23949 unsigned NumElems = VT.getVectorNumElements();
23950 // Extract the LHS vectors
23951 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
23952 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
23954 // Extract the RHS vectors
23955 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
23956 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
23958 // Create min/max for each subvector
23959 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
23960 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
23962 // Merge the result
23963 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
23965 return DAG.getNode(Opc, DL, VT, LHS, RHS);
23968 // Simplify vector selection if condition value type matches vselect
23970 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23971 assert(Cond.getValueType().isVector() &&
23972 "vector select expects a vector selector!");
23974 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23975 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23977 // Try invert the condition if true value is not all 1s and false value
23979 if (!TValIsAllOnes && !FValIsAllZeros &&
23980 // Check if the selector will be produced by CMPP*/PCMP*
23981 Cond.getOpcode() == ISD::SETCC &&
23982 // Check if SETCC has already been promoted
23983 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
23984 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23985 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23987 if (TValIsAllZeros || FValIsAllOnes) {
23988 SDValue CC = Cond.getOperand(2);
23989 ISD::CondCode NewCC =
23990 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23991 Cond.getOperand(0).getValueType().isInteger());
23992 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23993 std::swap(LHS, RHS);
23994 TValIsAllOnes = FValIsAllOnes;
23995 FValIsAllZeros = TValIsAllZeros;
23999 if (TValIsAllOnes || FValIsAllZeros) {
24002 if (TValIsAllOnes && FValIsAllZeros)
24004 else if (TValIsAllOnes)
24005 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
24006 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
24007 else if (FValIsAllZeros)
24008 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24009 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
24011 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
24015 // If we know that this node is legal then we know that it is going to be
24016 // matched by one of the SSE/AVX BLEND instructions. These instructions only
24017 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
24018 // to simplify previous instructions.
24019 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24020 !DCI.isBeforeLegalize() &&
24021 // We explicitly check against SSE4.1, v8i16 and v16i16 because, although
24022 // vselect nodes may be marked as Custom, they might only be legal when
24023 // Cond is a build_vector of constants. This will be taken care in
24024 // a later condition.
24025 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) &&
24026 Subtarget->hasSSE41() && VT != MVT::v16i16 && VT != MVT::v8i16) &&
24027 // Don't optimize vector of constants. Those are handled by
24028 // the generic code and all the bits must be properly set for
24029 // the generic optimizer.
24030 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24031 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
24033 // Don't optimize vector selects that map to mask-registers.
24037 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24038 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24040 APInt KnownZero, KnownOne;
24041 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24042 DCI.isBeforeLegalizeOps());
24043 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24044 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24046 // If we changed the computation somewhere in the DAG, this change
24047 // will affect all users of Cond.
24048 // Make sure it is fine and update all the nodes so that we do not
24049 // use the generic VSELECT anymore. Otherwise, we may perform
24050 // wrong optimizations as we messed up with the actual expectation
24051 // for the vector boolean values.
24052 if (Cond != TLO.Old) {
24053 // Check all uses of that condition operand to check whether it will be
24054 // consumed by non-BLEND instructions, which may depend on all bits are
24056 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24058 if (I->getOpcode() != ISD::VSELECT)
24059 // TODO: Add other opcodes eventually lowered into BLEND.
24062 // Update all the users of the condition, before committing the change,
24063 // so that the VSELECT optimizations that expect the correct vector
24064 // boolean value will not be triggered.
24065 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24067 DAG.ReplaceAllUsesOfValueWith(
24069 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24070 Cond, I->getOperand(1), I->getOperand(2)));
24071 DCI.CommitTargetLoweringOpt(TLO);
24074 // At this point, only Cond is changed. Change the condition
24075 // just for N to keep the opportunity to optimize all other
24076 // users their own way.
24077 DAG.ReplaceAllUsesOfValueWith(
24079 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24080 TLO.New, N->getOperand(1), N->getOperand(2)));
24085 // We should generate an X86ISD::BLENDI from a vselect if its argument
24086 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24087 // constants. This specific pattern gets generated when we split a
24088 // selector for a 512 bit vector in a machine without AVX512 (but with
24089 // 256-bit vectors), during legalization:
24091 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24093 // Iff we find this pattern and the build_vectors are built from
24094 // constants, we translate the vselect into a shuffle_vector that we
24095 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24096 if ((N->getOpcode() == ISD::VSELECT ||
24097 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24098 !DCI.isBeforeLegalize()) {
24099 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24100 if (Shuffle.getNode())
24107 // Check whether a boolean test is testing a boolean value generated by
24108 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24111 // Simplify the following patterns:
24112 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24113 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24114 // to (Op EFLAGS Cond)
24116 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24117 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24118 // to (Op EFLAGS !Cond)
24120 // where Op could be BRCOND or CMOV.
24122 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24123 // Quit if not CMP and SUB with its value result used.
24124 if (Cmp.getOpcode() != X86ISD::CMP &&
24125 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24128 // Quit if not used as a boolean value.
24129 if (CC != X86::COND_E && CC != X86::COND_NE)
24132 // Check CMP operands. One of them should be 0 or 1 and the other should be
24133 // an SetCC or extended from it.
24134 SDValue Op1 = Cmp.getOperand(0);
24135 SDValue Op2 = Cmp.getOperand(1);
24138 const ConstantSDNode* C = nullptr;
24139 bool needOppositeCond = (CC == X86::COND_E);
24140 bool checkAgainstTrue = false; // Is it a comparison against 1?
24142 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24144 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24146 else // Quit if all operands are not constants.
24149 if (C->getZExtValue() == 1) {
24150 needOppositeCond = !needOppositeCond;
24151 checkAgainstTrue = true;
24152 } else if (C->getZExtValue() != 0)
24153 // Quit if the constant is neither 0 or 1.
24156 bool truncatedToBoolWithAnd = false;
24157 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24158 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24159 SetCC.getOpcode() == ISD::TRUNCATE ||
24160 SetCC.getOpcode() == ISD::AND) {
24161 if (SetCC.getOpcode() == ISD::AND) {
24163 ConstantSDNode *CS;
24164 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
24165 CS->getZExtValue() == 1)
24167 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
24168 CS->getZExtValue() == 1)
24172 SetCC = SetCC.getOperand(OpIdx);
24173 truncatedToBoolWithAnd = true;
24175 SetCC = SetCC.getOperand(0);
24178 switch (SetCC.getOpcode()) {
24179 case X86ISD::SETCC_CARRY:
24180 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24181 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24182 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24183 // truncated to i1 using 'and'.
24184 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24186 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24187 "Invalid use of SETCC_CARRY!");
24189 case X86ISD::SETCC:
24190 // Set the condition code or opposite one if necessary.
24191 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24192 if (needOppositeCond)
24193 CC = X86::GetOppositeBranchCondition(CC);
24194 return SetCC.getOperand(1);
24195 case X86ISD::CMOV: {
24196 // Check whether false/true value has canonical one, i.e. 0 or 1.
24197 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24198 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24199 // Quit if true value is not a constant.
24202 // Quit if false value is not a constant.
24204 SDValue Op = SetCC.getOperand(0);
24205 // Skip 'zext' or 'trunc' node.
24206 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24207 Op.getOpcode() == ISD::TRUNCATE)
24208 Op = Op.getOperand(0);
24209 // A special case for rdrand/rdseed, where 0 is set if false cond is
24211 if ((Op.getOpcode() != X86ISD::RDRAND &&
24212 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24215 // Quit if false value is not the constant 0 or 1.
24216 bool FValIsFalse = true;
24217 if (FVal && FVal->getZExtValue() != 0) {
24218 if (FVal->getZExtValue() != 1)
24220 // If FVal is 1, opposite cond is needed.
24221 needOppositeCond = !needOppositeCond;
24222 FValIsFalse = false;
24224 // Quit if TVal is not the constant opposite of FVal.
24225 if (FValIsFalse && TVal->getZExtValue() != 1)
24227 if (!FValIsFalse && TVal->getZExtValue() != 0)
24229 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24230 if (needOppositeCond)
24231 CC = X86::GetOppositeBranchCondition(CC);
24232 return SetCC.getOperand(3);
24239 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24240 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24241 TargetLowering::DAGCombinerInfo &DCI,
24242 const X86Subtarget *Subtarget) {
24245 // If the flag operand isn't dead, don't touch this CMOV.
24246 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24249 SDValue FalseOp = N->getOperand(0);
24250 SDValue TrueOp = N->getOperand(1);
24251 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24252 SDValue Cond = N->getOperand(3);
24254 if (CC == X86::COND_E || CC == X86::COND_NE) {
24255 switch (Cond.getOpcode()) {
24259 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24260 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24261 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24267 Flags = checkBoolTestSetCCCombine(Cond, CC);
24268 if (Flags.getNode() &&
24269 // Extra check as FCMOV only supports a subset of X86 cond.
24270 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24271 SDValue Ops[] = { FalseOp, TrueOp,
24272 DAG.getConstant(CC, MVT::i8), Flags };
24273 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24276 // If this is a select between two integer constants, try to do some
24277 // optimizations. Note that the operands are ordered the opposite of SELECT
24279 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24280 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24281 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24282 // larger than FalseC (the false value).
24283 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24284 CC = X86::GetOppositeBranchCondition(CC);
24285 std::swap(TrueC, FalseC);
24286 std::swap(TrueOp, FalseOp);
24289 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24290 // This is efficient for any integer data type (including i8/i16) and
24292 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24293 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24294 DAG.getConstant(CC, MVT::i8), Cond);
24296 // Zero extend the condition if needed.
24297 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24299 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24300 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24301 DAG.getConstant(ShAmt, MVT::i8));
24302 if (N->getNumValues() == 2) // Dead flag value?
24303 return DCI.CombineTo(N, Cond, SDValue());
24307 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24308 // for any integer data type, including i8/i16.
24309 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24310 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24311 DAG.getConstant(CC, MVT::i8), Cond);
24313 // Zero extend the condition if needed.
24314 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24315 FalseC->getValueType(0), Cond);
24316 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24317 SDValue(FalseC, 0));
24319 if (N->getNumValues() == 2) // Dead flag value?
24320 return DCI.CombineTo(N, Cond, SDValue());
24324 // Optimize cases that will turn into an LEA instruction. This requires
24325 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24326 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24327 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24328 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24330 bool isFastMultiplier = false;
24332 switch ((unsigned char)Diff) {
24334 case 1: // result = add base, cond
24335 case 2: // result = lea base( , cond*2)
24336 case 3: // result = lea base(cond, cond*2)
24337 case 4: // result = lea base( , cond*4)
24338 case 5: // result = lea base(cond, cond*4)
24339 case 8: // result = lea base( , cond*8)
24340 case 9: // result = lea base(cond, cond*8)
24341 isFastMultiplier = true;
24346 if (isFastMultiplier) {
24347 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24348 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24349 DAG.getConstant(CC, MVT::i8), Cond);
24350 // Zero extend the condition if needed.
24351 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24353 // Scale the condition by the difference.
24355 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24356 DAG.getConstant(Diff, Cond.getValueType()));
24358 // Add the base if non-zero.
24359 if (FalseC->getAPIntValue() != 0)
24360 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24361 SDValue(FalseC, 0));
24362 if (N->getNumValues() == 2) // Dead flag value?
24363 return DCI.CombineTo(N, Cond, SDValue());
24370 // Handle these cases:
24371 // (select (x != c), e, c) -> select (x != c), e, x),
24372 // (select (x == c), c, e) -> select (x == c), x, e)
24373 // where the c is an integer constant, and the "select" is the combination
24374 // of CMOV and CMP.
24376 // The rationale for this change is that the conditional-move from a constant
24377 // needs two instructions, however, conditional-move from a register needs
24378 // only one instruction.
24380 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24381 // some instruction-combining opportunities. This opt needs to be
24382 // postponed as late as possible.
24384 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24385 // the DCI.xxxx conditions are provided to postpone the optimization as
24386 // late as possible.
24388 ConstantSDNode *CmpAgainst = nullptr;
24389 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24390 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24391 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24393 if (CC == X86::COND_NE &&
24394 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24395 CC = X86::GetOppositeBranchCondition(CC);
24396 std::swap(TrueOp, FalseOp);
24399 if (CC == X86::COND_E &&
24400 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24401 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24402 DAG.getConstant(CC, MVT::i8), Cond };
24403 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24411 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
24412 const X86Subtarget *Subtarget) {
24413 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
24415 default: return SDValue();
24416 // SSE/AVX/AVX2 blend intrinsics.
24417 case Intrinsic::x86_avx2_pblendvb:
24418 case Intrinsic::x86_avx2_pblendw:
24419 case Intrinsic::x86_avx2_pblendd_128:
24420 case Intrinsic::x86_avx2_pblendd_256:
24421 // Don't try to simplify this intrinsic if we don't have AVX2.
24422 if (!Subtarget->hasAVX2())
24425 case Intrinsic::x86_avx_blend_pd_256:
24426 case Intrinsic::x86_avx_blend_ps_256:
24427 case Intrinsic::x86_avx_blendv_pd_256:
24428 case Intrinsic::x86_avx_blendv_ps_256:
24429 // Don't try to simplify this intrinsic if we don't have AVX.
24430 if (!Subtarget->hasAVX())
24433 case Intrinsic::x86_sse41_pblendw:
24434 case Intrinsic::x86_sse41_blendpd:
24435 case Intrinsic::x86_sse41_blendps:
24436 case Intrinsic::x86_sse41_blendvps:
24437 case Intrinsic::x86_sse41_blendvpd:
24438 case Intrinsic::x86_sse41_pblendvb: {
24439 SDValue Op0 = N->getOperand(1);
24440 SDValue Op1 = N->getOperand(2);
24441 SDValue Mask = N->getOperand(3);
24443 // Don't try to simplify this intrinsic if we don't have SSE4.1.
24444 if (!Subtarget->hasSSE41())
24447 // fold (blend A, A, Mask) -> A
24450 // fold (blend A, B, allZeros) -> A
24451 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
24453 // fold (blend A, B, allOnes) -> B
24454 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
24457 // Simplify the case where the mask is a constant i32 value.
24458 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
24459 if (C->isNullValue())
24461 if (C->isAllOnesValue())
24468 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
24469 case Intrinsic::x86_sse2_psrai_w:
24470 case Intrinsic::x86_sse2_psrai_d:
24471 case Intrinsic::x86_avx2_psrai_w:
24472 case Intrinsic::x86_avx2_psrai_d:
24473 case Intrinsic::x86_sse2_psra_w:
24474 case Intrinsic::x86_sse2_psra_d:
24475 case Intrinsic::x86_avx2_psra_w:
24476 case Intrinsic::x86_avx2_psra_d: {
24477 SDValue Op0 = N->getOperand(1);
24478 SDValue Op1 = N->getOperand(2);
24479 EVT VT = Op0.getValueType();
24480 assert(VT.isVector() && "Expected a vector type!");
24482 if (isa<BuildVectorSDNode>(Op1))
24483 Op1 = Op1.getOperand(0);
24485 if (!isa<ConstantSDNode>(Op1))
24488 EVT SVT = VT.getVectorElementType();
24489 unsigned SVTBits = SVT.getSizeInBits();
24491 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
24492 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
24493 uint64_t ShAmt = C.getZExtValue();
24495 // Don't try to convert this shift into a ISD::SRA if the shift
24496 // count is bigger than or equal to the element size.
24497 if (ShAmt >= SVTBits)
24500 // Trivial case: if the shift count is zero, then fold this
24501 // into the first operand.
24505 // Replace this packed shift intrinsic with a target independent
24507 SDValue Splat = DAG.getConstant(C, VT);
24508 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
24513 /// PerformMulCombine - Optimize a single multiply with constant into two
24514 /// in order to implement it with two cheaper instructions, e.g.
24515 /// LEA + SHL, LEA + LEA.
24516 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24517 TargetLowering::DAGCombinerInfo &DCI) {
24518 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24521 EVT VT = N->getValueType(0);
24522 if (VT != MVT::i64 && VT != MVT::i32)
24525 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24528 uint64_t MulAmt = C->getZExtValue();
24529 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24532 uint64_t MulAmt1 = 0;
24533 uint64_t MulAmt2 = 0;
24534 if ((MulAmt % 9) == 0) {
24536 MulAmt2 = MulAmt / 9;
24537 } else if ((MulAmt % 5) == 0) {
24539 MulAmt2 = MulAmt / 5;
24540 } else if ((MulAmt % 3) == 0) {
24542 MulAmt2 = MulAmt / 3;
24545 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24548 if (isPowerOf2_64(MulAmt2) &&
24549 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24550 // If second multiplifer is pow2, issue it first. We want the multiply by
24551 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24553 std::swap(MulAmt1, MulAmt2);
24556 if (isPowerOf2_64(MulAmt1))
24557 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24558 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
24560 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24561 DAG.getConstant(MulAmt1, VT));
24563 if (isPowerOf2_64(MulAmt2))
24564 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24565 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
24567 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24568 DAG.getConstant(MulAmt2, VT));
24570 // Do not add new nodes to DAG combiner worklist.
24571 DCI.CombineTo(N, NewMul, false);
24576 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24577 SDValue N0 = N->getOperand(0);
24578 SDValue N1 = N->getOperand(1);
24579 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24580 EVT VT = N0.getValueType();
24582 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24583 // since the result of setcc_c is all zero's or all ones.
24584 if (VT.isInteger() && !VT.isVector() &&
24585 N1C && N0.getOpcode() == ISD::AND &&
24586 N0.getOperand(1).getOpcode() == ISD::Constant) {
24587 SDValue N00 = N0.getOperand(0);
24588 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
24589 ((N00.getOpcode() == ISD::ANY_EXTEND ||
24590 N00.getOpcode() == ISD::ZERO_EXTEND) &&
24591 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
24592 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24593 APInt ShAmt = N1C->getAPIntValue();
24594 Mask = Mask.shl(ShAmt);
24596 return DAG.getNode(ISD::AND, SDLoc(N), VT,
24597 N00, DAG.getConstant(Mask, VT));
24601 // Hardware support for vector shifts is sparse which makes us scalarize the
24602 // vector operations in many cases. Also, on sandybridge ADD is faster than
24604 // (shl V, 1) -> add V,V
24605 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24606 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24607 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24608 // We shift all of the values by one. In many cases we do not have
24609 // hardware support for this operation. This is better expressed as an ADD
24611 if (N1SplatC->getZExtValue() == 1)
24612 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24618 /// \brief Returns a vector of 0s if the node in input is a vector logical
24619 /// shift by a constant amount which is known to be bigger than or equal
24620 /// to the vector element size in bits.
24621 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24622 const X86Subtarget *Subtarget) {
24623 EVT VT = N->getValueType(0);
24625 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24626 (!Subtarget->hasInt256() ||
24627 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24630 SDValue Amt = N->getOperand(1);
24632 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24633 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24634 APInt ShiftAmt = AmtSplat->getAPIntValue();
24635 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24637 // SSE2/AVX2 logical shifts always return a vector of 0s
24638 // if the shift amount is bigger than or equal to
24639 // the element size. The constant shift amount will be
24640 // encoded as a 8-bit immediate.
24641 if (ShiftAmt.trunc(8).uge(MaxAmount))
24642 return getZeroVector(VT, Subtarget, DAG, DL);
24648 /// PerformShiftCombine - Combine shifts.
24649 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24650 TargetLowering::DAGCombinerInfo &DCI,
24651 const X86Subtarget *Subtarget) {
24652 if (N->getOpcode() == ISD::SHL) {
24653 SDValue V = PerformSHLCombine(N, DAG);
24654 if (V.getNode()) return V;
24657 if (N->getOpcode() != ISD::SRA) {
24658 // Try to fold this logical shift into a zero vector.
24659 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
24660 if (V.getNode()) return V;
24666 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24667 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24668 // and friends. Likewise for OR -> CMPNEQSS.
24669 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24670 TargetLowering::DAGCombinerInfo &DCI,
24671 const X86Subtarget *Subtarget) {
24674 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24675 // we're requiring SSE2 for both.
24676 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24677 SDValue N0 = N->getOperand(0);
24678 SDValue N1 = N->getOperand(1);
24679 SDValue CMP0 = N0->getOperand(1);
24680 SDValue CMP1 = N1->getOperand(1);
24683 // The SETCCs should both refer to the same CMP.
24684 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24687 SDValue CMP00 = CMP0->getOperand(0);
24688 SDValue CMP01 = CMP0->getOperand(1);
24689 EVT VT = CMP00.getValueType();
24691 if (VT == MVT::f32 || VT == MVT::f64) {
24692 bool ExpectingFlags = false;
24693 // Check for any users that want flags:
24694 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24695 !ExpectingFlags && UI != UE; ++UI)
24696 switch (UI->getOpcode()) {
24701 ExpectingFlags = true;
24703 case ISD::CopyToReg:
24704 case ISD::SIGN_EXTEND:
24705 case ISD::ZERO_EXTEND:
24706 case ISD::ANY_EXTEND:
24710 if (!ExpectingFlags) {
24711 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24712 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24714 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24715 X86::CondCode tmp = cc0;
24720 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24721 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24722 // FIXME: need symbolic constants for these magic numbers.
24723 // See X86ATTInstPrinter.cpp:printSSECC().
24724 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24725 if (Subtarget->hasAVX512()) {
24726 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24727 CMP01, DAG.getConstant(x86cc, MVT::i8));
24728 if (N->getValueType(0) != MVT::i1)
24729 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24733 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24734 CMP00.getValueType(), CMP00, CMP01,
24735 DAG.getConstant(x86cc, MVT::i8));
24737 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24738 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24740 if (is64BitFP && !Subtarget->is64Bit()) {
24741 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24742 // 64-bit integer, since that's not a legal type. Since
24743 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24744 // bits, but can do this little dance to extract the lowest 32 bits
24745 // and work with those going forward.
24746 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24748 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
24750 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24751 Vector32, DAG.getIntPtrConstant(0));
24755 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
24756 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24757 DAG.getConstant(1, IntVT));
24758 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
24759 return OneBitOfTruth;
24767 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24768 /// so it can be folded inside ANDNP.
24769 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24770 EVT VT = N->getValueType(0);
24772 // Match direct AllOnes for 128 and 256-bit vectors
24773 if (ISD::isBuildVectorAllOnes(N))
24776 // Look through a bit convert.
24777 if (N->getOpcode() == ISD::BITCAST)
24778 N = N->getOperand(0).getNode();
24780 // Sometimes the operand may come from a insert_subvector building a 256-bit
24782 if (VT.is256BitVector() &&
24783 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24784 SDValue V1 = N->getOperand(0);
24785 SDValue V2 = N->getOperand(1);
24787 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24788 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24789 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24790 ISD::isBuildVectorAllOnes(V2.getNode()))
24797 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24798 // register. In most cases we actually compare or select YMM-sized registers
24799 // and mixing the two types creates horrible code. This method optimizes
24800 // some of the transition sequences.
24801 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24802 TargetLowering::DAGCombinerInfo &DCI,
24803 const X86Subtarget *Subtarget) {
24804 EVT VT = N->getValueType(0);
24805 if (!VT.is256BitVector())
24808 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24809 N->getOpcode() == ISD::ZERO_EXTEND ||
24810 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24812 SDValue Narrow = N->getOperand(0);
24813 EVT NarrowVT = Narrow->getValueType(0);
24814 if (!NarrowVT.is128BitVector())
24817 if (Narrow->getOpcode() != ISD::XOR &&
24818 Narrow->getOpcode() != ISD::AND &&
24819 Narrow->getOpcode() != ISD::OR)
24822 SDValue N0 = Narrow->getOperand(0);
24823 SDValue N1 = Narrow->getOperand(1);
24826 // The Left side has to be a trunc.
24827 if (N0.getOpcode() != ISD::TRUNCATE)
24830 // The type of the truncated inputs.
24831 EVT WideVT = N0->getOperand(0)->getValueType(0);
24835 // The right side has to be a 'trunc' or a constant vector.
24836 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24837 ConstantSDNode *RHSConstSplat = nullptr;
24838 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24839 RHSConstSplat = RHSBV->getConstantSplatNode();
24840 if (!RHSTrunc && !RHSConstSplat)
24843 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24845 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24848 // Set N0 and N1 to hold the inputs to the new wide operation.
24849 N0 = N0->getOperand(0);
24850 if (RHSConstSplat) {
24851 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24852 SDValue(RHSConstSplat, 0));
24853 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24854 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24855 } else if (RHSTrunc) {
24856 N1 = N1->getOperand(0);
24859 // Generate the wide operation.
24860 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24861 unsigned Opcode = N->getOpcode();
24863 case ISD::ANY_EXTEND:
24865 case ISD::ZERO_EXTEND: {
24866 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24867 APInt Mask = APInt::getAllOnesValue(InBits);
24868 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24869 return DAG.getNode(ISD::AND, DL, VT,
24870 Op, DAG.getConstant(Mask, VT));
24872 case ISD::SIGN_EXTEND:
24873 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24874 Op, DAG.getValueType(NarrowVT));
24876 llvm_unreachable("Unexpected opcode");
24880 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24881 TargetLowering::DAGCombinerInfo &DCI,
24882 const X86Subtarget *Subtarget) {
24883 SDValue N0 = N->getOperand(0);
24884 SDValue N1 = N->getOperand(1);
24887 // A vector zext_in_reg may be represented as a shuffle,
24888 // feeding into a bitcast (this represents anyext) feeding into
24889 // an and with a mask.
24890 // We'd like to try to combine that into a shuffle with zero
24891 // plus a bitcast, removing the and.
24892 if (N0.getOpcode() != ISD::BITCAST ||
24893 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24896 // The other side of the AND should be a splat of 2^C, where C
24897 // is the number of bits in the source type.
24898 if (N1.getOpcode() == ISD::BITCAST)
24899 N1 = N1.getOperand(0);
24900 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24902 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24904 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24905 EVT SrcType = Shuffle->getValueType(0);
24907 // We expect a single-source shuffle
24908 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24911 unsigned SrcSize = SrcType.getScalarSizeInBits();
24913 APInt SplatValue, SplatUndef;
24914 unsigned SplatBitSize;
24916 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24917 SplatBitSize, HasAnyUndefs))
24920 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24921 // Make sure the splat matches the mask we expect
24922 if (SplatBitSize > ResSize ||
24923 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24926 // Make sure the input and output size make sense
24927 if (SrcSize >= ResSize || ResSize % SrcSize)
24930 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24931 // The number of u's between each two values depends on the ratio between
24932 // the source and dest type.
24933 unsigned ZextRatio = ResSize / SrcSize;
24934 bool IsZext = true;
24935 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24936 if (i % ZextRatio) {
24937 if (Shuffle->getMaskElt(i) > 0) {
24943 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24944 // Expected element number
24954 // Ok, perform the transformation - replace the shuffle with
24955 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24956 // (instead of undef) where the k elements come from the zero vector.
24957 SmallVector<int, 8> Mask;
24958 unsigned NumElems = SrcType.getVectorNumElements();
24959 for (unsigned i = 0; i < NumElems; ++i)
24961 Mask.push_back(NumElems);
24963 Mask.push_back(i / ZextRatio);
24965 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24966 Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
24967 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
24970 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24971 TargetLowering::DAGCombinerInfo &DCI,
24972 const X86Subtarget *Subtarget) {
24973 if (DCI.isBeforeLegalizeOps())
24976 SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget);
24977 if (Zext.getNode())
24980 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
24984 EVT VT = N->getValueType(0);
24985 SDValue N0 = N->getOperand(0);
24986 SDValue N1 = N->getOperand(1);
24989 // Create BEXTR instructions
24990 // BEXTR is ((X >> imm) & (2**size-1))
24991 if (VT == MVT::i32 || VT == MVT::i64) {
24992 // Check for BEXTR.
24993 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24994 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24995 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24996 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24997 if (MaskNode && ShiftNode) {
24998 uint64_t Mask = MaskNode->getZExtValue();
24999 uint64_t Shift = ShiftNode->getZExtValue();
25000 if (isMask_64(Mask)) {
25001 uint64_t MaskSize = countPopulation(Mask);
25002 if (Shift + MaskSize <= VT.getSizeInBits())
25003 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25004 DAG.getConstant(Shift | (MaskSize << 8), VT));
25012 // Want to form ANDNP nodes:
25013 // 1) In the hopes of then easily combining them with OR and AND nodes
25014 // to form PBLEND/PSIGN.
25015 // 2) To match ANDN packed intrinsics
25016 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25019 // Check LHS for vnot
25020 if (N0.getOpcode() == ISD::XOR &&
25021 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25022 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25023 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25025 // Check RHS for vnot
25026 if (N1.getOpcode() == ISD::XOR &&
25027 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25028 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25029 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25034 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25035 TargetLowering::DAGCombinerInfo &DCI,
25036 const X86Subtarget *Subtarget) {
25037 if (DCI.isBeforeLegalizeOps())
25040 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
25044 SDValue N0 = N->getOperand(0);
25045 SDValue N1 = N->getOperand(1);
25046 EVT VT = N->getValueType(0);
25048 // look for psign/blend
25049 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25050 if (!Subtarget->hasSSSE3() ||
25051 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25054 // Canonicalize pandn to RHS
25055 if (N0.getOpcode() == X86ISD::ANDNP)
25057 // or (and (m, y), (pandn m, x))
25058 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25059 SDValue Mask = N1.getOperand(0);
25060 SDValue X = N1.getOperand(1);
25062 if (N0.getOperand(0) == Mask)
25063 Y = N0.getOperand(1);
25064 if (N0.getOperand(1) == Mask)
25065 Y = N0.getOperand(0);
25067 // Check to see if the mask appeared in both the AND and ANDNP and
25071 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25072 // Look through mask bitcast.
25073 if (Mask.getOpcode() == ISD::BITCAST)
25074 Mask = Mask.getOperand(0);
25075 if (X.getOpcode() == ISD::BITCAST)
25076 X = X.getOperand(0);
25077 if (Y.getOpcode() == ISD::BITCAST)
25078 Y = Y.getOperand(0);
25080 EVT MaskVT = Mask.getValueType();
25082 // Validate that the Mask operand is a vector sra node.
25083 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25084 // there is no psrai.b
25085 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25086 unsigned SraAmt = ~0;
25087 if (Mask.getOpcode() == ISD::SRA) {
25088 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25089 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25090 SraAmt = AmtConst->getZExtValue();
25091 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25092 SDValue SraC = Mask.getOperand(1);
25093 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25095 if ((SraAmt + 1) != EltBits)
25100 // Now we know we at least have a plendvb with the mask val. See if
25101 // we can form a psignb/w/d.
25102 // psign = x.type == y.type == mask.type && y = sub(0, x);
25103 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25104 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25105 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25106 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25107 "Unsupported VT for PSIGN");
25108 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25109 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
25111 // PBLENDVB only available on SSE 4.1
25112 if (!Subtarget->hasSSE41())
25115 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25117 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
25118 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
25119 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
25120 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25121 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
25125 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25128 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25129 MachineFunction &MF = DAG.getMachineFunction();
25131 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
25133 // SHLD/SHRD instructions have lower register pressure, but on some
25134 // platforms they have higher latency than the equivalent
25135 // series of shifts/or that would otherwise be generated.
25136 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25137 // have higher latencies and we are not optimizing for size.
25138 if (!OptForSize && Subtarget->isSHLDSlow())
25141 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25143 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25145 if (!N0.hasOneUse() || !N1.hasOneUse())
25148 SDValue ShAmt0 = N0.getOperand(1);
25149 if (ShAmt0.getValueType() != MVT::i8)
25151 SDValue ShAmt1 = N1.getOperand(1);
25152 if (ShAmt1.getValueType() != MVT::i8)
25154 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25155 ShAmt0 = ShAmt0.getOperand(0);
25156 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25157 ShAmt1 = ShAmt1.getOperand(0);
25160 unsigned Opc = X86ISD::SHLD;
25161 SDValue Op0 = N0.getOperand(0);
25162 SDValue Op1 = N1.getOperand(0);
25163 if (ShAmt0.getOpcode() == ISD::SUB) {
25164 Opc = X86ISD::SHRD;
25165 std::swap(Op0, Op1);
25166 std::swap(ShAmt0, ShAmt1);
25169 unsigned Bits = VT.getSizeInBits();
25170 if (ShAmt1.getOpcode() == ISD::SUB) {
25171 SDValue Sum = ShAmt1.getOperand(0);
25172 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25173 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25174 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25175 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25176 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25177 return DAG.getNode(Opc, DL, VT,
25179 DAG.getNode(ISD::TRUNCATE, DL,
25182 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25183 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25185 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25186 return DAG.getNode(Opc, DL, VT,
25187 N0.getOperand(0), N1.getOperand(0),
25188 DAG.getNode(ISD::TRUNCATE, DL,
25195 // Generate NEG and CMOV for integer abs.
25196 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25197 EVT VT = N->getValueType(0);
25199 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25200 // 8-bit integer abs to NEG and CMOV.
25201 if (VT.isInteger() && VT.getSizeInBits() == 8)
25204 SDValue N0 = N->getOperand(0);
25205 SDValue N1 = N->getOperand(1);
25208 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25209 // and change it to SUB and CMOV.
25210 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25211 N0.getOpcode() == ISD::ADD &&
25212 N0.getOperand(1) == N1 &&
25213 N1.getOpcode() == ISD::SRA &&
25214 N1.getOperand(0) == N0.getOperand(0))
25215 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25216 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25217 // Generate SUB & CMOV.
25218 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25219 DAG.getConstant(0, VT), N0.getOperand(0));
25221 SDValue Ops[] = { N0.getOperand(0), Neg,
25222 DAG.getConstant(X86::COND_GE, MVT::i8),
25223 SDValue(Neg.getNode(), 1) };
25224 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25229 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
25230 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25231 TargetLowering::DAGCombinerInfo &DCI,
25232 const X86Subtarget *Subtarget) {
25233 if (DCI.isBeforeLegalizeOps())
25236 if (Subtarget->hasCMov()) {
25237 SDValue RV = performIntegerAbsCombine(N, DAG);
25245 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25246 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25247 TargetLowering::DAGCombinerInfo &DCI,
25248 const X86Subtarget *Subtarget) {
25249 LoadSDNode *Ld = cast<LoadSDNode>(N);
25250 EVT RegVT = Ld->getValueType(0);
25251 EVT MemVT = Ld->getMemoryVT();
25253 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25255 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25256 // into two 16-byte operations.
25257 ISD::LoadExtType Ext = Ld->getExtensionType();
25258 unsigned Alignment = Ld->getAlignment();
25259 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
25260 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
25261 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
25262 unsigned NumElems = RegVT.getVectorNumElements();
25266 SDValue Ptr = Ld->getBasePtr();
25267 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
25269 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25271 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25272 Ld->getPointerInfo(), Ld->isVolatile(),
25273 Ld->isNonTemporal(), Ld->isInvariant(),
25275 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25276 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25277 Ld->getPointerInfo(), Ld->isVolatile(),
25278 Ld->isNonTemporal(), Ld->isInvariant(),
25279 std::min(16U, Alignment));
25280 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25282 Load2.getValue(1));
25284 SDValue NewVec = DAG.getUNDEF(RegVT);
25285 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25286 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25287 return DCI.CombineTo(N, NewVec, TF, true);
25293 /// PerformMLOADCombine - Resolve extending loads
25294 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25295 TargetLowering::DAGCombinerInfo &DCI,
25296 const X86Subtarget *Subtarget) {
25297 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25298 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25301 EVT VT = Mld->getValueType(0);
25302 unsigned NumElems = VT.getVectorNumElements();
25303 EVT LdVT = Mld->getMemoryVT();
25306 assert(LdVT != VT && "Cannot extend to the same type");
25307 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25308 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25309 // From, To sizes and ElemCount must be pow of two
25310 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25311 "Unexpected size for extending masked load");
25313 unsigned SizeRatio = ToSz / FromSz;
25314 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25316 // Create a type on which we perform the shuffle
25317 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25318 LdVT.getScalarType(), NumElems*SizeRatio);
25319 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25321 // Convert Src0 value
25322 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
25323 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25324 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25325 for (unsigned i = 0; i != NumElems; ++i)
25326 ShuffleVec[i] = i * SizeRatio;
25328 // Can't shuffle using an illegal type.
25329 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
25330 && "WideVecVT should be legal");
25331 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25332 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25334 // Prepare the new mask
25336 SDValue Mask = Mld->getMask();
25337 if (Mask.getValueType() == VT) {
25338 // Mask and original value have the same type
25339 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
25340 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25341 for (unsigned i = 0; i != NumElems; ++i)
25342 ShuffleVec[i] = i * SizeRatio;
25343 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25344 ShuffleVec[i] = NumElems*SizeRatio;
25345 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25346 DAG.getConstant(0, WideVecVT),
25350 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25351 unsigned WidenNumElts = NumElems*SizeRatio;
25352 unsigned MaskNumElts = VT.getVectorNumElements();
25353 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25356 unsigned NumConcat = WidenNumElts / MaskNumElts;
25357 SmallVector<SDValue, 16> Ops(NumConcat);
25358 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
25360 for (unsigned i = 1; i != NumConcat; ++i)
25363 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25366 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25367 Mld->getBasePtr(), NewMask, WideSrc0,
25368 Mld->getMemoryVT(), Mld->getMemOperand(),
25370 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25371 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25374 /// PerformMSTORECombine - Resolve truncating stores
25375 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25376 const X86Subtarget *Subtarget) {
25377 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25378 if (!Mst->isTruncatingStore())
25381 EVT VT = Mst->getValue().getValueType();
25382 unsigned NumElems = VT.getVectorNumElements();
25383 EVT StVT = Mst->getMemoryVT();
25386 assert(StVT != VT && "Cannot truncate to the same type");
25387 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25388 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25390 // From, To sizes and ElemCount must be pow of two
25391 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25392 "Unexpected size for truncating masked store");
25393 // We are going to use the original vector elt for storing.
25394 // Accumulated smaller vector elements must be a multiple of the store size.
25395 assert (((NumElems * FromSz) % ToSz) == 0 &&
25396 "Unexpected ratio for truncating masked store");
25398 unsigned SizeRatio = FromSz / ToSz;
25399 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25401 // Create a type on which we perform the shuffle
25402 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25403 StVT.getScalarType(), NumElems*SizeRatio);
25405 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25407 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
25408 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25409 for (unsigned i = 0; i != NumElems; ++i)
25410 ShuffleVec[i] = i * SizeRatio;
25412 // Can't shuffle using an illegal type.
25413 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
25414 && "WideVecVT should be legal");
25416 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25417 DAG.getUNDEF(WideVecVT),
25421 SDValue Mask = Mst->getMask();
25422 if (Mask.getValueType() == VT) {
25423 // Mask and original value have the same type
25424 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
25425 for (unsigned i = 0; i != NumElems; ++i)
25426 ShuffleVec[i] = i * SizeRatio;
25427 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25428 ShuffleVec[i] = NumElems*SizeRatio;
25429 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25430 DAG.getConstant(0, WideVecVT),
25434 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25435 unsigned WidenNumElts = NumElems*SizeRatio;
25436 unsigned MaskNumElts = VT.getVectorNumElements();
25437 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25440 unsigned NumConcat = WidenNumElts / MaskNumElts;
25441 SmallVector<SDValue, 16> Ops(NumConcat);
25442 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
25444 for (unsigned i = 1; i != NumConcat; ++i)
25447 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25450 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25451 NewMask, StVT, Mst->getMemOperand(), false);
25453 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25454 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25455 const X86Subtarget *Subtarget) {
25456 StoreSDNode *St = cast<StoreSDNode>(N);
25457 EVT VT = St->getValue().getValueType();
25458 EVT StVT = St->getMemoryVT();
25460 SDValue StoredVal = St->getOperand(1);
25461 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25463 // If we are saving a concatenation of two XMM registers and 32-byte stores
25464 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25465 unsigned Alignment = St->getAlignment();
25466 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
25467 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
25468 StVT == VT && !IsAligned) {
25469 unsigned NumElems = VT.getVectorNumElements();
25473 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25474 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25476 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
25477 SDValue Ptr0 = St->getBasePtr();
25478 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25480 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25481 St->getPointerInfo(), St->isVolatile(),
25482 St->isNonTemporal(), Alignment);
25483 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25484 St->getPointerInfo(), St->isVolatile(),
25485 St->isNonTemporal(),
25486 std::min(16U, Alignment));
25487 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25490 // Optimize trunc store (of multiple scalars) to shuffle and store.
25491 // First, pack all of the elements in one place. Next, store to memory
25492 // in fewer chunks.
25493 if (St->isTruncatingStore() && VT.isVector()) {
25494 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25495 unsigned NumElems = VT.getVectorNumElements();
25496 assert(StVT != VT && "Cannot truncate to the same type");
25497 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25498 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25500 // From, To sizes and ElemCount must be pow of two
25501 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25502 // We are going to use the original vector elt for storing.
25503 // Accumulated smaller vector elements must be a multiple of the store size.
25504 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25506 unsigned SizeRatio = FromSz / ToSz;
25508 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25510 // Create a type on which we perform the shuffle
25511 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25512 StVT.getScalarType(), NumElems*SizeRatio);
25514 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25516 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
25517 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25518 for (unsigned i = 0; i != NumElems; ++i)
25519 ShuffleVec[i] = i * SizeRatio;
25521 // Can't shuffle using an illegal type.
25522 if (!TLI.isTypeLegal(WideVecVT))
25525 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25526 DAG.getUNDEF(WideVecVT),
25528 // At this point all of the data is stored at the bottom of the
25529 // register. We now need to save it to mem.
25531 // Find the largest store unit
25532 MVT StoreType = MVT::i8;
25533 for (MVT Tp : MVT::integer_valuetypes()) {
25534 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25538 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25539 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25540 (64 <= NumElems * ToSz))
25541 StoreType = MVT::f64;
25543 // Bitcast the original vector into a vector of store-size units
25544 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25545 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25546 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25547 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
25548 SmallVector<SDValue, 8> Chains;
25549 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
25550 TLI.getPointerTy());
25551 SDValue Ptr = St->getBasePtr();
25553 // Perform one or more big stores into memory.
25554 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25555 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25556 StoreType, ShuffWide,
25557 DAG.getIntPtrConstant(i));
25558 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25559 St->getPointerInfo(), St->isVolatile(),
25560 St->isNonTemporal(), St->getAlignment());
25561 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25562 Chains.push_back(Ch);
25565 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25568 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25569 // the FP state in cases where an emms may be missing.
25570 // A preferable solution to the general problem is to figure out the right
25571 // places to insert EMMS. This qualifies as a quick hack.
25573 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25574 if (VT.getSizeInBits() != 64)
25577 const Function *F = DAG.getMachineFunction().getFunction();
25578 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25579 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
25580 && Subtarget->hasSSE2();
25581 if ((VT.isVector() ||
25582 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25583 isa<LoadSDNode>(St->getValue()) &&
25584 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25585 St->getChain().hasOneUse() && !St->isVolatile()) {
25586 SDNode* LdVal = St->getValue().getNode();
25587 LoadSDNode *Ld = nullptr;
25588 int TokenFactorIndex = -1;
25589 SmallVector<SDValue, 8> Ops;
25590 SDNode* ChainVal = St->getChain().getNode();
25591 // Must be a store of a load. We currently handle two cases: the load
25592 // is a direct child, and it's under an intervening TokenFactor. It is
25593 // possible to dig deeper under nested TokenFactors.
25594 if (ChainVal == LdVal)
25595 Ld = cast<LoadSDNode>(St->getChain());
25596 else if (St->getValue().hasOneUse() &&
25597 ChainVal->getOpcode() == ISD::TokenFactor) {
25598 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25599 if (ChainVal->getOperand(i).getNode() == LdVal) {
25600 TokenFactorIndex = i;
25601 Ld = cast<LoadSDNode>(St->getValue());
25603 Ops.push_back(ChainVal->getOperand(i));
25607 if (!Ld || !ISD::isNormalLoad(Ld))
25610 // If this is not the MMX case, i.e. we are just turning i64 load/store
25611 // into f64 load/store, avoid the transformation if there are multiple
25612 // uses of the loaded value.
25613 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25618 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25619 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25621 if (Subtarget->is64Bit() || F64IsLegal) {
25622 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25623 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25624 Ld->getPointerInfo(), Ld->isVolatile(),
25625 Ld->isNonTemporal(), Ld->isInvariant(),
25626 Ld->getAlignment());
25627 SDValue NewChain = NewLd.getValue(1);
25628 if (TokenFactorIndex != -1) {
25629 Ops.push_back(NewChain);
25630 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25632 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25633 St->getPointerInfo(),
25634 St->isVolatile(), St->isNonTemporal(),
25635 St->getAlignment());
25638 // Otherwise, lower to two pairs of 32-bit loads / stores.
25639 SDValue LoAddr = Ld->getBasePtr();
25640 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25641 DAG.getConstant(4, MVT::i32));
25643 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25644 Ld->getPointerInfo(),
25645 Ld->isVolatile(), Ld->isNonTemporal(),
25646 Ld->isInvariant(), Ld->getAlignment());
25647 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25648 Ld->getPointerInfo().getWithOffset(4),
25649 Ld->isVolatile(), Ld->isNonTemporal(),
25651 MinAlign(Ld->getAlignment(), 4));
25653 SDValue NewChain = LoLd.getValue(1);
25654 if (TokenFactorIndex != -1) {
25655 Ops.push_back(LoLd);
25656 Ops.push_back(HiLd);
25657 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25660 LoAddr = St->getBasePtr();
25661 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25662 DAG.getConstant(4, MVT::i32));
25664 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25665 St->getPointerInfo(),
25666 St->isVolatile(), St->isNonTemporal(),
25667 St->getAlignment());
25668 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25669 St->getPointerInfo().getWithOffset(4),
25671 St->isNonTemporal(),
25672 MinAlign(St->getAlignment(), 4));
25673 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25678 /// Return 'true' if this vector operation is "horizontal"
25679 /// and return the operands for the horizontal operation in LHS and RHS. A
25680 /// horizontal operation performs the binary operation on successive elements
25681 /// of its first operand, then on successive elements of its second operand,
25682 /// returning the resulting values in a vector. For example, if
25683 /// A = < float a0, float a1, float a2, float a3 >
25685 /// B = < float b0, float b1, float b2, float b3 >
25686 /// then the result of doing a horizontal operation on A and B is
25687 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25688 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25689 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25690 /// set to A, RHS to B, and the routine returns 'true'.
25691 /// Note that the binary operation should have the property that if one of the
25692 /// operands is UNDEF then the result is UNDEF.
25693 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25694 // Look for the following pattern: if
25695 // A = < float a0, float a1, float a2, float a3 >
25696 // B = < float b0, float b1, float b2, float b3 >
25698 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25699 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25700 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25701 // which is A horizontal-op B.
25703 // At least one of the operands should be a vector shuffle.
25704 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25705 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25708 MVT VT = LHS.getSimpleValueType();
25710 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25711 "Unsupported vector type for horizontal add/sub");
25713 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25714 // operate independently on 128-bit lanes.
25715 unsigned NumElts = VT.getVectorNumElements();
25716 unsigned NumLanes = VT.getSizeInBits()/128;
25717 unsigned NumLaneElts = NumElts / NumLanes;
25718 assert((NumLaneElts % 2 == 0) &&
25719 "Vector type should have an even number of elements in each lane");
25720 unsigned HalfLaneElts = NumLaneElts/2;
25722 // View LHS in the form
25723 // LHS = VECTOR_SHUFFLE A, B, LMask
25724 // If LHS is not a shuffle then pretend it is the shuffle
25725 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25726 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25729 SmallVector<int, 16> LMask(NumElts);
25730 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25731 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25732 A = LHS.getOperand(0);
25733 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25734 B = LHS.getOperand(1);
25735 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25736 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25738 if (LHS.getOpcode() != ISD::UNDEF)
25740 for (unsigned i = 0; i != NumElts; ++i)
25744 // Likewise, view RHS in the form
25745 // RHS = VECTOR_SHUFFLE C, D, RMask
25747 SmallVector<int, 16> RMask(NumElts);
25748 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25749 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25750 C = RHS.getOperand(0);
25751 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25752 D = RHS.getOperand(1);
25753 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25754 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25756 if (RHS.getOpcode() != ISD::UNDEF)
25758 for (unsigned i = 0; i != NumElts; ++i)
25762 // Check that the shuffles are both shuffling the same vectors.
25763 if (!(A == C && B == D) && !(A == D && B == C))
25766 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25767 if (!A.getNode() && !B.getNode())
25770 // If A and B occur in reverse order in RHS, then "swap" them (which means
25771 // rewriting the mask).
25773 CommuteVectorShuffleMask(RMask, NumElts);
25775 // At this point LHS and RHS are equivalent to
25776 // LHS = VECTOR_SHUFFLE A, B, LMask
25777 // RHS = VECTOR_SHUFFLE A, B, RMask
25778 // Check that the masks correspond to performing a horizontal operation.
25779 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25780 for (unsigned i = 0; i != NumLaneElts; ++i) {
25781 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25783 // Ignore any UNDEF components.
25784 if (LIdx < 0 || RIdx < 0 ||
25785 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25786 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25789 // Check that successive elements are being operated on. If not, this is
25790 // not a horizontal operation.
25791 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25792 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25793 if (!(LIdx == Index && RIdx == Index + 1) &&
25794 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25799 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25800 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25804 /// Do target-specific dag combines on floating point adds.
25805 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25806 const X86Subtarget *Subtarget) {
25807 EVT VT = N->getValueType(0);
25808 SDValue LHS = N->getOperand(0);
25809 SDValue RHS = N->getOperand(1);
25811 // Try to synthesize horizontal adds from adds of shuffles.
25812 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25813 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25814 isHorizontalBinOp(LHS, RHS, true))
25815 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25819 /// Do target-specific dag combines on floating point subs.
25820 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25821 const X86Subtarget *Subtarget) {
25822 EVT VT = N->getValueType(0);
25823 SDValue LHS = N->getOperand(0);
25824 SDValue RHS = N->getOperand(1);
25826 // Try to synthesize horizontal subs from subs of shuffles.
25827 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25828 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25829 isHorizontalBinOp(LHS, RHS, false))
25830 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25834 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25835 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
25836 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25838 // F[X]OR(0.0, x) -> x
25839 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25840 if (C->getValueAPF().isPosZero())
25841 return N->getOperand(1);
25843 // F[X]OR(x, 0.0) -> x
25844 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25845 if (C->getValueAPF().isPosZero())
25846 return N->getOperand(0);
25850 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25851 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25852 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25854 // Only perform optimizations if UnsafeMath is used.
25855 if (!DAG.getTarget().Options.UnsafeFPMath)
25858 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25859 // into FMINC and FMAXC, which are Commutative operations.
25860 unsigned NewOp = 0;
25861 switch (N->getOpcode()) {
25862 default: llvm_unreachable("unknown opcode");
25863 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25864 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25867 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25868 N->getOperand(0), N->getOperand(1));
25871 /// Do target-specific dag combines on X86ISD::FAND nodes.
25872 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25873 // FAND(0.0, x) -> 0.0
25874 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25875 if (C->getValueAPF().isPosZero())
25876 return N->getOperand(0);
25878 // FAND(x, 0.0) -> 0.0
25879 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25880 if (C->getValueAPF().isPosZero())
25881 return N->getOperand(1);
25886 /// Do target-specific dag combines on X86ISD::FANDN nodes
25887 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25888 // FANDN(0.0, x) -> x
25889 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25890 if (C->getValueAPF().isPosZero())
25891 return N->getOperand(1);
25893 // FANDN(x, 0.0) -> 0.0
25894 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25895 if (C->getValueAPF().isPosZero())
25896 return N->getOperand(1);
25901 static SDValue PerformBTCombine(SDNode *N,
25903 TargetLowering::DAGCombinerInfo &DCI) {
25904 // BT ignores high bits in the bit index operand.
25905 SDValue Op1 = N->getOperand(1);
25906 if (Op1.hasOneUse()) {
25907 unsigned BitWidth = Op1.getValueSizeInBits();
25908 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25909 APInt KnownZero, KnownOne;
25910 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25911 !DCI.isBeforeLegalizeOps());
25912 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25913 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25914 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25915 DCI.CommitTargetLoweringOpt(TLO);
25920 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25921 SDValue Op = N->getOperand(0);
25922 if (Op.getOpcode() == ISD::BITCAST)
25923 Op = Op.getOperand(0);
25924 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25925 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25926 VT.getVectorElementType().getSizeInBits() ==
25927 OpVT.getVectorElementType().getSizeInBits()) {
25928 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25933 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25934 const X86Subtarget *Subtarget) {
25935 EVT VT = N->getValueType(0);
25936 if (!VT.isVector())
25939 SDValue N0 = N->getOperand(0);
25940 SDValue N1 = N->getOperand(1);
25941 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25944 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25945 // both SSE and AVX2 since there is no sign-extended shift right
25946 // operation on a vector with 64-bit elements.
25947 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25948 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25949 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25950 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25951 SDValue N00 = N0.getOperand(0);
25953 // EXTLOAD has a better solution on AVX2,
25954 // it may be replaced with X86ISD::VSEXT node.
25955 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25956 if (!ISD::isNormalLoad(N00.getNode()))
25959 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25960 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25962 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25968 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25969 TargetLowering::DAGCombinerInfo &DCI,
25970 const X86Subtarget *Subtarget) {
25971 SDValue N0 = N->getOperand(0);
25972 EVT VT = N->getValueType(0);
25974 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25975 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25976 // This exposes the sext to the sdivrem lowering, so that it directly extends
25977 // from AH (which we otherwise need to do contortions to access).
25978 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25979 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
25981 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25982 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
25983 N0.getOperand(0), N0.getOperand(1));
25984 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25985 return R.getValue(1);
25988 if (!DCI.isBeforeLegalizeOps())
25991 if (!Subtarget->hasFp256())
25994 if (VT.isVector() && VT.getSizeInBits() == 256) {
25995 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
26003 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26004 const X86Subtarget* Subtarget) {
26006 EVT VT = N->getValueType(0);
26008 // Let legalize expand this if it isn't a legal type yet.
26009 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26012 EVT ScalarVT = VT.getScalarType();
26013 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26014 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
26017 SDValue A = N->getOperand(0);
26018 SDValue B = N->getOperand(1);
26019 SDValue C = N->getOperand(2);
26021 bool NegA = (A.getOpcode() == ISD::FNEG);
26022 bool NegB = (B.getOpcode() == ISD::FNEG);
26023 bool NegC = (C.getOpcode() == ISD::FNEG);
26025 // Negative multiplication when NegA xor NegB
26026 bool NegMul = (NegA != NegB);
26028 A = A.getOperand(0);
26030 B = B.getOperand(0);
26032 C = C.getOperand(0);
26036 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26038 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26040 return DAG.getNode(Opcode, dl, VT, A, B, C);
26043 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26044 TargetLowering::DAGCombinerInfo &DCI,
26045 const X86Subtarget *Subtarget) {
26046 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26047 // (and (i32 x86isd::setcc_carry), 1)
26048 // This eliminates the zext. This transformation is necessary because
26049 // ISD::SETCC is always legalized to i8.
26051 SDValue N0 = N->getOperand(0);
26052 EVT VT = N->getValueType(0);
26054 if (N0.getOpcode() == ISD::AND &&
26056 N0.getOperand(0).hasOneUse()) {
26057 SDValue N00 = N0.getOperand(0);
26058 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26059 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26060 if (!C || C->getZExtValue() != 1)
26062 return DAG.getNode(ISD::AND, dl, VT,
26063 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26064 N00.getOperand(0), N00.getOperand(1)),
26065 DAG.getConstant(1, VT));
26069 if (N0.getOpcode() == ISD::TRUNCATE &&
26071 N0.getOperand(0).hasOneUse()) {
26072 SDValue N00 = N0.getOperand(0);
26073 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26074 return DAG.getNode(ISD::AND, dl, VT,
26075 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26076 N00.getOperand(0), N00.getOperand(1)),
26077 DAG.getConstant(1, VT));
26080 if (VT.is256BitVector()) {
26081 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
26086 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26087 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26088 // This exposes the zext to the udivrem lowering, so that it directly extends
26089 // from AH (which we otherwise need to do contortions to access).
26090 if (N0.getOpcode() == ISD::UDIVREM &&
26091 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26092 (VT == MVT::i32 || VT == MVT::i64)) {
26093 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26094 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26095 N0.getOperand(0), N0.getOperand(1));
26096 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26097 return R.getValue(1);
26103 // Optimize x == -y --> x+y == 0
26104 // x != -y --> x+y != 0
26105 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26106 const X86Subtarget* Subtarget) {
26107 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26108 SDValue LHS = N->getOperand(0);
26109 SDValue RHS = N->getOperand(1);
26110 EVT VT = N->getValueType(0);
26113 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26114 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26115 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26116 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
26117 LHS.getValueType(), RHS, LHS.getOperand(1));
26118 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
26119 addV, DAG.getConstant(0, addV.getValueType()), CC);
26121 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26122 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26123 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26124 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
26125 RHS.getValueType(), LHS, RHS.getOperand(1));
26126 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
26127 addV, DAG.getConstant(0, addV.getValueType()), CC);
26130 if (VT.getScalarType() == MVT::i1) {
26131 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26132 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26133 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
26134 if (!IsSEXT0 && !IsVZero0)
26136 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
26137 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26138 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26140 if (!IsSEXT1 && !IsVZero1)
26143 if (IsSEXT0 && IsVZero1) {
26144 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
26145 if (CC == ISD::SETEQ)
26146 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26147 return LHS.getOperand(0);
26149 if (IsSEXT1 && IsVZero0) {
26150 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
26151 if (CC == ISD::SETEQ)
26152 return DAG.getNOT(DL, RHS.getOperand(0), VT);
26153 return RHS.getOperand(0);
26160 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26161 const X86Subtarget *Subtarget) {
26163 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26164 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26165 "X86insertps is only defined for v4x32");
26167 SDValue Ld = N->getOperand(1);
26168 if (MayFoldLoad(Ld)) {
26169 // Extract the countS bits from the immediate so we can get the proper
26170 // address when narrowing the vector load to a specific element.
26171 // When the second source op is a memory address, interps doesn't use
26172 // countS and just gets an f32 from that address.
26173 unsigned DestIndex =
26174 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26175 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26179 // Create this as a scalar to vector to match the instruction pattern.
26180 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26181 // countS bits are ignored when loading from memory on insertps, which
26182 // means we don't need to explicitly set them to 0.
26183 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26184 LoadScalarToVector, N->getOperand(2));
26187 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26188 // as "sbb reg,reg", since it can be extended without zext and produces
26189 // an all-ones bit which is more useful than 0/1 in some cases.
26190 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26193 return DAG.getNode(ISD::AND, DL, VT,
26194 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26195 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
26196 DAG.getConstant(1, VT));
26197 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26198 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26199 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26200 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
26203 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26204 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26205 TargetLowering::DAGCombinerInfo &DCI,
26206 const X86Subtarget *Subtarget) {
26208 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26209 SDValue EFLAGS = N->getOperand(1);
26211 if (CC == X86::COND_A) {
26212 // Try to convert COND_A into COND_B in an attempt to facilitate
26213 // materializing "setb reg".
26215 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26216 // cannot take an immediate as its first operand.
26218 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26219 EFLAGS.getValueType().isInteger() &&
26220 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26221 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26222 EFLAGS.getNode()->getVTList(),
26223 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26224 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26225 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26229 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26230 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26232 if (CC == X86::COND_B)
26233 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26237 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
26238 if (Flags.getNode()) {
26239 SDValue Cond = DAG.getConstant(CC, MVT::i8);
26240 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26246 // Optimize branch condition evaluation.
26248 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26249 TargetLowering::DAGCombinerInfo &DCI,
26250 const X86Subtarget *Subtarget) {
26252 SDValue Chain = N->getOperand(0);
26253 SDValue Dest = N->getOperand(1);
26254 SDValue EFLAGS = N->getOperand(3);
26255 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26259 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
26260 if (Flags.getNode()) {
26261 SDValue Cond = DAG.getConstant(CC, MVT::i8);
26262 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26269 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26270 SelectionDAG &DAG) {
26271 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26272 // optimize away operation when it's from a constant.
26274 // The general transformation is:
26275 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26276 // AND(VECTOR_CMP(x,y), constant2)
26277 // constant2 = UNARYOP(constant)
26279 // Early exit if this isn't a vector operation, the operand of the
26280 // unary operation isn't a bitwise AND, or if the sizes of the operations
26281 // aren't the same.
26282 EVT VT = N->getValueType(0);
26283 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26284 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26285 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26288 // Now check that the other operand of the AND is a constant. We could
26289 // make the transformation for non-constant splats as well, but it's unclear
26290 // that would be a benefit as it would not eliminate any operations, just
26291 // perform one more step in scalar code before moving to the vector unit.
26292 if (BuildVectorSDNode *BV =
26293 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26294 // Bail out if the vector isn't a constant.
26295 if (!BV->isConstant())
26298 // Everything checks out. Build up the new and improved node.
26300 EVT IntVT = BV->getValueType(0);
26301 // Create a new constant of the appropriate type for the transformed
26303 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26304 // The AND node needs bitcasts to/from an integer vector type around it.
26305 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
26306 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26307 N->getOperand(0)->getOperand(0), MaskConst);
26308 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
26315 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26316 const X86Subtarget *Subtarget) {
26317 // First try to optimize away the conversion entirely when it's
26318 // conditionally from a constant. Vectors only.
26319 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
26320 if (Res != SDValue())
26323 // Now move on to more general possibilities.
26324 SDValue Op0 = N->getOperand(0);
26325 EVT InVT = Op0->getValueType(0);
26327 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
26328 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
26330 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
26331 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26332 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
26335 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26336 // a 32-bit target where SSE doesn't support i64->FP operations.
26337 if (Op0.getOpcode() == ISD::LOAD) {
26338 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26339 EVT VT = Ld->getValueType(0);
26340 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
26341 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26342 !Subtarget->is64Bit() && VT == MVT::i64) {
26343 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26344 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
26345 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26352 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26353 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26354 X86TargetLowering::DAGCombinerInfo &DCI) {
26355 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26356 // the result is either zero or one (depending on the input carry bit).
26357 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26358 if (X86::isZeroNode(N->getOperand(0)) &&
26359 X86::isZeroNode(N->getOperand(1)) &&
26360 // We don't have a good way to replace an EFLAGS use, so only do this when
26362 SDValue(N, 1).use_empty()) {
26364 EVT VT = N->getValueType(0);
26365 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
26366 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26367 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26368 DAG.getConstant(X86::COND_B,MVT::i8),
26370 DAG.getConstant(1, VT));
26371 return DCI.CombineTo(N, Res1, CarryOut);
26377 // fold (add Y, (sete X, 0)) -> adc 0, Y
26378 // (add Y, (setne X, 0)) -> sbb -1, Y
26379 // (sub (sete X, 0), Y) -> sbb 0, Y
26380 // (sub (setne X, 0), Y) -> adc -1, Y
26381 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26384 // Look through ZExts.
26385 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26386 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26389 SDValue SetCC = Ext.getOperand(0);
26390 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26393 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26394 if (CC != X86::COND_E && CC != X86::COND_NE)
26397 SDValue Cmp = SetCC.getOperand(1);
26398 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26399 !X86::isZeroNode(Cmp.getOperand(1)) ||
26400 !Cmp.getOperand(0).getValueType().isInteger())
26403 SDValue CmpOp0 = Cmp.getOperand(0);
26404 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26405 DAG.getConstant(1, CmpOp0.getValueType()));
26407 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26408 if (CC == X86::COND_NE)
26409 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26410 DL, OtherVal.getValueType(), OtherVal,
26411 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
26412 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26413 DL, OtherVal.getValueType(), OtherVal,
26414 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
26417 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26418 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26419 const X86Subtarget *Subtarget) {
26420 EVT VT = N->getValueType(0);
26421 SDValue Op0 = N->getOperand(0);
26422 SDValue Op1 = N->getOperand(1);
26424 // Try to synthesize horizontal adds from adds of shuffles.
26425 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26426 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26427 isHorizontalBinOp(Op0, Op1, true))
26428 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26430 return OptimizeConditionalInDecrement(N, DAG);
26433 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26434 const X86Subtarget *Subtarget) {
26435 SDValue Op0 = N->getOperand(0);
26436 SDValue Op1 = N->getOperand(1);
26438 // X86 can't encode an immediate LHS of a sub. See if we can push the
26439 // negation into a preceding instruction.
26440 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26441 // If the RHS of the sub is a XOR with one use and a constant, invert the
26442 // immediate. Then add one to the LHS of the sub so we can turn
26443 // X-Y -> X+~Y+1, saving one register.
26444 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26445 isa<ConstantSDNode>(Op1.getOperand(1))) {
26446 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26447 EVT VT = Op0.getValueType();
26448 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26450 DAG.getConstant(~XorC, VT));
26451 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26452 DAG.getConstant(C->getAPIntValue()+1, VT));
26456 // Try to synthesize horizontal adds from adds of shuffles.
26457 EVT VT = N->getValueType(0);
26458 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26459 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26460 isHorizontalBinOp(Op0, Op1, true))
26461 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26463 return OptimizeConditionalInDecrement(N, DAG);
26466 /// performVZEXTCombine - Performs build vector combines
26467 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26468 TargetLowering::DAGCombinerInfo &DCI,
26469 const X86Subtarget *Subtarget) {
26471 MVT VT = N->getSimpleValueType(0);
26472 SDValue Op = N->getOperand(0);
26473 MVT OpVT = Op.getSimpleValueType();
26474 MVT OpEltVT = OpVT.getVectorElementType();
26475 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26477 // (vzext (bitcast (vzext (x)) -> (vzext x)
26479 while (V.getOpcode() == ISD::BITCAST)
26480 V = V.getOperand(0);
26482 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26483 MVT InnerVT = V.getSimpleValueType();
26484 MVT InnerEltVT = InnerVT.getVectorElementType();
26486 // If the element sizes match exactly, we can just do one larger vzext. This
26487 // is always an exact type match as vzext operates on integer types.
26488 if (OpEltVT == InnerEltVT) {
26489 assert(OpVT == InnerVT && "Types must match for vzext!");
26490 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26493 // The only other way we can combine them is if only a single element of the
26494 // inner vzext is used in the input to the outer vzext.
26495 if (InnerEltVT.getSizeInBits() < InputBits)
26498 // In this case, the inner vzext is completely dead because we're going to
26499 // only look at bits inside of the low element. Just do the outer vzext on
26500 // a bitcast of the input to the inner.
26501 return DAG.getNode(X86ISD::VZEXT, DL, VT,
26502 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
26505 // Check if we can bypass extracting and re-inserting an element of an input
26506 // vector. Essentialy:
26507 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26508 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26509 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26510 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26511 SDValue ExtractedV = V.getOperand(0);
26512 SDValue OrigV = ExtractedV.getOperand(0);
26513 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26514 if (ExtractIdx->getZExtValue() == 0) {
26515 MVT OrigVT = OrigV.getSimpleValueType();
26516 // Extract a subvector if necessary...
26517 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26518 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26519 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26520 OrigVT.getVectorNumElements() / Ratio);
26521 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26522 DAG.getIntPtrConstant(0));
26524 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
26525 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26532 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26533 DAGCombinerInfo &DCI) const {
26534 SelectionDAG &DAG = DCI.DAG;
26535 switch (N->getOpcode()) {
26537 case ISD::EXTRACT_VECTOR_ELT:
26538 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26541 case X86ISD::SHRUNKBLEND:
26542 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26543 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
26544 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26545 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26546 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26547 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26548 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26551 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26552 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26553 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26554 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26555 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26556 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26557 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26558 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26559 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26560 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26561 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26563 case X86ISD::FOR: return PerformFORCombine(N, DAG);
26565 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26566 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26567 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26568 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26569 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26570 case ISD::ANY_EXTEND:
26571 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26572 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26573 case ISD::SIGN_EXTEND_INREG:
26574 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26575 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
26576 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26577 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26578 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26579 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26580 case X86ISD::SHUFP: // Handle all target specific shuffles
26581 case X86ISD::PALIGNR:
26582 case X86ISD::UNPCKH:
26583 case X86ISD::UNPCKL:
26584 case X86ISD::MOVHLPS:
26585 case X86ISD::MOVLHPS:
26586 case X86ISD::PSHUFB:
26587 case X86ISD::PSHUFD:
26588 case X86ISD::PSHUFHW:
26589 case X86ISD::PSHUFLW:
26590 case X86ISD::MOVSS:
26591 case X86ISD::MOVSD:
26592 case X86ISD::VPERMILPI:
26593 case X86ISD::VPERM2X128:
26594 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26595 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26596 case ISD::INTRINSIC_WO_CHAIN:
26597 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
26598 case X86ISD::INSERTPS: {
26599 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26600 return PerformINSERTPSCombine(N, DAG, Subtarget);
26603 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
26609 /// isTypeDesirableForOp - Return true if the target has native support for
26610 /// the specified value type and it is 'desirable' to use the type for the
26611 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26612 /// instruction encodings are longer and some i16 instructions are slow.
26613 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26614 if (!isTypeLegal(VT))
26616 if (VT != MVT::i16)
26623 case ISD::SIGN_EXTEND:
26624 case ISD::ZERO_EXTEND:
26625 case ISD::ANY_EXTEND:
26638 /// IsDesirableToPromoteOp - This method query the target whether it is
26639 /// beneficial for dag combiner to promote the specified node. If true, it
26640 /// should return the desired promotion type by reference.
26641 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26642 EVT VT = Op.getValueType();
26643 if (VT != MVT::i16)
26646 bool Promote = false;
26647 bool Commute = false;
26648 switch (Op.getOpcode()) {
26651 LoadSDNode *LD = cast<LoadSDNode>(Op);
26652 // If the non-extending load has a single use and it's not live out, then it
26653 // might be folded.
26654 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26655 Op.hasOneUse()*/) {
26656 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26657 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26658 // The only case where we'd want to promote LOAD (rather then it being
26659 // promoted as an operand is when it's only use is liveout.
26660 if (UI->getOpcode() != ISD::CopyToReg)
26667 case ISD::SIGN_EXTEND:
26668 case ISD::ZERO_EXTEND:
26669 case ISD::ANY_EXTEND:
26674 SDValue N0 = Op.getOperand(0);
26675 // Look out for (store (shl (load), x)).
26676 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26689 SDValue N0 = Op.getOperand(0);
26690 SDValue N1 = Op.getOperand(1);
26691 if (!Commute && MayFoldLoad(N1))
26693 // Avoid disabling potential load folding opportunities.
26694 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26696 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26706 //===----------------------------------------------------------------------===//
26707 // X86 Inline Assembly Support
26708 //===----------------------------------------------------------------------===//
26711 // Helper to match a string separated by whitespace.
26712 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
26713 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
26715 for (unsigned i = 0, e = args.size(); i != e; ++i) {
26716 StringRef piece(*args[i]);
26717 if (!s.startswith(piece)) // Check if the piece matches.
26720 s = s.substr(piece.size());
26721 StringRef::size_type pos = s.find_first_not_of(" \t");
26722 if (pos == 0) // We matched a prefix.
26730 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
26733 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26735 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26736 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26737 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26738 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26740 if (AsmPieces.size() == 3)
26742 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26749 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26750 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26752 std::string AsmStr = IA->getAsmString();
26754 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26755 if (!Ty || Ty->getBitWidth() % 16 != 0)
26758 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26759 SmallVector<StringRef, 4> AsmPieces;
26760 SplitString(AsmStr, AsmPieces, ";\n");
26762 switch (AsmPieces.size()) {
26763 default: return false;
26765 // FIXME: this should verify that we are targeting a 486 or better. If not,
26766 // we will turn this bswap into something that will be lowered to logical
26767 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26768 // lower so don't worry about this.
26770 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
26771 matchAsm(AsmPieces[0], "bswapl", "$0") ||
26772 matchAsm(AsmPieces[0], "bswapq", "$0") ||
26773 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
26774 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
26775 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
26776 // No need to check constraints, nothing other than the equivalent of
26777 // "=r,0" would be valid here.
26778 return IntrinsicLowering::LowerToByteSwap(CI);
26781 // rorw $$8, ${0:w} --> llvm.bswap.i16
26782 if (CI->getType()->isIntegerTy(16) &&
26783 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26784 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
26785 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
26787 const std::string &ConstraintsStr = IA->getConstraintString();
26788 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26789 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26790 if (clobbersFlagRegisters(AsmPieces))
26791 return IntrinsicLowering::LowerToByteSwap(CI);
26795 if (CI->getType()->isIntegerTy(32) &&
26796 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26797 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
26798 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
26799 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
26801 const std::string &ConstraintsStr = IA->getConstraintString();
26802 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26803 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26804 if (clobbersFlagRegisters(AsmPieces))
26805 return IntrinsicLowering::LowerToByteSwap(CI);
26808 if (CI->getType()->isIntegerTy(64)) {
26809 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26810 if (Constraints.size() >= 2 &&
26811 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26812 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26813 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26814 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
26815 matchAsm(AsmPieces[1], "bswap", "%edx") &&
26816 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
26817 return IntrinsicLowering::LowerToByteSwap(CI);
26825 /// getConstraintType - Given a constraint letter, return the type of
26826 /// constraint it is for this target.
26827 X86TargetLowering::ConstraintType
26828 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
26829 if (Constraint.size() == 1) {
26830 switch (Constraint[0]) {
26841 return C_RegisterClass;
26865 return TargetLowering::getConstraintType(Constraint);
26868 /// Examine constraint type and operand type and determine a weight value.
26869 /// This object must already have been set up with the operand type
26870 /// and the current alternative constraint selected.
26871 TargetLowering::ConstraintWeight
26872 X86TargetLowering::getSingleConstraintMatchWeight(
26873 AsmOperandInfo &info, const char *constraint) const {
26874 ConstraintWeight weight = CW_Invalid;
26875 Value *CallOperandVal = info.CallOperandVal;
26876 // If we don't have a value, we can't do a match,
26877 // but allow it at the lowest weight.
26878 if (!CallOperandVal)
26880 Type *type = CallOperandVal->getType();
26881 // Look at the constraint type.
26882 switch (*constraint) {
26884 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26895 if (CallOperandVal->getType()->isIntegerTy())
26896 weight = CW_SpecificReg;
26901 if (type->isFloatingPointTy())
26902 weight = CW_SpecificReg;
26905 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26906 weight = CW_SpecificReg;
26910 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26911 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26912 weight = CW_Register;
26915 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26916 if (C->getZExtValue() <= 31)
26917 weight = CW_Constant;
26921 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26922 if (C->getZExtValue() <= 63)
26923 weight = CW_Constant;
26927 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26928 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26929 weight = CW_Constant;
26933 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26934 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26935 weight = CW_Constant;
26939 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26940 if (C->getZExtValue() <= 3)
26941 weight = CW_Constant;
26945 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26946 if (C->getZExtValue() <= 0xff)
26947 weight = CW_Constant;
26952 if (dyn_cast<ConstantFP>(CallOperandVal)) {
26953 weight = CW_Constant;
26957 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26958 if ((C->getSExtValue() >= -0x80000000LL) &&
26959 (C->getSExtValue() <= 0x7fffffffLL))
26960 weight = CW_Constant;
26964 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26965 if (C->getZExtValue() <= 0xffffffff)
26966 weight = CW_Constant;
26973 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26974 /// with another that has more specific requirements based on the type of the
26975 /// corresponding operand.
26976 const char *X86TargetLowering::
26977 LowerXConstraint(EVT ConstraintVT) const {
26978 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26979 // 'f' like normal targets.
26980 if (ConstraintVT.isFloatingPoint()) {
26981 if (Subtarget->hasSSE2())
26983 if (Subtarget->hasSSE1())
26987 return TargetLowering::LowerXConstraint(ConstraintVT);
26990 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26991 /// vector. If it is invalid, don't add anything to Ops.
26992 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26993 std::string &Constraint,
26994 std::vector<SDValue>&Ops,
26995 SelectionDAG &DAG) const {
26998 // Only support length 1 constraints for now.
26999 if (Constraint.length() > 1) return;
27001 char ConstraintLetter = Constraint[0];
27002 switch (ConstraintLetter) {
27005 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27006 if (C->getZExtValue() <= 31) {
27007 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27013 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27014 if (C->getZExtValue() <= 63) {
27015 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27021 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27022 if (isInt<8>(C->getSExtValue())) {
27023 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27030 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27031 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27032 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
27038 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27039 if (C->getZExtValue() <= 3) {
27040 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27046 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27047 if (C->getZExtValue() <= 255) {
27048 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27054 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27055 if (C->getZExtValue() <= 127) {
27056 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27062 // 32-bit signed value
27063 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27064 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27065 C->getSExtValue())) {
27066 // Widen to 64 bits here to get it sign extended.
27067 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
27070 // FIXME gcc accepts some relocatable values here too, but only in certain
27071 // memory models; it's complicated.
27076 // 32-bit unsigned value
27077 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27078 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27079 C->getZExtValue())) {
27080 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
27084 // FIXME gcc accepts some relocatable values here too, but only in certain
27085 // memory models; it's complicated.
27089 // Literal immediates are always ok.
27090 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27091 // Widen to 64 bits here to get it sign extended.
27092 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
27096 // In any sort of PIC mode addresses need to be computed at runtime by
27097 // adding in a register or some sort of table lookup. These can't
27098 // be used as immediates.
27099 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27102 // If we are in non-pic codegen mode, we allow the address of a global (with
27103 // an optional displacement) to be used with 'i'.
27104 GlobalAddressSDNode *GA = nullptr;
27105 int64_t Offset = 0;
27107 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27109 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27110 Offset += GA->getOffset();
27112 } else if (Op.getOpcode() == ISD::ADD) {
27113 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27114 Offset += C->getZExtValue();
27115 Op = Op.getOperand(0);
27118 } else if (Op.getOpcode() == ISD::SUB) {
27119 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27120 Offset += -C->getZExtValue();
27121 Op = Op.getOperand(0);
27126 // Otherwise, this isn't something we can handle, reject it.
27130 const GlobalValue *GV = GA->getGlobal();
27131 // If we require an extra load to get this address, as in PIC mode, we
27132 // can't accept it.
27133 if (isGlobalStubReference(
27134 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27137 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27138 GA->getValueType(0), Offset);
27143 if (Result.getNode()) {
27144 Ops.push_back(Result);
27147 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27150 std::pair<unsigned, const TargetRegisterClass*>
27151 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
27153 // First, see if this is a constraint that directly corresponds to an LLVM
27155 if (Constraint.size() == 1) {
27156 // GCC Constraint Letters
27157 switch (Constraint[0]) {
27159 // TODO: Slight differences here in allocation order and leaving
27160 // RIP in the class. Do they matter any more here than they do
27161 // in the normal allocation?
27162 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27163 if (Subtarget->is64Bit()) {
27164 if (VT == MVT::i32 || VT == MVT::f32)
27165 return std::make_pair(0U, &X86::GR32RegClass);
27166 if (VT == MVT::i16)
27167 return std::make_pair(0U, &X86::GR16RegClass);
27168 if (VT == MVT::i8 || VT == MVT::i1)
27169 return std::make_pair(0U, &X86::GR8RegClass);
27170 if (VT == MVT::i64 || VT == MVT::f64)
27171 return std::make_pair(0U, &X86::GR64RegClass);
27174 // 32-bit fallthrough
27175 case 'Q': // Q_REGS
27176 if (VT == MVT::i32 || VT == MVT::f32)
27177 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27178 if (VT == MVT::i16)
27179 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27180 if (VT == MVT::i8 || VT == MVT::i1)
27181 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27182 if (VT == MVT::i64)
27183 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27185 case 'r': // GENERAL_REGS
27186 case 'l': // INDEX_REGS
27187 if (VT == MVT::i8 || VT == MVT::i1)
27188 return std::make_pair(0U, &X86::GR8RegClass);
27189 if (VT == MVT::i16)
27190 return std::make_pair(0U, &X86::GR16RegClass);
27191 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27192 return std::make_pair(0U, &X86::GR32RegClass);
27193 return std::make_pair(0U, &X86::GR64RegClass);
27194 case 'R': // LEGACY_REGS
27195 if (VT == MVT::i8 || VT == MVT::i1)
27196 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27197 if (VT == MVT::i16)
27198 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27199 if (VT == MVT::i32 || !Subtarget->is64Bit())
27200 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27201 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27202 case 'f': // FP Stack registers.
27203 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27204 // value to the correct fpstack register class.
27205 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27206 return std::make_pair(0U, &X86::RFP32RegClass);
27207 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27208 return std::make_pair(0U, &X86::RFP64RegClass);
27209 return std::make_pair(0U, &X86::RFP80RegClass);
27210 case 'y': // MMX_REGS if MMX allowed.
27211 if (!Subtarget->hasMMX()) break;
27212 return std::make_pair(0U, &X86::VR64RegClass);
27213 case 'Y': // SSE_REGS if SSE2 allowed
27214 if (!Subtarget->hasSSE2()) break;
27216 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27217 if (!Subtarget->hasSSE1()) break;
27219 switch (VT.SimpleTy) {
27221 // Scalar SSE types.
27224 return std::make_pair(0U, &X86::FR32RegClass);
27227 return std::make_pair(0U, &X86::FR64RegClass);
27235 return std::make_pair(0U, &X86::VR128RegClass);
27243 return std::make_pair(0U, &X86::VR256RegClass);
27248 return std::make_pair(0U, &X86::VR512RegClass);
27254 // Use the default implementation in TargetLowering to convert the register
27255 // constraint into a member of a register class.
27256 std::pair<unsigned, const TargetRegisterClass*> Res;
27257 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
27259 // Not found as a standard register?
27261 // Map st(0) -> st(7) -> ST0
27262 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27263 tolower(Constraint[1]) == 's' &&
27264 tolower(Constraint[2]) == 't' &&
27265 Constraint[3] == '(' &&
27266 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27267 Constraint[5] == ')' &&
27268 Constraint[6] == '}') {
27270 Res.first = X86::FP0+Constraint[4]-'0';
27271 Res.second = &X86::RFP80RegClass;
27275 // GCC allows "st(0)" to be called just plain "st".
27276 if (StringRef("{st}").equals_lower(Constraint)) {
27277 Res.first = X86::FP0;
27278 Res.second = &X86::RFP80RegClass;
27283 if (StringRef("{flags}").equals_lower(Constraint)) {
27284 Res.first = X86::EFLAGS;
27285 Res.second = &X86::CCRRegClass;
27289 // 'A' means EAX + EDX.
27290 if (Constraint == "A") {
27291 Res.first = X86::EAX;
27292 Res.second = &X86::GR32_ADRegClass;
27298 // Otherwise, check to see if this is a register class of the wrong value
27299 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27300 // turn into {ax},{dx}.
27301 if (Res.second->hasType(VT))
27302 return Res; // Correct type already, nothing to do.
27304 // All of the single-register GCC register classes map their values onto
27305 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
27306 // really want an 8-bit or 32-bit register, map to the appropriate register
27307 // class and return the appropriate register.
27308 if (Res.second == &X86::GR16RegClass) {
27309 if (VT == MVT::i8 || VT == MVT::i1) {
27310 unsigned DestReg = 0;
27311 switch (Res.first) {
27313 case X86::AX: DestReg = X86::AL; break;
27314 case X86::DX: DestReg = X86::DL; break;
27315 case X86::CX: DestReg = X86::CL; break;
27316 case X86::BX: DestReg = X86::BL; break;
27319 Res.first = DestReg;
27320 Res.second = &X86::GR8RegClass;
27322 } else if (VT == MVT::i32 || VT == MVT::f32) {
27323 unsigned DestReg = 0;
27324 switch (Res.first) {
27326 case X86::AX: DestReg = X86::EAX; break;
27327 case X86::DX: DestReg = X86::EDX; break;
27328 case X86::CX: DestReg = X86::ECX; break;
27329 case X86::BX: DestReg = X86::EBX; break;
27330 case X86::SI: DestReg = X86::ESI; break;
27331 case X86::DI: DestReg = X86::EDI; break;
27332 case X86::BP: DestReg = X86::EBP; break;
27333 case X86::SP: DestReg = X86::ESP; break;
27336 Res.first = DestReg;
27337 Res.second = &X86::GR32RegClass;
27339 } else if (VT == MVT::i64 || VT == MVT::f64) {
27340 unsigned DestReg = 0;
27341 switch (Res.first) {
27343 case X86::AX: DestReg = X86::RAX; break;
27344 case X86::DX: DestReg = X86::RDX; break;
27345 case X86::CX: DestReg = X86::RCX; break;
27346 case X86::BX: DestReg = X86::RBX; break;
27347 case X86::SI: DestReg = X86::RSI; break;
27348 case X86::DI: DestReg = X86::RDI; break;
27349 case X86::BP: DestReg = X86::RBP; break;
27350 case X86::SP: DestReg = X86::RSP; break;
27353 Res.first = DestReg;
27354 Res.second = &X86::GR64RegClass;
27357 } else if (Res.second == &X86::FR32RegClass ||
27358 Res.second == &X86::FR64RegClass ||
27359 Res.second == &X86::VR128RegClass ||
27360 Res.second == &X86::VR256RegClass ||
27361 Res.second == &X86::FR32XRegClass ||
27362 Res.second == &X86::FR64XRegClass ||
27363 Res.second == &X86::VR128XRegClass ||
27364 Res.second == &X86::VR256XRegClass ||
27365 Res.second == &X86::VR512RegClass) {
27366 // Handle references to XMM physical registers that got mapped into the
27367 // wrong class. This can happen with constraints like {xmm0} where the
27368 // target independent register mapper will just pick the first match it can
27369 // find, ignoring the required type.
27371 if (VT == MVT::f32 || VT == MVT::i32)
27372 Res.second = &X86::FR32RegClass;
27373 else if (VT == MVT::f64 || VT == MVT::i64)
27374 Res.second = &X86::FR64RegClass;
27375 else if (X86::VR128RegClass.hasType(VT))
27376 Res.second = &X86::VR128RegClass;
27377 else if (X86::VR256RegClass.hasType(VT))
27378 Res.second = &X86::VR256RegClass;
27379 else if (X86::VR512RegClass.hasType(VT))
27380 Res.second = &X86::VR512RegClass;
27386 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
27388 // Scaling factors are not free at all.
27389 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27390 // will take 2 allocations in the out of order engine instead of 1
27391 // for plain addressing mode, i.e. inst (reg1).
27393 // vaddps (%rsi,%drx), %ymm0, %ymm1
27394 // Requires two allocations (one for the load, one for the computation)
27396 // vaddps (%rsi), %ymm0, %ymm1
27397 // Requires just 1 allocation, i.e., freeing allocations for other operations
27398 // and having less micro operations to execute.
27400 // For some X86 architectures, this is even worse because for instance for
27401 // stores, the complex addressing mode forces the instruction to use the
27402 // "load" ports instead of the dedicated "store" port.
27403 // E.g., on Haswell:
27404 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27405 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27406 if (isLegalAddressingMode(AM, Ty))
27407 // Scale represents reg2 * scale, thus account for 1
27408 // as soon as we use a second register.
27409 return AM.Scale != 0;
27413 bool X86TargetLowering::isTargetFTOL() const {
27414 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();