1 //===- InstCombineCalls.cpp -----------------------------------------------===//
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 implements the visitCall and visitInvoke functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/IR/CallSite.h"
19 #include "llvm/IR/Dominators.h"
20 #include "llvm/IR/PatternMatch.h"
21 #include "llvm/IR/Statepoint.h"
22 #include "llvm/Transforms/Utils/BuildLibCalls.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
26 using namespace PatternMatch;
28 #define DEBUG_TYPE "instcombine"
30 STATISTIC(NumSimplified, "Number of library calls simplified");
32 /// getPromotedType - Return the specified type promoted as it would be to pass
33 /// though a va_arg area.
34 static Type *getPromotedType(Type *Ty) {
35 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
36 if (ITy->getBitWidth() < 32)
37 return Type::getInt32Ty(Ty->getContext());
42 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
43 /// single scalar element, like {{{type}}} or [1 x type], return type.
44 static Type *reduceToSingleValueType(Type *T) {
45 while (!T->isSingleValueType()) {
46 if (StructType *STy = dyn_cast<StructType>(T)) {
47 if (STy->getNumElements() == 1)
48 T = STy->getElementType(0);
51 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
52 if (ATy->getNumElements() == 1)
53 T = ATy->getElementType();
63 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
64 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, MI, AC, DT);
65 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, AC, DT);
66 unsigned MinAlign = std::min(DstAlign, SrcAlign);
67 unsigned CopyAlign = MI->getAlignment();
69 if (CopyAlign < MinAlign) {
70 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), MinAlign, false));
74 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
76 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
77 if (!MemOpLength) return nullptr;
79 // Source and destination pointer types are always "i8*" for intrinsic. See
80 // if the size is something we can handle with a single primitive load/store.
81 // A single load+store correctly handles overlapping memory in the memmove
83 uint64_t Size = MemOpLength->getLimitedValue();
84 assert(Size && "0-sized memory transferring should be removed already.");
86 if (Size > 8 || (Size&(Size-1)))
87 return nullptr; // If not 1/2/4/8 bytes, exit.
89 // Use an integer load+store unless we can find something better.
91 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
93 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
95 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
96 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
97 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
99 // Memcpy forces the use of i8* for the source and destination. That means
100 // that if you're using memcpy to move one double around, you'll get a cast
101 // from double* to i8*. We'd much rather use a double load+store rather than
102 // an i64 load+store, here because this improves the odds that the source or
103 // dest address will be promotable. See if we can find a better type than the
105 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
106 MDNode *CopyMD = nullptr;
107 if (StrippedDest != MI->getArgOperand(0)) {
108 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
110 if (SrcETy->isSized() && DL.getTypeStoreSize(SrcETy) == Size) {
111 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
112 // down through these levels if so.
113 SrcETy = reduceToSingleValueType(SrcETy);
115 if (SrcETy->isSingleValueType()) {
116 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
117 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
119 // If the memcpy has metadata describing the members, see if we can
120 // get the TBAA tag describing our copy.
121 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
122 if (M->getNumOperands() == 3 && M->getOperand(0) &&
123 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
124 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() &&
126 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
127 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
129 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
130 CopyMD = cast<MDNode>(M->getOperand(2));
136 // If the memcpy/memmove provides better alignment info than we can
138 SrcAlign = std::max(SrcAlign, CopyAlign);
139 DstAlign = std::max(DstAlign, CopyAlign);
141 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
142 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
143 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
144 L->setAlignment(SrcAlign);
146 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
147 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
148 S->setAlignment(DstAlign);
150 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
152 // Set the size of the copy to 0, it will be deleted on the next iteration.
153 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
157 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
158 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, AC, DT);
159 if (MI->getAlignment() < Alignment) {
160 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
165 // Extract the length and alignment and fill if they are constant.
166 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
167 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
168 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
170 uint64_t Len = LenC->getLimitedValue();
171 Alignment = MI->getAlignment();
172 assert(Len && "0-sized memory setting should be removed already.");
174 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
175 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
176 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
178 Value *Dest = MI->getDest();
179 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
180 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
181 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
183 // Alignment 0 is identity for alignment 1 for memset, but not store.
184 if (Alignment == 0) Alignment = 1;
186 // Extract the fill value and store.
187 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
188 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
190 S->setAlignment(Alignment);
192 // Set the size of the copy to 0, it will be deleted on the next iteration.
193 MI->setLength(Constant::getNullValue(LenC->getType()));
200 static Value *SimplifyX86immshift(const IntrinsicInst &II,
201 InstCombiner::BuilderTy &Builder) {
202 bool LogicalShift = false;
203 bool ShiftLeft = false;
205 switch (II.getIntrinsicID()) {
208 case Intrinsic::x86_sse2_psra_d:
209 case Intrinsic::x86_sse2_psra_w:
210 case Intrinsic::x86_sse2_psrai_d:
211 case Intrinsic::x86_sse2_psrai_w:
212 case Intrinsic::x86_avx2_psra_d:
213 case Intrinsic::x86_avx2_psra_w:
214 case Intrinsic::x86_avx2_psrai_d:
215 case Intrinsic::x86_avx2_psrai_w:
216 LogicalShift = false; ShiftLeft = false;
218 case Intrinsic::x86_sse2_psrl_d:
219 case Intrinsic::x86_sse2_psrl_q:
220 case Intrinsic::x86_sse2_psrl_w:
221 case Intrinsic::x86_sse2_psrli_d:
222 case Intrinsic::x86_sse2_psrli_q:
223 case Intrinsic::x86_sse2_psrli_w:
224 case Intrinsic::x86_avx2_psrl_d:
225 case Intrinsic::x86_avx2_psrl_q:
226 case Intrinsic::x86_avx2_psrl_w:
227 case Intrinsic::x86_avx2_psrli_d:
228 case Intrinsic::x86_avx2_psrli_q:
229 case Intrinsic::x86_avx2_psrli_w:
230 LogicalShift = true; ShiftLeft = false;
232 case Intrinsic::x86_sse2_psll_d:
233 case Intrinsic::x86_sse2_psll_q:
234 case Intrinsic::x86_sse2_psll_w:
235 case Intrinsic::x86_sse2_pslli_d:
236 case Intrinsic::x86_sse2_pslli_q:
237 case Intrinsic::x86_sse2_pslli_w:
238 case Intrinsic::x86_avx2_psll_d:
239 case Intrinsic::x86_avx2_psll_q:
240 case Intrinsic::x86_avx2_psll_w:
241 case Intrinsic::x86_avx2_pslli_d:
242 case Intrinsic::x86_avx2_pslli_q:
243 case Intrinsic::x86_avx2_pslli_w:
244 LogicalShift = true; ShiftLeft = true;
247 assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
249 // Simplify if count is constant.
250 auto Arg1 = II.getArgOperand(1);
251 auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1);
252 auto CDV = dyn_cast<ConstantDataVector>(Arg1);
253 auto CInt = dyn_cast<ConstantInt>(Arg1);
254 if (!CAZ && !CDV && !CInt)
259 // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
260 // operand to compute the shift amount.
261 auto VT = cast<VectorType>(CDV->getType());
262 unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits();
263 assert((64 % BitWidth) == 0 && "Unexpected packed shift size");
264 unsigned NumSubElts = 64 / BitWidth;
266 // Concatenate the sub-elements to create the 64-bit value.
267 for (unsigned i = 0; i != NumSubElts; ++i) {
268 unsigned SubEltIdx = (NumSubElts - 1) - i;
269 auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
270 Count = Count.shl(BitWidth);
271 Count |= SubElt->getValue().zextOrTrunc(64);
275 Count = CInt->getValue();
277 auto Vec = II.getArgOperand(0);
278 auto VT = cast<VectorType>(Vec->getType());
279 auto SVT = VT->getElementType();
280 unsigned VWidth = VT->getNumElements();
281 unsigned BitWidth = SVT->getPrimitiveSizeInBits();
283 // If shift-by-zero then just return the original value.
287 // Handle cases when Shift >= BitWidth.
288 if (Count.uge(BitWidth)) {
289 // If LogicalShift - just return zero.
291 return ConstantAggregateZero::get(VT);
293 // If ArithmeticShift - clamp Shift to (BitWidth - 1).
294 Count = APInt(64, BitWidth - 1);
297 // Get a constant vector of the same type as the first operand.
298 auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
299 auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
302 return Builder.CreateShl(Vec, ShiftVec);
305 return Builder.CreateLShr(Vec, ShiftVec);
307 return Builder.CreateAShr(Vec, ShiftVec);
310 static Value *SimplifyX86extend(const IntrinsicInst &II,
311 InstCombiner::BuilderTy &Builder,
313 VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType());
314 VectorType *DstTy = cast<VectorType>(II.getType());
315 unsigned NumDstElts = DstTy->getNumElements();
317 // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
318 SmallVector<int, 8> ShuffleMask;
319 for (int i = 0; i != (int)NumDstElts; ++i)
320 ShuffleMask.push_back(i);
322 Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0),
323 UndefValue::get(SrcTy), ShuffleMask);
324 return SignExtend ? Builder.CreateSExt(SV, DstTy)
325 : Builder.CreateZExt(SV, DstTy);
328 static Value *SimplifyX86insertps(const IntrinsicInst &II,
329 InstCombiner::BuilderTy &Builder) {
330 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
331 VectorType *VecTy = cast<VectorType>(II.getType());
332 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
334 // The immediate permute control byte looks like this:
335 // [3:0] - zero mask for each 32-bit lane
336 // [5:4] - select one 32-bit destination lane
337 // [7:6] - select one 32-bit source lane
339 uint8_t Imm = CInt->getZExtValue();
340 uint8_t ZMask = Imm & 0xf;
341 uint8_t DestLane = (Imm >> 4) & 0x3;
342 uint8_t SourceLane = (Imm >> 6) & 0x3;
344 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
346 // If all zero mask bits are set, this was just a weird way to
347 // generate a zero vector.
351 // Initialize by passing all of the first source bits through.
352 int ShuffleMask[4] = { 0, 1, 2, 3 };
354 // We may replace the second operand with the zero vector.
355 Value *V1 = II.getArgOperand(1);
358 // If the zero mask is being used with a single input or the zero mask
359 // overrides the destination lane, this is a shuffle with the zero vector.
360 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
361 (ZMask & (1 << DestLane))) {
363 // We may still move 32-bits of the first source vector from one lane
365 ShuffleMask[DestLane] = SourceLane;
366 // The zero mask may override the previous insert operation.
367 for (unsigned i = 0; i < 4; ++i)
368 if ((ZMask >> i) & 0x1)
369 ShuffleMask[i] = i + 4;
371 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
375 // Replace the selected destination lane with the selected source lane.
376 ShuffleMask[DestLane] = SourceLane + 4;
379 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
384 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
385 /// source vectors, unless a zero bit is set. If a zero bit is set,
386 /// then ignore that half of the mask and clear that half of the vector.
387 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
388 InstCombiner::BuilderTy &Builder) {
389 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
390 VectorType *VecTy = cast<VectorType>(II.getType());
391 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
393 // The immediate permute control byte looks like this:
394 // [1:0] - select 128 bits from sources for low half of destination
396 // [3] - zero low half of destination
397 // [5:4] - select 128 bits from sources for high half of destination
399 // [7] - zero high half of destination
401 uint8_t Imm = CInt->getZExtValue();
403 bool LowHalfZero = Imm & 0x08;
404 bool HighHalfZero = Imm & 0x80;
406 // If both zero mask bits are set, this was just a weird way to
407 // generate a zero vector.
408 if (LowHalfZero && HighHalfZero)
411 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
412 unsigned NumElts = VecTy->getNumElements();
413 unsigned HalfSize = NumElts / 2;
414 SmallVector<int, 8> ShuffleMask(NumElts);
416 // The high bit of the selection field chooses the 1st or 2nd operand.
417 bool LowInputSelect = Imm & 0x02;
418 bool HighInputSelect = Imm & 0x20;
420 // The low bit of the selection field chooses the low or high half
421 // of the selected operand.
422 bool LowHalfSelect = Imm & 0x01;
423 bool HighHalfSelect = Imm & 0x10;
425 // Determine which operand(s) are actually in use for this instruction.
426 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
427 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
429 // If needed, replace operands based on zero mask.
430 V0 = LowHalfZero ? ZeroVector : V0;
431 V1 = HighHalfZero ? ZeroVector : V1;
433 // Permute low half of result.
434 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
435 for (unsigned i = 0; i < HalfSize; ++i)
436 ShuffleMask[i] = StartIndex + i;
438 // Permute high half of result.
439 StartIndex = HighHalfSelect ? HalfSize : 0;
440 StartIndex += NumElts;
441 for (unsigned i = 0; i < HalfSize; ++i)
442 ShuffleMask[i + HalfSize] = StartIndex + i;
444 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
449 /// Decode XOP integer vector comparison intrinsics.
450 static Value *SimplifyX86vpcom(const IntrinsicInst &II,
451 InstCombiner::BuilderTy &Builder, bool IsSigned) {
452 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
453 uint64_t Imm = CInt->getZExtValue() & 0x7;
454 VectorType *VecTy = cast<VectorType>(II.getType());
455 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
459 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
462 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
465 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
468 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
471 Pred = ICmpInst::ICMP_EQ; break;
473 Pred = ICmpInst::ICMP_NE; break;
475 return ConstantInt::getSigned(VecTy, 0); // FALSE
477 return ConstantInt::getSigned(VecTy, -1); // TRUE
480 if (Value *Cmp = Builder.CreateICmp(Pred, II.getArgOperand(0), II.getArgOperand(1)))
481 return Builder.CreateSExtOrTrunc(Cmp, VecTy);
486 /// visitCallInst - CallInst simplification. This mostly only handles folding
487 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
488 /// the heavy lifting.
490 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
491 auto Args = CI.arg_operands();
492 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
494 return ReplaceInstUsesWith(CI, V);
496 if (isFreeCall(&CI, TLI))
497 return visitFree(CI);
499 // If the caller function is nounwind, mark the call as nounwind, even if the
501 if (CI.getParent()->getParent()->doesNotThrow() &&
502 !CI.doesNotThrow()) {
503 CI.setDoesNotThrow();
507 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
508 if (!II) return visitCallSite(&CI);
510 // Intrinsics cannot occur in an invoke, so handle them here instead of in
512 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
513 bool Changed = false;
515 // memmove/cpy/set of zero bytes is a noop.
516 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
517 if (NumBytes->isNullValue())
518 return EraseInstFromFunction(CI);
520 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
521 if (CI->getZExtValue() == 1) {
522 // Replace the instruction with just byte operations. We would
523 // transform other cases to loads/stores, but we don't know if
524 // alignment is sufficient.
528 // No other transformations apply to volatile transfers.
529 if (MI->isVolatile())
532 // If we have a memmove and the source operation is a constant global,
533 // then the source and dest pointers can't alias, so we can change this
534 // into a call to memcpy.
535 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
536 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
537 if (GVSrc->isConstant()) {
538 Module *M = CI.getParent()->getParent()->getParent();
539 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
540 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
541 CI.getArgOperand(1)->getType(),
542 CI.getArgOperand(2)->getType() };
543 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
548 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
549 // memmove(x,x,size) -> noop.
550 if (MTI->getSource() == MTI->getDest())
551 return EraseInstFromFunction(CI);
554 // If we can determine a pointer alignment that is bigger than currently
555 // set, update the alignment.
556 if (isa<MemTransferInst>(MI)) {
557 if (Instruction *I = SimplifyMemTransfer(MI))
559 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
560 if (Instruction *I = SimplifyMemSet(MSI))
564 if (Changed) return II;
567 auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width, unsigned DemandedWidth)
569 APInt UndefElts(Width, 0);
570 APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
571 return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
574 switch (II->getIntrinsicID()) {
576 case Intrinsic::objectsize: {
578 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
579 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
582 case Intrinsic::bswap: {
583 Value *IIOperand = II->getArgOperand(0);
586 // bswap(bswap(x)) -> x
587 if (match(IIOperand, m_BSwap(m_Value(X))))
588 return ReplaceInstUsesWith(CI, X);
590 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
591 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
592 unsigned C = X->getType()->getPrimitiveSizeInBits() -
593 IIOperand->getType()->getPrimitiveSizeInBits();
594 Value *CV = ConstantInt::get(X->getType(), C);
595 Value *V = Builder->CreateLShr(X, CV);
596 return new TruncInst(V, IIOperand->getType());
601 case Intrinsic::powi:
602 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
605 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
608 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
609 // powi(x, -1) -> 1/x
610 if (Power->isAllOnesValue())
611 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
612 II->getArgOperand(0));
615 case Intrinsic::cttz: {
616 // If all bits below the first known one are known zero,
617 // this value is constant.
618 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
619 // FIXME: Try to simplify vectors of integers.
621 uint32_t BitWidth = IT->getBitWidth();
622 APInt KnownZero(BitWidth, 0);
623 APInt KnownOne(BitWidth, 0);
624 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
625 unsigned TrailingZeros = KnownOne.countTrailingZeros();
626 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
627 if ((Mask & KnownZero) == Mask)
628 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
629 APInt(BitWidth, TrailingZeros)));
633 case Intrinsic::ctlz: {
634 // If all bits above the first known one are known zero,
635 // this value is constant.
636 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
637 // FIXME: Try to simplify vectors of integers.
639 uint32_t BitWidth = IT->getBitWidth();
640 APInt KnownZero(BitWidth, 0);
641 APInt KnownOne(BitWidth, 0);
642 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
643 unsigned LeadingZeros = KnownOne.countLeadingZeros();
644 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
645 if ((Mask & KnownZero) == Mask)
646 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
647 APInt(BitWidth, LeadingZeros)));
652 case Intrinsic::uadd_with_overflow:
653 case Intrinsic::sadd_with_overflow:
654 case Intrinsic::umul_with_overflow:
655 case Intrinsic::smul_with_overflow:
656 if (isa<Constant>(II->getArgOperand(0)) &&
657 !isa<Constant>(II->getArgOperand(1))) {
658 // Canonicalize constants into the RHS.
659 Value *LHS = II->getArgOperand(0);
660 II->setArgOperand(0, II->getArgOperand(1));
661 II->setArgOperand(1, LHS);
666 case Intrinsic::usub_with_overflow:
667 case Intrinsic::ssub_with_overflow: {
668 OverflowCheckFlavor OCF =
669 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
670 assert(OCF != OCF_INVALID && "unexpected!");
672 Value *OperationResult = nullptr;
673 Constant *OverflowResult = nullptr;
674 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
675 *II, OperationResult, OverflowResult))
676 return CreateOverflowTuple(II, OperationResult, OverflowResult);
681 case Intrinsic::minnum:
682 case Intrinsic::maxnum: {
683 Value *Arg0 = II->getArgOperand(0);
684 Value *Arg1 = II->getArgOperand(1);
688 return ReplaceInstUsesWith(CI, Arg0);
690 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
691 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
693 // Canonicalize constants into the RHS.
695 II->setArgOperand(0, Arg1);
696 II->setArgOperand(1, Arg0);
701 if (C1 && C1->isNaN())
702 return ReplaceInstUsesWith(CI, Arg0);
704 // This is the value because if undef were NaN, we would return the other
705 // value and cannot return a NaN unless both operands are.
707 // fmin(undef, x) -> x
708 if (isa<UndefValue>(Arg0))
709 return ReplaceInstUsesWith(CI, Arg1);
711 // fmin(x, undef) -> x
712 if (isa<UndefValue>(Arg1))
713 return ReplaceInstUsesWith(CI, Arg0);
717 if (II->getIntrinsicID() == Intrinsic::minnum) {
718 // fmin(x, fmin(x, y)) -> fmin(x, y)
719 // fmin(y, fmin(x, y)) -> fmin(x, y)
720 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
721 if (Arg0 == X || Arg0 == Y)
722 return ReplaceInstUsesWith(CI, Arg1);
725 // fmin(fmin(x, y), x) -> fmin(x, y)
726 // fmin(fmin(x, y), y) -> fmin(x, y)
727 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
728 if (Arg1 == X || Arg1 == Y)
729 return ReplaceInstUsesWith(CI, Arg0);
732 // TODO: fmin(nnan x, inf) -> x
733 // TODO: fmin(nnan ninf x, flt_max) -> x
734 if (C1 && C1->isInfinity()) {
735 // fmin(x, -inf) -> -inf
736 if (C1->isNegative())
737 return ReplaceInstUsesWith(CI, Arg1);
740 assert(II->getIntrinsicID() == Intrinsic::maxnum);
741 // fmax(x, fmax(x, y)) -> fmax(x, y)
742 // fmax(y, fmax(x, y)) -> fmax(x, y)
743 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
744 if (Arg0 == X || Arg0 == Y)
745 return ReplaceInstUsesWith(CI, Arg1);
748 // fmax(fmax(x, y), x) -> fmax(x, y)
749 // fmax(fmax(x, y), y) -> fmax(x, y)
750 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
751 if (Arg1 == X || Arg1 == Y)
752 return ReplaceInstUsesWith(CI, Arg0);
755 // TODO: fmax(nnan x, -inf) -> x
756 // TODO: fmax(nnan ninf x, -flt_max) -> x
757 if (C1 && C1->isInfinity()) {
758 // fmax(x, inf) -> inf
759 if (!C1->isNegative())
760 return ReplaceInstUsesWith(CI, Arg1);
765 case Intrinsic::ppc_altivec_lvx:
766 case Intrinsic::ppc_altivec_lvxl:
767 // Turn PPC lvx -> load if the pointer is known aligned.
768 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
770 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
771 PointerType::getUnqual(II->getType()));
772 return new LoadInst(Ptr);
775 case Intrinsic::ppc_vsx_lxvw4x:
776 case Intrinsic::ppc_vsx_lxvd2x: {
777 // Turn PPC VSX loads into normal loads.
778 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
779 PointerType::getUnqual(II->getType()));
780 return new LoadInst(Ptr, Twine(""), false, 1);
782 case Intrinsic::ppc_altivec_stvx:
783 case Intrinsic::ppc_altivec_stvxl:
784 // Turn stvx -> store if the pointer is known aligned.
785 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
788 PointerType::getUnqual(II->getArgOperand(0)->getType());
789 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
790 return new StoreInst(II->getArgOperand(0), Ptr);
793 case Intrinsic::ppc_vsx_stxvw4x:
794 case Intrinsic::ppc_vsx_stxvd2x: {
795 // Turn PPC VSX stores into normal stores.
796 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
797 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
798 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
800 case Intrinsic::ppc_qpx_qvlfs:
801 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
802 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
804 Type *VTy = VectorType::get(Builder->getFloatTy(),
805 II->getType()->getVectorNumElements());
806 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
807 PointerType::getUnqual(VTy));
808 Value *Load = Builder->CreateLoad(Ptr);
809 return new FPExtInst(Load, II->getType());
812 case Intrinsic::ppc_qpx_qvlfd:
813 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
814 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
816 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
817 PointerType::getUnqual(II->getType()));
818 return new LoadInst(Ptr);
821 case Intrinsic::ppc_qpx_qvstfs:
822 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
823 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
825 Type *VTy = VectorType::get(Builder->getFloatTy(),
826 II->getArgOperand(0)->getType()->getVectorNumElements());
827 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
828 Type *OpPtrTy = PointerType::getUnqual(VTy);
829 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
830 return new StoreInst(TOp, Ptr);
833 case Intrinsic::ppc_qpx_qvstfd:
834 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
835 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
838 PointerType::getUnqual(II->getArgOperand(0)->getType());
839 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
840 return new StoreInst(II->getArgOperand(0), Ptr);
844 case Intrinsic::x86_sse_storeu_ps:
845 case Intrinsic::x86_sse2_storeu_pd:
846 case Intrinsic::x86_sse2_storeu_dq:
847 // Turn X86 storeu -> store if the pointer is known aligned.
848 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
851 PointerType::getUnqual(II->getArgOperand(1)->getType());
852 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
853 return new StoreInst(II->getArgOperand(1), Ptr);
857 case Intrinsic::x86_vcvtph2ps_128:
858 case Intrinsic::x86_vcvtph2ps_256: {
859 auto Arg = II->getArgOperand(0);
860 auto ArgType = cast<VectorType>(Arg->getType());
861 auto RetType = cast<VectorType>(II->getType());
862 unsigned ArgWidth = ArgType->getNumElements();
863 unsigned RetWidth = RetType->getNumElements();
864 assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
865 assert(ArgType->isIntOrIntVectorTy() &&
866 ArgType->getScalarSizeInBits() == 16 &&
867 "CVTPH2PS input type should be 16-bit integer vector");
868 assert(RetType->getScalarType()->isFloatTy() &&
869 "CVTPH2PS output type should be 32-bit float vector");
871 // Constant folding: Convert to generic half to single conversion.
872 if (isa<ConstantAggregateZero>(Arg))
873 return ReplaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
875 if (isa<ConstantDataVector>(Arg)) {
876 auto VectorHalfAsShorts = Arg;
877 if (RetWidth < ArgWidth) {
878 SmallVector<int, 8> SubVecMask;
879 for (unsigned i = 0; i != RetWidth; ++i)
880 SubVecMask.push_back((int)i);
881 VectorHalfAsShorts = Builder->CreateShuffleVector(
882 Arg, UndefValue::get(ArgType), SubVecMask);
885 auto VectorHalfType =
886 VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
888 Builder->CreateBitCast(VectorHalfAsShorts, VectorHalfType);
889 auto VectorFloats = Builder->CreateFPExt(VectorHalfs, RetType);
890 return ReplaceInstUsesWith(*II, VectorFloats);
893 // We only use the lowest lanes of the argument.
894 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) {
895 II->setArgOperand(0, V);
901 case Intrinsic::x86_sse_cvtss2si:
902 case Intrinsic::x86_sse_cvtss2si64:
903 case Intrinsic::x86_sse_cvttss2si:
904 case Intrinsic::x86_sse_cvttss2si64:
905 case Intrinsic::x86_sse2_cvtsd2si:
906 case Intrinsic::x86_sse2_cvtsd2si64:
907 case Intrinsic::x86_sse2_cvttsd2si:
908 case Intrinsic::x86_sse2_cvttsd2si64: {
909 // These intrinsics only demand the 0th element of their input vectors. If
910 // we can simplify the input based on that, do so now.
911 Value *Arg = II->getArgOperand(0);
912 unsigned VWidth = Arg->getType()->getVectorNumElements();
913 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
914 II->setArgOperand(0, V);
920 // Constant fold ashr( <A x Bi>, Ci ).
921 // Constant fold lshr( <A x Bi>, Ci ).
922 // Constant fold shl( <A x Bi>, Ci ).
923 case Intrinsic::x86_sse2_psrai_d:
924 case Intrinsic::x86_sse2_psrai_w:
925 case Intrinsic::x86_avx2_psrai_d:
926 case Intrinsic::x86_avx2_psrai_w:
927 case Intrinsic::x86_sse2_psrli_d:
928 case Intrinsic::x86_sse2_psrli_q:
929 case Intrinsic::x86_sse2_psrli_w:
930 case Intrinsic::x86_avx2_psrli_d:
931 case Intrinsic::x86_avx2_psrli_q:
932 case Intrinsic::x86_avx2_psrli_w:
933 case Intrinsic::x86_sse2_pslli_d:
934 case Intrinsic::x86_sse2_pslli_q:
935 case Intrinsic::x86_sse2_pslli_w:
936 case Intrinsic::x86_avx2_pslli_d:
937 case Intrinsic::x86_avx2_pslli_q:
938 case Intrinsic::x86_avx2_pslli_w:
939 if (Value *V = SimplifyX86immshift(*II, *Builder))
940 return ReplaceInstUsesWith(*II, V);
943 case Intrinsic::x86_sse2_psra_d:
944 case Intrinsic::x86_sse2_psra_w:
945 case Intrinsic::x86_avx2_psra_d:
946 case Intrinsic::x86_avx2_psra_w:
947 case Intrinsic::x86_sse2_psrl_d:
948 case Intrinsic::x86_sse2_psrl_q:
949 case Intrinsic::x86_sse2_psrl_w:
950 case Intrinsic::x86_avx2_psrl_d:
951 case Intrinsic::x86_avx2_psrl_q:
952 case Intrinsic::x86_avx2_psrl_w:
953 case Intrinsic::x86_sse2_psll_d:
954 case Intrinsic::x86_sse2_psll_q:
955 case Intrinsic::x86_sse2_psll_w:
956 case Intrinsic::x86_avx2_psll_d:
957 case Intrinsic::x86_avx2_psll_q:
958 case Intrinsic::x86_avx2_psll_w: {
959 if (Value *V = SimplifyX86immshift(*II, *Builder))
960 return ReplaceInstUsesWith(*II, V);
962 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
963 // operand to compute the shift amount.
964 Value *Arg1 = II->getArgOperand(1);
965 assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 &&
966 "Unexpected packed shift size");
967 unsigned VWidth = Arg1->getType()->getVectorNumElements();
969 if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) {
970 II->setArgOperand(1, V);
976 case Intrinsic::x86_avx2_pmovsxbd:
977 case Intrinsic::x86_avx2_pmovsxbq:
978 case Intrinsic::x86_avx2_pmovsxbw:
979 case Intrinsic::x86_avx2_pmovsxdq:
980 case Intrinsic::x86_avx2_pmovsxwd:
981 case Intrinsic::x86_avx2_pmovsxwq:
982 if (Value *V = SimplifyX86extend(*II, *Builder, true))
983 return ReplaceInstUsesWith(*II, V);
986 case Intrinsic::x86_sse41_pmovzxbd:
987 case Intrinsic::x86_sse41_pmovzxbq:
988 case Intrinsic::x86_sse41_pmovzxbw:
989 case Intrinsic::x86_sse41_pmovzxdq:
990 case Intrinsic::x86_sse41_pmovzxwd:
991 case Intrinsic::x86_sse41_pmovzxwq:
992 case Intrinsic::x86_avx2_pmovzxbd:
993 case Intrinsic::x86_avx2_pmovzxbq:
994 case Intrinsic::x86_avx2_pmovzxbw:
995 case Intrinsic::x86_avx2_pmovzxdq:
996 case Intrinsic::x86_avx2_pmovzxwd:
997 case Intrinsic::x86_avx2_pmovzxwq:
998 if (Value *V = SimplifyX86extend(*II, *Builder, false))
999 return ReplaceInstUsesWith(*II, V);
1002 case Intrinsic::x86_sse41_insertps:
1003 if (Value *V = SimplifyX86insertps(*II, *Builder))
1004 return ReplaceInstUsesWith(*II, V);
1007 case Intrinsic::x86_sse4a_extrq: {
1008 // EXTRQ uses only the lowest 64-bits of the first 128-bit vector
1009 // operands and the lowest 16-bits of the second.
1010 Value *Op0 = II->getArgOperand(0);
1011 Value *Op1 = II->getArgOperand(1);
1012 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
1013 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
1014 assert(VWidth0 == 2 && VWidth1 == 16 && "Unexpected operand sizes");
1016 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
1017 II->setArgOperand(0, V);
1020 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
1021 II->setArgOperand(1, V);
1027 case Intrinsic::x86_sse4a_extrqi: {
1028 // EXTRQI uses only the lowest 64-bits of the first 128-bit vector
1030 Value *Op = II->getArgOperand(0);
1031 unsigned VWidth = Op->getType()->getVectorNumElements();
1032 assert(VWidth == 2 && "Unexpected operand size");
1034 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1035 II->setArgOperand(0, V);
1041 case Intrinsic::x86_sse4a_insertq: {
1042 // INSERTQ uses only the lowest 64-bits of the first 128-bit vector
1044 Value *Op = II->getArgOperand(0);
1045 unsigned VWidth = Op->getType()->getVectorNumElements();
1046 assert(VWidth == 2 && "Unexpected operand size");
1048 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1049 II->setArgOperand(0, V);
1055 case Intrinsic::x86_sse4a_insertqi: {
1056 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
1058 // TODO: eventually we should lower this intrinsic to IR
1059 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
1060 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
1061 unsigned Index = CIIndex->getZExtValue();
1063 // From AMD documentation: "a value of zero in the field length is
1064 // defined as length of 64".
1065 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
1067 // From AMD documentation: "If the sum of the bit index + length field
1068 // is greater than 64, the results are undefined".
1069 unsigned End = Index + Length;
1071 // Note that both field index and field length are 8-bit quantities.
1072 // Since variables 'Index' and 'Length' are unsigned values
1073 // obtained from zero-extending field index and field length
1074 // respectively, their sum should never wrap around.
1076 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
1078 if (Length == 64 && Index == 0) {
1079 Value *Vec = II->getArgOperand(1);
1080 Value *Undef = UndefValue::get(Vec->getType());
1081 const uint32_t Mask[] = {0, 2};
1082 return ReplaceInstUsesWith(
1084 Builder->CreateShuffleVector(
1085 Vec, Undef, ConstantDataVector::get(
1086 II->getContext(), makeArrayRef(Mask))));
1091 // INSERTQI uses only the lowest 64-bits of the first two 128-bit vector
1093 Value *Op0 = II->getArgOperand(0);
1094 Value *Op1 = II->getArgOperand(1);
1095 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
1096 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
1097 assert(VWidth0 == 2 && VWidth1 == 2 && "Unexpected operand sizes");
1099 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
1100 II->setArgOperand(0, V);
1104 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
1105 II->setArgOperand(1, V);
1111 case Intrinsic::x86_sse41_pblendvb:
1112 case Intrinsic::x86_sse41_blendvps:
1113 case Intrinsic::x86_sse41_blendvpd:
1114 case Intrinsic::x86_avx_blendv_ps_256:
1115 case Intrinsic::x86_avx_blendv_pd_256:
1116 case Intrinsic::x86_avx2_pblendvb: {
1117 // Convert blendv* to vector selects if the mask is constant.
1118 // This optimization is convoluted because the intrinsic is defined as
1119 // getting a vector of floats or doubles for the ps and pd versions.
1120 // FIXME: That should be changed.
1122 Value *Op0 = II->getArgOperand(0);
1123 Value *Op1 = II->getArgOperand(1);
1124 Value *Mask = II->getArgOperand(2);
1126 // fold (blend A, A, Mask) -> A
1128 return ReplaceInstUsesWith(CI, Op0);
1130 // Zero Mask - select 1st argument.
1131 if (isa<ConstantAggregateZero>(Mask))
1132 return ReplaceInstUsesWith(CI, Op0);
1134 // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
1135 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
1136 auto Tyi1 = Builder->getInt1Ty();
1137 auto SelectorType = cast<VectorType>(Mask->getType());
1138 auto EltTy = SelectorType->getElementType();
1139 unsigned Size = SelectorType->getNumElements();
1143 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
1144 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
1145 "Wrong arguments for variable blend intrinsic");
1146 SmallVector<Constant *, 32> Selectors;
1147 for (unsigned I = 0; I < Size; ++I) {
1148 // The intrinsics only read the top bit
1151 Selector = C->getElementAsInteger(I);
1153 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
1154 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
1156 auto NewSelector = ConstantVector::get(Selectors);
1157 return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
1162 case Intrinsic::x86_ssse3_pshuf_b_128:
1163 case Intrinsic::x86_avx2_pshuf_b: {
1164 // Turn pshufb(V1,mask) -> shuffle(V1,Zero,mask) if mask is a constant.
1165 auto *V = II->getArgOperand(1);
1166 auto *VTy = cast<VectorType>(V->getType());
1167 unsigned NumElts = VTy->getNumElements();
1168 assert((NumElts == 16 || NumElts == 32) &&
1169 "Unexpected number of elements in shuffle mask!");
1170 // Initialize the resulting shuffle mask to all zeroes.
1171 uint32_t Indexes[32] = {0};
1173 if (auto *Mask = dyn_cast<ConstantDataVector>(V)) {
1174 // Each byte in the shuffle control mask forms an index to permute the
1175 // corresponding byte in the destination operand.
1176 for (unsigned I = 0; I < NumElts; ++I) {
1177 int8_t Index = Mask->getElementAsInteger(I);
1178 // If the most significant bit (bit[7]) of each byte of the shuffle
1179 // control mask is set, then zero is written in the result byte.
1180 // The zero vector is in the right-hand side of the resulting
1183 // The value of each index is the least significant 4 bits of the
1184 // shuffle control byte.
1185 Indexes[I] = (Index < 0) ? NumElts : Index & 0xF;
1187 } else if (!isa<ConstantAggregateZero>(V))
1190 // The value of each index for the high 128-bit lane is the least
1191 // significant 4 bits of the respective shuffle control byte.
1192 for (unsigned I = 16; I < NumElts; ++I)
1193 Indexes[I] += I & 0xF0;
1195 auto NewC = ConstantDataVector::get(V->getContext(),
1196 makeArrayRef(Indexes, NumElts));
1197 auto V1 = II->getArgOperand(0);
1198 auto V2 = Constant::getNullValue(II->getType());
1199 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1200 return ReplaceInstUsesWith(CI, Shuffle);
1203 case Intrinsic::x86_avx_vpermilvar_ps:
1204 case Intrinsic::x86_avx_vpermilvar_ps_256:
1205 case Intrinsic::x86_avx_vpermilvar_pd:
1206 case Intrinsic::x86_avx_vpermilvar_pd_256: {
1207 // Convert vpermil* to shufflevector if the mask is constant.
1208 Value *V = II->getArgOperand(1);
1209 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1210 assert(Size == 8 || Size == 4 || Size == 2);
1211 uint32_t Indexes[8];
1212 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1213 // The intrinsics only read one or two bits, clear the rest.
1214 for (unsigned I = 0; I < Size; ++I) {
1215 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1216 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1217 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1221 } else if (isa<ConstantAggregateZero>(V)) {
1222 for (unsigned I = 0; I < Size; ++I)
1227 // The _256 variants are a bit trickier since the mask bits always index
1228 // into the corresponding 128 half. In order to convert to a generic
1229 // shuffle, we have to make that explicit.
1230 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1231 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1232 for (unsigned I = Size / 2; I < Size; ++I)
1233 Indexes[I] += Size / 2;
1236 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1237 auto V1 = II->getArgOperand(0);
1238 auto V2 = UndefValue::get(V1->getType());
1239 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1240 return ReplaceInstUsesWith(CI, Shuffle);
1243 case Intrinsic::x86_avx_vperm2f128_pd_256:
1244 case Intrinsic::x86_avx_vperm2f128_ps_256:
1245 case Intrinsic::x86_avx_vperm2f128_si_256:
1246 case Intrinsic::x86_avx2_vperm2i128:
1247 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1248 return ReplaceInstUsesWith(*II, V);
1251 case Intrinsic::x86_xop_vpcomb:
1252 case Intrinsic::x86_xop_vpcomd:
1253 case Intrinsic::x86_xop_vpcomq:
1254 case Intrinsic::x86_xop_vpcomw:
1255 if (Value *V = SimplifyX86vpcom(*II, *Builder, true))
1256 return ReplaceInstUsesWith(*II, V);
1259 case Intrinsic::x86_xop_vpcomub:
1260 case Intrinsic::x86_xop_vpcomud:
1261 case Intrinsic::x86_xop_vpcomuq:
1262 case Intrinsic::x86_xop_vpcomuw:
1263 if (Value *V = SimplifyX86vpcom(*II, *Builder, false))
1264 return ReplaceInstUsesWith(*II, V);
1267 case Intrinsic::ppc_altivec_vperm:
1268 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1269 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1270 // a vectorshuffle for little endian, we must undo the transformation
1271 // performed on vec_perm in altivec.h. That is, we must complement
1272 // the permutation mask with respect to 31 and reverse the order of
1274 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1275 assert(Mask->getType()->getVectorNumElements() == 16 &&
1276 "Bad type for intrinsic!");
1278 // Check that all of the elements are integer constants or undefs.
1279 bool AllEltsOk = true;
1280 for (unsigned i = 0; i != 16; ++i) {
1281 Constant *Elt = Mask->getAggregateElement(i);
1282 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1289 // Cast the input vectors to byte vectors.
1290 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1292 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1294 Value *Result = UndefValue::get(Op0->getType());
1296 // Only extract each element once.
1297 Value *ExtractedElts[32];
1298 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1300 for (unsigned i = 0; i != 16; ++i) {
1301 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1304 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1305 Idx &= 31; // Match the hardware behavior.
1306 if (DL.isLittleEndian())
1309 if (!ExtractedElts[Idx]) {
1310 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1311 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1312 ExtractedElts[Idx] =
1313 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1314 Builder->getInt32(Idx&15));
1317 // Insert this value into the result vector.
1318 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1319 Builder->getInt32(i));
1321 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1326 case Intrinsic::arm_neon_vld1:
1327 case Intrinsic::arm_neon_vld2:
1328 case Intrinsic::arm_neon_vld3:
1329 case Intrinsic::arm_neon_vld4:
1330 case Intrinsic::arm_neon_vld2lane:
1331 case Intrinsic::arm_neon_vld3lane:
1332 case Intrinsic::arm_neon_vld4lane:
1333 case Intrinsic::arm_neon_vst1:
1334 case Intrinsic::arm_neon_vst2:
1335 case Intrinsic::arm_neon_vst3:
1336 case Intrinsic::arm_neon_vst4:
1337 case Intrinsic::arm_neon_vst2lane:
1338 case Intrinsic::arm_neon_vst3lane:
1339 case Intrinsic::arm_neon_vst4lane: {
1340 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1341 unsigned AlignArg = II->getNumArgOperands() - 1;
1342 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1343 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1344 II->setArgOperand(AlignArg,
1345 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1352 case Intrinsic::arm_neon_vmulls:
1353 case Intrinsic::arm_neon_vmullu:
1354 case Intrinsic::aarch64_neon_smull:
1355 case Intrinsic::aarch64_neon_umull: {
1356 Value *Arg0 = II->getArgOperand(0);
1357 Value *Arg1 = II->getArgOperand(1);
1359 // Handle mul by zero first:
1360 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1361 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1364 // Check for constant LHS & RHS - in this case we just simplify.
1365 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1366 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1367 VectorType *NewVT = cast<VectorType>(II->getType());
1368 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1369 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1370 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1371 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1373 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1376 // Couldn't simplify - canonicalize constant to the RHS.
1377 std::swap(Arg0, Arg1);
1380 // Handle mul by one:
1381 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1382 if (ConstantInt *Splat =
1383 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1385 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1386 /*isSigned=*/!Zext);
1391 case Intrinsic::AMDGPU_rcp: {
1392 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1393 const APFloat &ArgVal = C->getValueAPF();
1394 APFloat Val(ArgVal.getSemantics(), 1.0);
1395 APFloat::opStatus Status = Val.divide(ArgVal,
1396 APFloat::rmNearestTiesToEven);
1397 // Only do this if it was exact and therefore not dependent on the
1399 if (Status == APFloat::opOK)
1400 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1405 case Intrinsic::stackrestore: {
1406 // If the save is right next to the restore, remove the restore. This can
1407 // happen when variable allocas are DCE'd.
1408 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1409 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1410 if (&*++SS->getIterator() == II)
1411 return EraseInstFromFunction(CI);
1415 // Scan down this block to see if there is another stack restore in the
1416 // same block without an intervening call/alloca.
1417 BasicBlock::iterator BI(II);
1418 TerminatorInst *TI = II->getParent()->getTerminator();
1419 bool CannotRemove = false;
1420 for (++BI; &*BI != TI; ++BI) {
1421 if (isa<AllocaInst>(BI)) {
1422 CannotRemove = true;
1425 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1426 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1427 // If there is a stackrestore below this one, remove this one.
1428 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1429 return EraseInstFromFunction(CI);
1430 // Otherwise, ignore the intrinsic.
1432 // If we found a non-intrinsic call, we can't remove the stack
1434 CannotRemove = true;
1440 // If the stack restore is in a return, resume, or unwind block and if there
1441 // are no allocas or calls between the restore and the return, nuke the
1443 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1444 return EraseInstFromFunction(CI);
1447 case Intrinsic::lifetime_start: {
1448 // Remove trivially empty lifetime_start/end ranges, i.e. a start
1449 // immediately followed by an end (ignoring debuginfo or other
1450 // lifetime markers in between).
1451 BasicBlock::iterator BI = II->getIterator(), BE = II->getParent()->end();
1452 for (++BI; BI != BE; ++BI) {
1453 if (IntrinsicInst *LTE = dyn_cast<IntrinsicInst>(BI)) {
1454 if (isa<DbgInfoIntrinsic>(LTE) ||
1455 LTE->getIntrinsicID() == Intrinsic::lifetime_start)
1457 if (LTE->getIntrinsicID() == Intrinsic::lifetime_end) {
1458 if (II->getOperand(0) == LTE->getOperand(0) &&
1459 II->getOperand(1) == LTE->getOperand(1)) {
1460 EraseInstFromFunction(*LTE);
1461 return EraseInstFromFunction(*II);
1470 case Intrinsic::assume: {
1471 // Canonicalize assume(a && b) -> assume(a); assume(b);
1472 // Note: New assumption intrinsics created here are registered by
1473 // the InstCombineIRInserter object.
1474 Value *IIOperand = II->getArgOperand(0), *A, *B,
1475 *AssumeIntrinsic = II->getCalledValue();
1476 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1477 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1478 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1479 return EraseInstFromFunction(*II);
1481 // assume(!(a || b)) -> assume(!a); assume(!b);
1482 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1483 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1485 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1487 return EraseInstFromFunction(*II);
1490 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1491 // (if assume is valid at the load)
1492 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1493 Value *LHS = ICmp->getOperand(0);
1494 Value *RHS = ICmp->getOperand(1);
1495 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1496 isa<LoadInst>(LHS) &&
1497 isa<Constant>(RHS) &&
1498 RHS->getType()->isPointerTy() &&
1499 cast<Constant>(RHS)->isNullValue()) {
1500 LoadInst* LI = cast<LoadInst>(LHS);
1501 if (isValidAssumeForContext(II, LI, DT)) {
1502 MDNode *MD = MDNode::get(II->getContext(), None);
1503 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1504 return EraseInstFromFunction(*II);
1507 // TODO: apply nonnull return attributes to calls and invokes
1508 // TODO: apply range metadata for range check patterns?
1510 // If there is a dominating assume with the same condition as this one,
1511 // then this one is redundant, and should be removed.
1512 APInt KnownZero(1, 0), KnownOne(1, 0);
1513 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1514 if (KnownOne.isAllOnesValue())
1515 return EraseInstFromFunction(*II);
1519 case Intrinsic::experimental_gc_relocate: {
1520 // Translate facts known about a pointer before relocating into
1521 // facts about the relocate value, while being careful to
1522 // preserve relocation semantics.
1523 GCRelocateOperands Operands(II);
1524 Value *DerivedPtr = Operands.getDerivedPtr();
1525 auto *GCRelocateType = cast<PointerType>(II->getType());
1527 // Remove the relocation if unused, note that this check is required
1528 // to prevent the cases below from looping forever.
1529 if (II->use_empty())
1530 return EraseInstFromFunction(*II);
1532 // Undef is undef, even after relocation.
1533 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1534 // most practical collectors, but there was discussion in the review thread
1535 // about whether it was legal for all possible collectors.
1536 if (isa<UndefValue>(DerivedPtr)) {
1537 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1538 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1541 // The relocation of null will be null for most any collector.
1542 // TODO: provide a hook for this in GCStrategy. There might be some weird
1543 // collector this property does not hold for.
1544 if (isa<ConstantPointerNull>(DerivedPtr)) {
1545 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1546 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1549 // isKnownNonNull -> nonnull attribute
1550 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI))
1551 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1553 // isDereferenceablePointer -> deref attribute
1554 if (isDereferenceablePointer(DerivedPtr, DL)) {
1555 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1556 uint64_t Bytes = A->getDereferenceableBytes();
1557 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1561 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1562 // Canonicalize on the type from the uses to the defs
1564 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1568 return visitCallSite(II);
1571 // InvokeInst simplification
1573 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1574 return visitCallSite(&II);
1577 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1578 /// passed through the varargs area, we can eliminate the use of the cast.
1579 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1580 const DataLayout &DL,
1581 const CastInst *const CI,
1583 if (!CI->isLosslessCast())
1586 // If this is a GC intrinsic, avoid munging types. We need types for
1587 // statepoint reconstruction in SelectionDAG.
1588 // TODO: This is probably something which should be expanded to all
1589 // intrinsics since the entire point of intrinsics is that
1590 // they are understandable by the optimizer.
1591 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1594 // The size of ByVal or InAlloca arguments is derived from the type, so we
1595 // can't change to a type with a different size. If the size were
1596 // passed explicitly we could avoid this check.
1597 if (!CS.isByValOrInAllocaArgument(ix))
1601 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1602 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1603 if (!SrcTy->isSized() || !DstTy->isSized())
1605 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1610 // Try to fold some different type of calls here.
1611 // Currently we're only working with the checking functions, memcpy_chk,
1612 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1613 // strcat_chk and strncat_chk.
1614 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1615 if (!CI->getCalledFunction()) return nullptr;
1617 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1618 ReplaceInstUsesWith(*From, With);
1620 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1621 if (Value *With = Simplifier.optimizeCall(CI)) {
1623 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1629 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1630 // Strip off at most one level of pointer casts, looking for an alloca. This
1631 // is good enough in practice and simpler than handling any number of casts.
1632 Value *Underlying = TrampMem->stripPointerCasts();
1633 if (Underlying != TrampMem &&
1634 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1636 if (!isa<AllocaInst>(Underlying))
1639 IntrinsicInst *InitTrampoline = nullptr;
1640 for (User *U : TrampMem->users()) {
1641 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1644 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1646 // More than one init_trampoline writes to this value. Give up.
1648 InitTrampoline = II;
1651 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1652 // Allow any number of calls to adjust.trampoline.
1657 // No call to init.trampoline found.
1658 if (!InitTrampoline)
1661 // Check that the alloca is being used in the expected way.
1662 if (InitTrampoline->getOperand(0) != TrampMem)
1665 return InitTrampoline;
1668 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1670 // Visit all the previous instructions in the basic block, and try to find a
1671 // init.trampoline which has a direct path to the adjust.trampoline.
1672 for (BasicBlock::iterator I = AdjustTramp->getIterator(),
1673 E = AdjustTramp->getParent()->begin();
1675 Instruction *Inst = &*--I;
1676 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1677 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1678 II->getOperand(0) == TrampMem)
1680 if (Inst->mayWriteToMemory())
1686 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1687 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1688 // to a direct call to a function. Otherwise return NULL.
1690 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1691 Callee = Callee->stripPointerCasts();
1692 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1694 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1697 Value *TrampMem = AdjustTramp->getOperand(0);
1699 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1701 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1706 // visitCallSite - Improvements for call and invoke instructions.
1708 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1710 if (isAllocLikeFn(CS.getInstruction(), TLI))
1711 return visitAllocSite(*CS.getInstruction());
1713 bool Changed = false;
1715 // Mark any parameters that are known to be non-null with the nonnull
1716 // attribute. This is helpful for inlining calls to functions with null
1717 // checks on their arguments.
1719 for (Value *V : CS.args()) {
1720 if (V->getType()->isPointerTy() && !CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1721 isKnownNonNullAt(V, CS.getInstruction(), DT, TLI)) {
1722 AttributeSet AS = CS.getAttributes();
1723 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1724 Attribute::NonNull);
1725 CS.setAttributes(AS);
1730 assert(ArgNo == CS.arg_size() && "sanity check");
1732 // If the callee is a pointer to a function, attempt to move any casts to the
1733 // arguments of the call/invoke.
1734 Value *Callee = CS.getCalledValue();
1735 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1738 if (Function *CalleeF = dyn_cast<Function>(Callee))
1739 // If the call and callee calling conventions don't match, this call must
1740 // be unreachable, as the call is undefined.
1741 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1742 // Only do this for calls to a function with a body. A prototype may
1743 // not actually end up matching the implementation's calling conv for a
1744 // variety of reasons (e.g. it may be written in assembly).
1745 !CalleeF->isDeclaration()) {
1746 Instruction *OldCall = CS.getInstruction();
1747 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1748 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1750 // If OldCall does not return void then replaceAllUsesWith undef.
1751 // This allows ValueHandlers and custom metadata to adjust itself.
1752 if (!OldCall->getType()->isVoidTy())
1753 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1754 if (isa<CallInst>(OldCall))
1755 return EraseInstFromFunction(*OldCall);
1757 // We cannot remove an invoke, because it would change the CFG, just
1758 // change the callee to a null pointer.
1759 cast<InvokeInst>(OldCall)->setCalledFunction(
1760 Constant::getNullValue(CalleeF->getType()));
1764 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1765 // If CS does not return void then replaceAllUsesWith undef.
1766 // This allows ValueHandlers and custom metadata to adjust itself.
1767 if (!CS.getInstruction()->getType()->isVoidTy())
1768 ReplaceInstUsesWith(*CS.getInstruction(),
1769 UndefValue::get(CS.getInstruction()->getType()));
1771 if (isa<InvokeInst>(CS.getInstruction())) {
1772 // Can't remove an invoke because we cannot change the CFG.
1776 // This instruction is not reachable, just remove it. We insert a store to
1777 // undef so that we know that this code is not reachable, despite the fact
1778 // that we can't modify the CFG here.
1779 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1780 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1781 CS.getInstruction());
1783 return EraseInstFromFunction(*CS.getInstruction());
1786 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1787 return transformCallThroughTrampoline(CS, II);
1789 PointerType *PTy = cast<PointerType>(Callee->getType());
1790 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1791 if (FTy->isVarArg()) {
1792 int ix = FTy->getNumParams();
1793 // See if we can optimize any arguments passed through the varargs area of
1795 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1796 E = CS.arg_end(); I != E; ++I, ++ix) {
1797 CastInst *CI = dyn_cast<CastInst>(*I);
1798 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1799 *I = CI->getOperand(0);
1805 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1806 // Inline asm calls cannot throw - mark them 'nounwind'.
1807 CS.setDoesNotThrow();
1811 // Try to optimize the call if possible, we require DataLayout for most of
1812 // this. None of these calls are seen as possibly dead so go ahead and
1813 // delete the instruction now.
1814 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1815 Instruction *I = tryOptimizeCall(CI);
1816 // If we changed something return the result, etc. Otherwise let
1817 // the fallthrough check.
1818 if (I) return EraseInstFromFunction(*I);
1821 return Changed ? CS.getInstruction() : nullptr;
1824 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1825 // attempt to move the cast to the arguments of the call/invoke.
1827 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1829 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1832 // The prototype of thunks are a lie, don't try to directly call such
1834 if (Callee->hasFnAttribute("thunk"))
1836 Instruction *Caller = CS.getInstruction();
1837 const AttributeSet &CallerPAL = CS.getAttributes();
1839 // Okay, this is a cast from a function to a different type. Unless doing so
1840 // would cause a type conversion of one of our arguments, change this call to
1841 // be a direct call with arguments casted to the appropriate types.
1843 FunctionType *FT = Callee->getFunctionType();
1844 Type *OldRetTy = Caller->getType();
1845 Type *NewRetTy = FT->getReturnType();
1847 // Check to see if we are changing the return type...
1848 if (OldRetTy != NewRetTy) {
1850 if (NewRetTy->isStructTy())
1851 return false; // TODO: Handle multiple return values.
1853 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1854 if (Callee->isDeclaration())
1855 return false; // Cannot transform this return value.
1857 if (!Caller->use_empty() &&
1858 // void -> non-void is handled specially
1859 !NewRetTy->isVoidTy())
1860 return false; // Cannot transform this return value.
1863 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1864 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1865 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1866 return false; // Attribute not compatible with transformed value.
1869 // If the callsite is an invoke instruction, and the return value is used by
1870 // a PHI node in a successor, we cannot change the return type of the call
1871 // because there is no place to put the cast instruction (without breaking
1872 // the critical edge). Bail out in this case.
1873 if (!Caller->use_empty())
1874 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1875 for (User *U : II->users())
1876 if (PHINode *PN = dyn_cast<PHINode>(U))
1877 if (PN->getParent() == II->getNormalDest() ||
1878 PN->getParent() == II->getUnwindDest())
1882 unsigned NumActualArgs = CS.arg_size();
1883 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1885 // Prevent us turning:
1886 // declare void @takes_i32_inalloca(i32* inalloca)
1887 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1890 // call void @takes_i32_inalloca(i32* null)
1892 // Similarly, avoid folding away bitcasts of byval calls.
1893 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1894 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1897 CallSite::arg_iterator AI = CS.arg_begin();
1898 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1899 Type *ParamTy = FT->getParamType(i);
1900 Type *ActTy = (*AI)->getType();
1902 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1903 return false; // Cannot transform this parameter value.
1905 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1906 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1907 return false; // Attribute not compatible with transformed value.
1909 if (CS.isInAllocaArgument(i))
1910 return false; // Cannot transform to and from inalloca.
1912 // If the parameter is passed as a byval argument, then we have to have a
1913 // sized type and the sized type has to have the same size as the old type.
1914 if (ParamTy != ActTy &&
1915 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1916 Attribute::ByVal)) {
1917 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1918 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1921 Type *CurElTy = ActTy->getPointerElementType();
1922 if (DL.getTypeAllocSize(CurElTy) !=
1923 DL.getTypeAllocSize(ParamPTy->getElementType()))
1928 if (Callee->isDeclaration()) {
1929 // Do not delete arguments unless we have a function body.
1930 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1933 // If the callee is just a declaration, don't change the varargsness of the
1934 // call. We don't want to introduce a varargs call where one doesn't
1936 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1937 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1940 // If both the callee and the cast type are varargs, we still have to make
1941 // sure the number of fixed parameters are the same or we have the same
1942 // ABI issues as if we introduce a varargs call.
1943 if (FT->isVarArg() &&
1944 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1945 FT->getNumParams() !=
1946 cast<FunctionType>(APTy->getElementType())->getNumParams())
1950 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1951 !CallerPAL.isEmpty())
1952 // In this case we have more arguments than the new function type, but we
1953 // won't be dropping them. Check that these extra arguments have attributes
1954 // that are compatible with being a vararg call argument.
1955 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1956 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1957 if (Index <= FT->getNumParams())
1960 // Check if it has an attribute that's incompatible with varargs.
1961 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1962 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1967 // Okay, we decided that this is a safe thing to do: go ahead and start
1968 // inserting cast instructions as necessary.
1969 std::vector<Value*> Args;
1970 Args.reserve(NumActualArgs);
1971 SmallVector<AttributeSet, 8> attrVec;
1972 attrVec.reserve(NumCommonArgs);
1974 // Get any return attributes.
1975 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1977 // If the return value is not being used, the type may not be compatible
1978 // with the existing attributes. Wipe out any problematic attributes.
1979 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1981 // Add the new return attributes.
1982 if (RAttrs.hasAttributes())
1983 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1984 AttributeSet::ReturnIndex, RAttrs));
1986 AI = CS.arg_begin();
1987 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1988 Type *ParamTy = FT->getParamType(i);
1990 if ((*AI)->getType() == ParamTy) {
1991 Args.push_back(*AI);
1993 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1996 // Add any parameter attributes.
1997 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1998 if (PAttrs.hasAttributes())
1999 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
2003 // If the function takes more arguments than the call was taking, add them
2005 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
2006 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
2008 // If we are removing arguments to the function, emit an obnoxious warning.
2009 if (FT->getNumParams() < NumActualArgs) {
2010 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
2011 if (FT->isVarArg()) {
2012 // Add all of the arguments in their promoted form to the arg list.
2013 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
2014 Type *PTy = getPromotedType((*AI)->getType());
2015 if (PTy != (*AI)->getType()) {
2016 // Must promote to pass through va_arg area!
2017 Instruction::CastOps opcode =
2018 CastInst::getCastOpcode(*AI, false, PTy, false);
2019 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
2021 Args.push_back(*AI);
2024 // Add any parameter attributes.
2025 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
2026 if (PAttrs.hasAttributes())
2027 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
2033 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
2034 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
2035 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
2037 if (NewRetTy->isVoidTy())
2038 Caller->setName(""); // Void type should not have a name.
2040 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
2044 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2045 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
2046 II->getUnwindDest(), Args);
2048 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
2049 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
2051 CallInst *CI = cast<CallInst>(Caller);
2052 NC = Builder->CreateCall(Callee, Args);
2054 if (CI->isTailCall())
2055 cast<CallInst>(NC)->setTailCall();
2056 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
2057 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
2060 // Insert a cast of the return type as necessary.
2062 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
2063 if (!NV->getType()->isVoidTy()) {
2064 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
2065 NC->setDebugLoc(Caller->getDebugLoc());
2067 // If this is an invoke instruction, we should insert it after the first
2068 // non-phi, instruction in the normal successor block.
2069 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2070 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
2071 InsertNewInstBefore(NC, *I);
2073 // Otherwise, it's a call, just insert cast right after the call.
2074 InsertNewInstBefore(NC, *Caller);
2076 Worklist.AddUsersToWorkList(*Caller);
2078 NV = UndefValue::get(Caller->getType());
2082 if (!Caller->use_empty())
2083 ReplaceInstUsesWith(*Caller, NV);
2084 else if (Caller->hasValueHandle()) {
2085 if (OldRetTy == NV->getType())
2086 ValueHandleBase::ValueIsRAUWd(Caller, NV);
2088 // We cannot call ValueIsRAUWd with a different type, and the
2089 // actual tracked value will disappear.
2090 ValueHandleBase::ValueIsDeleted(Caller);
2093 EraseInstFromFunction(*Caller);
2097 // transformCallThroughTrampoline - Turn a call to a function created by
2098 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
2099 // underlying function.
2102 InstCombiner::transformCallThroughTrampoline(CallSite CS,
2103 IntrinsicInst *Tramp) {
2104 Value *Callee = CS.getCalledValue();
2105 PointerType *PTy = cast<PointerType>(Callee->getType());
2106 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
2107 const AttributeSet &Attrs = CS.getAttributes();
2109 // If the call already has the 'nest' attribute somewhere then give up -
2110 // otherwise 'nest' would occur twice after splicing in the chain.
2111 if (Attrs.hasAttrSomewhere(Attribute::Nest))
2115 "transformCallThroughTrampoline called with incorrect CallSite.");
2117 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
2118 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
2119 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
2121 const AttributeSet &NestAttrs = NestF->getAttributes();
2122 if (!NestAttrs.isEmpty()) {
2123 unsigned NestIdx = 1;
2124 Type *NestTy = nullptr;
2125 AttributeSet NestAttr;
2127 // Look for a parameter marked with the 'nest' attribute.
2128 for (FunctionType::param_iterator I = NestFTy->param_begin(),
2129 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
2130 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
2131 // Record the parameter type and any other attributes.
2133 NestAttr = NestAttrs.getParamAttributes(NestIdx);
2138 Instruction *Caller = CS.getInstruction();
2139 std::vector<Value*> NewArgs;
2140 NewArgs.reserve(CS.arg_size() + 1);
2142 SmallVector<AttributeSet, 8> NewAttrs;
2143 NewAttrs.reserve(Attrs.getNumSlots() + 1);
2145 // Insert the nest argument into the call argument list, which may
2146 // mean appending it. Likewise for attributes.
2148 // Add any result attributes.
2149 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
2150 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2151 Attrs.getRetAttributes()));
2155 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
2157 if (Idx == NestIdx) {
2158 // Add the chain argument and attributes.
2159 Value *NestVal = Tramp->getArgOperand(2);
2160 if (NestVal->getType() != NestTy)
2161 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
2162 NewArgs.push_back(NestVal);
2163 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2170 // Add the original argument and attributes.
2171 NewArgs.push_back(*I);
2172 AttributeSet Attr = Attrs.getParamAttributes(Idx);
2173 if (Attr.hasAttributes(Idx)) {
2174 AttrBuilder B(Attr, Idx);
2175 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2176 Idx + (Idx >= NestIdx), B));
2183 // Add any function attributes.
2184 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
2185 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
2186 Attrs.getFnAttributes()));
2188 // The trampoline may have been bitcast to a bogus type (FTy).
2189 // Handle this by synthesizing a new function type, equal to FTy
2190 // with the chain parameter inserted.
2192 std::vector<Type*> NewTypes;
2193 NewTypes.reserve(FTy->getNumParams()+1);
2195 // Insert the chain's type into the list of parameter types, which may
2196 // mean appending it.
2199 FunctionType::param_iterator I = FTy->param_begin(),
2200 E = FTy->param_end();
2204 // Add the chain's type.
2205 NewTypes.push_back(NestTy);
2210 // Add the original type.
2211 NewTypes.push_back(*I);
2217 // Replace the trampoline call with a direct call. Let the generic
2218 // code sort out any function type mismatches.
2219 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
2221 Constant *NewCallee =
2222 NestF->getType() == PointerType::getUnqual(NewFTy) ?
2223 NestF : ConstantExpr::getBitCast(NestF,
2224 PointerType::getUnqual(NewFTy));
2225 const AttributeSet &NewPAL =
2226 AttributeSet::get(FTy->getContext(), NewAttrs);
2228 Instruction *NewCaller;
2229 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2230 NewCaller = InvokeInst::Create(NewCallee,
2231 II->getNormalDest(), II->getUnwindDest(),
2233 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
2234 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
2236 NewCaller = CallInst::Create(NewCallee, NewArgs);
2237 if (cast<CallInst>(Caller)->isTailCall())
2238 cast<CallInst>(NewCaller)->setTailCall();
2239 cast<CallInst>(NewCaller)->
2240 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
2241 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
2248 // Replace the trampoline call with a direct call. Since there is no 'nest'
2249 // parameter, there is no need to adjust the argument list. Let the generic
2250 // code sort out any function type mismatches.
2251 Constant *NewCallee =
2252 NestF->getType() == PTy ? NestF :
2253 ConstantExpr::getBitCast(NestF, PTy);
2254 CS.setCalledFunction(NewCallee);
2255 return CS.getInstruction();