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 /// visitCallInst - CallInst simplification. This mostly only handles folding
450 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
451 /// the heavy lifting.
453 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
454 auto Args = CI.arg_operands();
455 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
457 return ReplaceInstUsesWith(CI, V);
459 if (isFreeCall(&CI, TLI))
460 return visitFree(CI);
462 // If the caller function is nounwind, mark the call as nounwind, even if the
464 if (CI.getParent()->getParent()->doesNotThrow() &&
465 !CI.doesNotThrow()) {
466 CI.setDoesNotThrow();
470 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
471 if (!II) return visitCallSite(&CI);
473 // Intrinsics cannot occur in an invoke, so handle them here instead of in
475 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
476 bool Changed = false;
478 // memmove/cpy/set of zero bytes is a noop.
479 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
480 if (NumBytes->isNullValue())
481 return EraseInstFromFunction(CI);
483 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
484 if (CI->getZExtValue() == 1) {
485 // Replace the instruction with just byte operations. We would
486 // transform other cases to loads/stores, but we don't know if
487 // alignment is sufficient.
491 // No other transformations apply to volatile transfers.
492 if (MI->isVolatile())
495 // If we have a memmove and the source operation is a constant global,
496 // then the source and dest pointers can't alias, so we can change this
497 // into a call to memcpy.
498 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
499 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
500 if (GVSrc->isConstant()) {
501 Module *M = CI.getParent()->getParent()->getParent();
502 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
503 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
504 CI.getArgOperand(1)->getType(),
505 CI.getArgOperand(2)->getType() };
506 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
511 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
512 // memmove(x,x,size) -> noop.
513 if (MTI->getSource() == MTI->getDest())
514 return EraseInstFromFunction(CI);
517 // If we can determine a pointer alignment that is bigger than currently
518 // set, update the alignment.
519 if (isa<MemTransferInst>(MI)) {
520 if (Instruction *I = SimplifyMemTransfer(MI))
522 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
523 if (Instruction *I = SimplifyMemSet(MSI))
527 if (Changed) return II;
530 auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width, unsigned DemandedWidth)
532 APInt UndefElts(Width, 0);
533 APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
534 return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
537 switch (II->getIntrinsicID()) {
539 case Intrinsic::objectsize: {
541 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
542 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
545 case Intrinsic::bswap: {
546 Value *IIOperand = II->getArgOperand(0);
549 // bswap(bswap(x)) -> x
550 if (match(IIOperand, m_BSwap(m_Value(X))))
551 return ReplaceInstUsesWith(CI, X);
553 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
554 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
555 unsigned C = X->getType()->getPrimitiveSizeInBits() -
556 IIOperand->getType()->getPrimitiveSizeInBits();
557 Value *CV = ConstantInt::get(X->getType(), C);
558 Value *V = Builder->CreateLShr(X, CV);
559 return new TruncInst(V, IIOperand->getType());
564 case Intrinsic::powi:
565 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
568 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
571 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
572 // powi(x, -1) -> 1/x
573 if (Power->isAllOnesValue())
574 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
575 II->getArgOperand(0));
578 case Intrinsic::cttz: {
579 // If all bits below the first known one are known zero,
580 // this value is constant.
581 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
582 // FIXME: Try to simplify vectors of integers.
584 uint32_t BitWidth = IT->getBitWidth();
585 APInt KnownZero(BitWidth, 0);
586 APInt KnownOne(BitWidth, 0);
587 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
588 unsigned TrailingZeros = KnownOne.countTrailingZeros();
589 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
590 if ((Mask & KnownZero) == Mask)
591 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
592 APInt(BitWidth, TrailingZeros)));
596 case Intrinsic::ctlz: {
597 // If all bits above the first known one are known zero,
598 // this value is constant.
599 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
600 // FIXME: Try to simplify vectors of integers.
602 uint32_t BitWidth = IT->getBitWidth();
603 APInt KnownZero(BitWidth, 0);
604 APInt KnownOne(BitWidth, 0);
605 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
606 unsigned LeadingZeros = KnownOne.countLeadingZeros();
607 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
608 if ((Mask & KnownZero) == Mask)
609 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
610 APInt(BitWidth, LeadingZeros)));
615 case Intrinsic::uadd_with_overflow:
616 case Intrinsic::sadd_with_overflow:
617 case Intrinsic::umul_with_overflow:
618 case Intrinsic::smul_with_overflow:
619 if (isa<Constant>(II->getArgOperand(0)) &&
620 !isa<Constant>(II->getArgOperand(1))) {
621 // Canonicalize constants into the RHS.
622 Value *LHS = II->getArgOperand(0);
623 II->setArgOperand(0, II->getArgOperand(1));
624 II->setArgOperand(1, LHS);
629 case Intrinsic::usub_with_overflow:
630 case Intrinsic::ssub_with_overflow: {
631 OverflowCheckFlavor OCF =
632 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
633 assert(OCF != OCF_INVALID && "unexpected!");
635 Value *OperationResult = nullptr;
636 Constant *OverflowResult = nullptr;
637 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
638 *II, OperationResult, OverflowResult))
639 return CreateOverflowTuple(II, OperationResult, OverflowResult);
644 case Intrinsic::minnum:
645 case Intrinsic::maxnum: {
646 Value *Arg0 = II->getArgOperand(0);
647 Value *Arg1 = II->getArgOperand(1);
651 return ReplaceInstUsesWith(CI, Arg0);
653 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
654 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
656 // Canonicalize constants into the RHS.
658 II->setArgOperand(0, Arg1);
659 II->setArgOperand(1, Arg0);
664 if (C1 && C1->isNaN())
665 return ReplaceInstUsesWith(CI, Arg0);
667 // This is the value because if undef were NaN, we would return the other
668 // value and cannot return a NaN unless both operands are.
670 // fmin(undef, x) -> x
671 if (isa<UndefValue>(Arg0))
672 return ReplaceInstUsesWith(CI, Arg1);
674 // fmin(x, undef) -> x
675 if (isa<UndefValue>(Arg1))
676 return ReplaceInstUsesWith(CI, Arg0);
680 if (II->getIntrinsicID() == Intrinsic::minnum) {
681 // fmin(x, fmin(x, y)) -> fmin(x, y)
682 // fmin(y, fmin(x, y)) -> fmin(x, y)
683 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
684 if (Arg0 == X || Arg0 == Y)
685 return ReplaceInstUsesWith(CI, Arg1);
688 // fmin(fmin(x, y), x) -> fmin(x, y)
689 // fmin(fmin(x, y), y) -> fmin(x, y)
690 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
691 if (Arg1 == X || Arg1 == Y)
692 return ReplaceInstUsesWith(CI, Arg0);
695 // TODO: fmin(nnan x, inf) -> x
696 // TODO: fmin(nnan ninf x, flt_max) -> x
697 if (C1 && C1->isInfinity()) {
698 // fmin(x, -inf) -> -inf
699 if (C1->isNegative())
700 return ReplaceInstUsesWith(CI, Arg1);
703 assert(II->getIntrinsicID() == Intrinsic::maxnum);
704 // fmax(x, fmax(x, y)) -> fmax(x, y)
705 // fmax(y, fmax(x, y)) -> fmax(x, y)
706 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
707 if (Arg0 == X || Arg0 == Y)
708 return ReplaceInstUsesWith(CI, Arg1);
711 // fmax(fmax(x, y), x) -> fmax(x, y)
712 // fmax(fmax(x, y), y) -> fmax(x, y)
713 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
714 if (Arg1 == X || Arg1 == Y)
715 return ReplaceInstUsesWith(CI, Arg0);
718 // TODO: fmax(nnan x, -inf) -> x
719 // TODO: fmax(nnan ninf x, -flt_max) -> x
720 if (C1 && C1->isInfinity()) {
721 // fmax(x, inf) -> inf
722 if (!C1->isNegative())
723 return ReplaceInstUsesWith(CI, Arg1);
728 case Intrinsic::ppc_altivec_lvx:
729 case Intrinsic::ppc_altivec_lvxl:
730 // Turn PPC lvx -> load if the pointer is known aligned.
731 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
733 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
734 PointerType::getUnqual(II->getType()));
735 return new LoadInst(Ptr);
738 case Intrinsic::ppc_vsx_lxvw4x:
739 case Intrinsic::ppc_vsx_lxvd2x: {
740 // Turn PPC VSX loads into normal loads.
741 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
742 PointerType::getUnqual(II->getType()));
743 return new LoadInst(Ptr, Twine(""), false, 1);
745 case Intrinsic::ppc_altivec_stvx:
746 case Intrinsic::ppc_altivec_stvxl:
747 // Turn stvx -> store if the pointer is known aligned.
748 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
751 PointerType::getUnqual(II->getArgOperand(0)->getType());
752 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
753 return new StoreInst(II->getArgOperand(0), Ptr);
756 case Intrinsic::ppc_vsx_stxvw4x:
757 case Intrinsic::ppc_vsx_stxvd2x: {
758 // Turn PPC VSX stores into normal stores.
759 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
760 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
761 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
763 case Intrinsic::ppc_qpx_qvlfs:
764 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
765 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
767 Type *VTy = VectorType::get(Builder->getFloatTy(),
768 II->getType()->getVectorNumElements());
769 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
770 PointerType::getUnqual(VTy));
771 Value *Load = Builder->CreateLoad(Ptr);
772 return new FPExtInst(Load, II->getType());
775 case Intrinsic::ppc_qpx_qvlfd:
776 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
777 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
779 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
780 PointerType::getUnqual(II->getType()));
781 return new LoadInst(Ptr);
784 case Intrinsic::ppc_qpx_qvstfs:
785 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
786 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
788 Type *VTy = VectorType::get(Builder->getFloatTy(),
789 II->getArgOperand(0)->getType()->getVectorNumElements());
790 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
791 Type *OpPtrTy = PointerType::getUnqual(VTy);
792 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
793 return new StoreInst(TOp, Ptr);
796 case Intrinsic::ppc_qpx_qvstfd:
797 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
798 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
801 PointerType::getUnqual(II->getArgOperand(0)->getType());
802 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
803 return new StoreInst(II->getArgOperand(0), Ptr);
807 case Intrinsic::x86_sse_storeu_ps:
808 case Intrinsic::x86_sse2_storeu_pd:
809 case Intrinsic::x86_sse2_storeu_dq:
810 // Turn X86 storeu -> store if the pointer is known aligned.
811 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
814 PointerType::getUnqual(II->getArgOperand(1)->getType());
815 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
816 return new StoreInst(II->getArgOperand(1), Ptr);
820 case Intrinsic::x86_vcvtph2ps_128:
821 case Intrinsic::x86_vcvtph2ps_256: {
822 auto Arg = II->getArgOperand(0);
823 auto ArgType = cast<VectorType>(Arg->getType());
824 auto RetType = cast<VectorType>(II->getType());
825 unsigned ArgWidth = ArgType->getNumElements();
826 unsigned RetWidth = RetType->getNumElements();
827 assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
828 assert(ArgType->isIntOrIntVectorTy() &&
829 ArgType->getScalarSizeInBits() == 16 &&
830 "CVTPH2PS input type should be 16-bit integer vector");
831 assert(RetType->getScalarType()->isFloatTy() &&
832 "CVTPH2PS output type should be 32-bit float vector");
834 // Constant folding: Convert to generic half to single conversion.
835 if (isa<ConstantAggregateZero>(Arg))
836 return ReplaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
838 if (isa<ConstantDataVector>(Arg)) {
839 auto VectorHalfAsShorts = Arg;
840 if (RetWidth < ArgWidth) {
841 SmallVector<int, 8> SubVecMask;
842 for (unsigned i = 0; i != RetWidth; ++i)
843 SubVecMask.push_back((int)i);
844 VectorHalfAsShorts = Builder->CreateShuffleVector(
845 Arg, UndefValue::get(ArgType), SubVecMask);
848 auto VectorHalfType =
849 VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
851 Builder->CreateBitCast(VectorHalfAsShorts, VectorHalfType);
852 auto VectorFloats = Builder->CreateFPExt(VectorHalfs, RetType);
853 return ReplaceInstUsesWith(*II, VectorFloats);
856 // We only use the lowest lanes of the argument.
857 APInt DemandedElts = APInt::getLowBitsSet(ArgWidth, RetWidth);
858 APInt UndefElts(ArgWidth, 0);
859 if (Value *V = SimplifyDemandedVectorElts(Arg, DemandedElts, UndefElts)) {
860 II->setArgOperand(0, V);
866 case Intrinsic::x86_sse_cvtss2si:
867 case Intrinsic::x86_sse_cvtss2si64:
868 case Intrinsic::x86_sse_cvttss2si:
869 case Intrinsic::x86_sse_cvttss2si64:
870 case Intrinsic::x86_sse2_cvtsd2si:
871 case Intrinsic::x86_sse2_cvtsd2si64:
872 case Intrinsic::x86_sse2_cvttsd2si:
873 case Intrinsic::x86_sse2_cvttsd2si64: {
874 // These intrinsics only demand the 0th element of their input vectors. If
875 // we can simplify the input based on that, do so now.
877 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
878 APInt DemandedElts(VWidth, 1);
879 APInt UndefElts(VWidth, 0);
880 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
881 DemandedElts, UndefElts)) {
882 II->setArgOperand(0, V);
888 // Constant fold ashr( <A x Bi>, Ci ).
889 // Constant fold lshr( <A x Bi>, Ci ).
890 // Constant fold shl( <A x Bi>, Ci ).
891 case Intrinsic::x86_sse2_psrai_d:
892 case Intrinsic::x86_sse2_psrai_w:
893 case Intrinsic::x86_avx2_psrai_d:
894 case Intrinsic::x86_avx2_psrai_w:
895 case Intrinsic::x86_sse2_psrli_d:
896 case Intrinsic::x86_sse2_psrli_q:
897 case Intrinsic::x86_sse2_psrli_w:
898 case Intrinsic::x86_avx2_psrli_d:
899 case Intrinsic::x86_avx2_psrli_q:
900 case Intrinsic::x86_avx2_psrli_w:
901 case Intrinsic::x86_sse2_pslli_d:
902 case Intrinsic::x86_sse2_pslli_q:
903 case Intrinsic::x86_sse2_pslli_w:
904 case Intrinsic::x86_avx2_pslli_d:
905 case Intrinsic::x86_avx2_pslli_q:
906 case Intrinsic::x86_avx2_pslli_w:
907 if (Value *V = SimplifyX86immshift(*II, *Builder))
908 return ReplaceInstUsesWith(*II, V);
911 case Intrinsic::x86_sse2_psra_d:
912 case Intrinsic::x86_sse2_psra_w:
913 case Intrinsic::x86_avx2_psra_d:
914 case Intrinsic::x86_avx2_psra_w:
915 case Intrinsic::x86_sse2_psrl_d:
916 case Intrinsic::x86_sse2_psrl_q:
917 case Intrinsic::x86_sse2_psrl_w:
918 case Intrinsic::x86_avx2_psrl_d:
919 case Intrinsic::x86_avx2_psrl_q:
920 case Intrinsic::x86_avx2_psrl_w:
921 case Intrinsic::x86_sse2_psll_d:
922 case Intrinsic::x86_sse2_psll_q:
923 case Intrinsic::x86_sse2_psll_w:
924 case Intrinsic::x86_avx2_psll_d:
925 case Intrinsic::x86_avx2_psll_q:
926 case Intrinsic::x86_avx2_psll_w: {
927 if (Value *V = SimplifyX86immshift(*II, *Builder))
928 return ReplaceInstUsesWith(*II, V);
930 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
931 // operand to compute the shift amount.
932 auto ShiftAmt = II->getArgOperand(1);
933 auto ShiftType = cast<VectorType>(ShiftAmt->getType());
934 assert(ShiftType->getPrimitiveSizeInBits() == 128 &&
935 "Unexpected packed shift size");
936 unsigned VWidth = ShiftType->getNumElements();
938 APInt DemandedElts = APInt::getLowBitsSet(VWidth, VWidth / 2);
939 APInt UndefElts(VWidth, 0);
941 SimplifyDemandedVectorElts(ShiftAmt, DemandedElts, UndefElts)) {
942 II->setArgOperand(1, V);
948 case Intrinsic::x86_sse41_pmovsxbd:
949 case Intrinsic::x86_sse41_pmovsxbq:
950 case Intrinsic::x86_sse41_pmovsxbw:
951 case Intrinsic::x86_sse41_pmovsxdq:
952 case Intrinsic::x86_sse41_pmovsxwd:
953 case Intrinsic::x86_sse41_pmovsxwq:
954 case Intrinsic::x86_avx2_pmovsxbd:
955 case Intrinsic::x86_avx2_pmovsxbq:
956 case Intrinsic::x86_avx2_pmovsxbw:
957 case Intrinsic::x86_avx2_pmovsxdq:
958 case Intrinsic::x86_avx2_pmovsxwd:
959 case Intrinsic::x86_avx2_pmovsxwq:
960 if (Value *V = SimplifyX86extend(*II, *Builder, true))
961 return ReplaceInstUsesWith(*II, V);
964 case Intrinsic::x86_sse41_pmovzxbd:
965 case Intrinsic::x86_sse41_pmovzxbq:
966 case Intrinsic::x86_sse41_pmovzxbw:
967 case Intrinsic::x86_sse41_pmovzxdq:
968 case Intrinsic::x86_sse41_pmovzxwd:
969 case Intrinsic::x86_sse41_pmovzxwq:
970 case Intrinsic::x86_avx2_pmovzxbd:
971 case Intrinsic::x86_avx2_pmovzxbq:
972 case Intrinsic::x86_avx2_pmovzxbw:
973 case Intrinsic::x86_avx2_pmovzxdq:
974 case Intrinsic::x86_avx2_pmovzxwd:
975 case Intrinsic::x86_avx2_pmovzxwq:
976 if (Value *V = SimplifyX86extend(*II, *Builder, false))
977 return ReplaceInstUsesWith(*II, V);
980 case Intrinsic::x86_sse41_insertps:
981 if (Value *V = SimplifyX86insertps(*II, *Builder))
982 return ReplaceInstUsesWith(*II, V);
985 case Intrinsic::x86_sse4a_extrq: {
986 // EXTRQ uses only the lowest 64-bits of the first 128-bit vector
987 // operands and the lowest 16-bits of the second.
988 Value *Op0 = II->getArgOperand(0);
989 Value *Op1 = II->getArgOperand(1);
990 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
991 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
992 assert(VWidth0 == 2 && VWidth1 == 16 && "Unexpected operand sizes");
994 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
995 II->setArgOperand(0, V);
998 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
999 II->setArgOperand(1, V);
1005 case Intrinsic::x86_sse4a_extrqi: {
1006 // EXTRQI uses only the lowest 64-bits of the first 128-bit vector
1008 Value *Op = II->getArgOperand(0);
1009 unsigned VWidth = Op->getType()->getVectorNumElements();
1010 assert(VWidth == 2 && "Unexpected operand size");
1012 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1013 II->setArgOperand(0, V);
1019 case Intrinsic::x86_sse4a_insertq: {
1020 // INSERTQ uses only the lowest 64-bits of the first 128-bit vector
1022 Value *Op = II->getArgOperand(0);
1023 unsigned VWidth = Op->getType()->getVectorNumElements();
1024 assert(VWidth == 2 && "Unexpected operand size");
1026 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1027 II->setArgOperand(0, V);
1033 case Intrinsic::x86_sse4a_insertqi: {
1034 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
1036 // TODO: eventually we should lower this intrinsic to IR
1037 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
1038 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
1039 unsigned Index = CIIndex->getZExtValue();
1040 // From AMD documentation: "a value of zero in the field length is
1041 // defined as length of 64".
1042 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
1044 // From AMD documentation: "If the sum of the bit index + length field
1045 // is greater than 64, the results are undefined".
1046 unsigned End = Index + Length;
1048 // Note that both field index and field length are 8-bit quantities.
1049 // Since variables 'Index' and 'Length' are unsigned values
1050 // obtained from zero-extending field index and field length
1051 // respectively, their sum should never wrap around.
1053 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
1055 if (Length == 64 && Index == 0) {
1056 Value *Vec = II->getArgOperand(1);
1057 Value *Undef = UndefValue::get(Vec->getType());
1058 const uint32_t Mask[] = { 0, 2 };
1059 return ReplaceInstUsesWith(
1061 Builder->CreateShuffleVector(
1062 Vec, Undef, ConstantDataVector::get(
1063 II->getContext(), makeArrayRef(Mask))));
1064 } else if (auto Source =
1065 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1066 if (Source->hasOneUse() &&
1067 Source->getArgOperand(1) == II->getArgOperand(1)) {
1068 // If the source of the insert has only one use and it's another
1069 // insert (and they're both inserting from the same vector), try to
1070 // bundle both together.
1071 auto CISourceLength =
1072 dyn_cast<ConstantInt>(Source->getArgOperand(2));
1073 auto CISourceIndex =
1074 dyn_cast<ConstantInt>(Source->getArgOperand(3));
1075 if (CISourceIndex && CISourceLength) {
1076 unsigned SourceIndex = CISourceIndex->getZExtValue();
1077 unsigned SourceLength = CISourceLength->getZExtValue();
1078 unsigned SourceEnd = SourceIndex + SourceLength;
1079 unsigned NewIndex, NewLength;
1080 bool ShouldReplace = false;
1081 if (Index <= SourceIndex && SourceIndex <= End) {
1083 NewLength = std::max(End, SourceEnd) - NewIndex;
1084 ShouldReplace = true;
1085 } else if (SourceIndex <= Index && Index <= SourceEnd) {
1086 NewIndex = SourceIndex;
1087 NewLength = std::max(SourceEnd, End) - NewIndex;
1088 ShouldReplace = true;
1091 if (ShouldReplace) {
1092 Constant *ConstantLength = ConstantInt::get(
1093 II->getArgOperand(2)->getType(), NewLength, false);
1094 Constant *ConstantIndex = ConstantInt::get(
1095 II->getArgOperand(3)->getType(), NewIndex, false);
1096 Value *Args[4] = { Source->getArgOperand(0),
1097 II->getArgOperand(1), ConstantLength,
1099 Module *M = CI.getParent()->getParent()->getParent();
1101 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1102 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
1110 // INSERTQI uses only the lowest 64-bits of the first two 128-bit vector
1112 Value *Op0 = II->getArgOperand(0);
1113 Value *Op1 = II->getArgOperand(1);
1114 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
1115 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
1116 assert(VWidth0 == 2 && VWidth1 == 2 && "Unexpected operand sizes");
1118 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
1119 II->setArgOperand(0, V);
1123 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
1124 II->setArgOperand(1, V);
1130 case Intrinsic::x86_sse41_pblendvb:
1131 case Intrinsic::x86_sse41_blendvps:
1132 case Intrinsic::x86_sse41_blendvpd:
1133 case Intrinsic::x86_avx_blendv_ps_256:
1134 case Intrinsic::x86_avx_blendv_pd_256:
1135 case Intrinsic::x86_avx2_pblendvb: {
1136 // Convert blendv* to vector selects if the mask is constant.
1137 // This optimization is convoluted because the intrinsic is defined as
1138 // getting a vector of floats or doubles for the ps and pd versions.
1139 // FIXME: That should be changed.
1141 Value *Op0 = II->getArgOperand(0);
1142 Value *Op1 = II->getArgOperand(1);
1143 Value *Mask = II->getArgOperand(2);
1145 // fold (blend A, A, Mask) -> A
1147 return ReplaceInstUsesWith(CI, Op0);
1149 // Zero Mask - select 1st argument.
1150 if (isa<ConstantAggregateZero>(Mask))
1151 return ReplaceInstUsesWith(CI, Op0);
1153 // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
1154 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
1155 auto Tyi1 = Builder->getInt1Ty();
1156 auto SelectorType = cast<VectorType>(Mask->getType());
1157 auto EltTy = SelectorType->getElementType();
1158 unsigned Size = SelectorType->getNumElements();
1162 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
1163 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
1164 "Wrong arguments for variable blend intrinsic");
1165 SmallVector<Constant *, 32> Selectors;
1166 for (unsigned I = 0; I < Size; ++I) {
1167 // The intrinsics only read the top bit
1170 Selector = C->getElementAsInteger(I);
1172 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
1173 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
1175 auto NewSelector = ConstantVector::get(Selectors);
1176 return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
1181 case Intrinsic::x86_avx_vpermilvar_ps:
1182 case Intrinsic::x86_avx_vpermilvar_ps_256:
1183 case Intrinsic::x86_avx_vpermilvar_pd:
1184 case Intrinsic::x86_avx_vpermilvar_pd_256: {
1185 // Convert vpermil* to shufflevector if the mask is constant.
1186 Value *V = II->getArgOperand(1);
1187 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1188 assert(Size == 8 || Size == 4 || Size == 2);
1189 uint32_t Indexes[8];
1190 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1191 // The intrinsics only read one or two bits, clear the rest.
1192 for (unsigned I = 0; I < Size; ++I) {
1193 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1194 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1195 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1199 } else if (isa<ConstantAggregateZero>(V)) {
1200 for (unsigned I = 0; I < Size; ++I)
1205 // The _256 variants are a bit trickier since the mask bits always index
1206 // into the corresponding 128 half. In order to convert to a generic
1207 // shuffle, we have to make that explicit.
1208 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1209 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1210 for (unsigned I = Size / 2; I < Size; ++I)
1211 Indexes[I] += Size / 2;
1214 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1215 auto V1 = II->getArgOperand(0);
1216 auto V2 = UndefValue::get(V1->getType());
1217 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1218 return ReplaceInstUsesWith(CI, Shuffle);
1221 case Intrinsic::x86_avx_vperm2f128_pd_256:
1222 case Intrinsic::x86_avx_vperm2f128_ps_256:
1223 case Intrinsic::x86_avx_vperm2f128_si_256:
1224 case Intrinsic::x86_avx2_vperm2i128:
1225 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1226 return ReplaceInstUsesWith(*II, V);
1229 case Intrinsic::ppc_altivec_vperm:
1230 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1231 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1232 // a vectorshuffle for little endian, we must undo the transformation
1233 // performed on vec_perm in altivec.h. That is, we must complement
1234 // the permutation mask with respect to 31 and reverse the order of
1236 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1237 assert(Mask->getType()->getVectorNumElements() == 16 &&
1238 "Bad type for intrinsic!");
1240 // Check that all of the elements are integer constants or undefs.
1241 bool AllEltsOk = true;
1242 for (unsigned i = 0; i != 16; ++i) {
1243 Constant *Elt = Mask->getAggregateElement(i);
1244 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1251 // Cast the input vectors to byte vectors.
1252 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1254 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1256 Value *Result = UndefValue::get(Op0->getType());
1258 // Only extract each element once.
1259 Value *ExtractedElts[32];
1260 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1262 for (unsigned i = 0; i != 16; ++i) {
1263 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1266 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1267 Idx &= 31; // Match the hardware behavior.
1268 if (DL.isLittleEndian())
1271 if (!ExtractedElts[Idx]) {
1272 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1273 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1274 ExtractedElts[Idx] =
1275 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1276 Builder->getInt32(Idx&15));
1279 // Insert this value into the result vector.
1280 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1281 Builder->getInt32(i));
1283 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1288 case Intrinsic::arm_neon_vld1:
1289 case Intrinsic::arm_neon_vld2:
1290 case Intrinsic::arm_neon_vld3:
1291 case Intrinsic::arm_neon_vld4:
1292 case Intrinsic::arm_neon_vld2lane:
1293 case Intrinsic::arm_neon_vld3lane:
1294 case Intrinsic::arm_neon_vld4lane:
1295 case Intrinsic::arm_neon_vst1:
1296 case Intrinsic::arm_neon_vst2:
1297 case Intrinsic::arm_neon_vst3:
1298 case Intrinsic::arm_neon_vst4:
1299 case Intrinsic::arm_neon_vst2lane:
1300 case Intrinsic::arm_neon_vst3lane:
1301 case Intrinsic::arm_neon_vst4lane: {
1302 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1303 unsigned AlignArg = II->getNumArgOperands() - 1;
1304 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1305 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1306 II->setArgOperand(AlignArg,
1307 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1314 case Intrinsic::arm_neon_vmulls:
1315 case Intrinsic::arm_neon_vmullu:
1316 case Intrinsic::aarch64_neon_smull:
1317 case Intrinsic::aarch64_neon_umull: {
1318 Value *Arg0 = II->getArgOperand(0);
1319 Value *Arg1 = II->getArgOperand(1);
1321 // Handle mul by zero first:
1322 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1323 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1326 // Check for constant LHS & RHS - in this case we just simplify.
1327 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1328 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1329 VectorType *NewVT = cast<VectorType>(II->getType());
1330 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1331 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1332 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1333 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1335 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1338 // Couldn't simplify - canonicalize constant to the RHS.
1339 std::swap(Arg0, Arg1);
1342 // Handle mul by one:
1343 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1344 if (ConstantInt *Splat =
1345 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1347 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1348 /*isSigned=*/!Zext);
1353 case Intrinsic::AMDGPU_rcp: {
1354 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1355 const APFloat &ArgVal = C->getValueAPF();
1356 APFloat Val(ArgVal.getSemantics(), 1.0);
1357 APFloat::opStatus Status = Val.divide(ArgVal,
1358 APFloat::rmNearestTiesToEven);
1359 // Only do this if it was exact and therefore not dependent on the
1361 if (Status == APFloat::opOK)
1362 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1367 case Intrinsic::stackrestore: {
1368 // If the save is right next to the restore, remove the restore. This can
1369 // happen when variable allocas are DCE'd.
1370 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1371 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1372 BasicBlock::iterator BI = SS;
1374 return EraseInstFromFunction(CI);
1378 // Scan down this block to see if there is another stack restore in the
1379 // same block without an intervening call/alloca.
1380 BasicBlock::iterator BI = II;
1381 TerminatorInst *TI = II->getParent()->getTerminator();
1382 bool CannotRemove = false;
1383 for (++BI; &*BI != TI; ++BI) {
1384 if (isa<AllocaInst>(BI)) {
1385 CannotRemove = true;
1388 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1389 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1390 // If there is a stackrestore below this one, remove this one.
1391 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1392 return EraseInstFromFunction(CI);
1393 // Otherwise, ignore the intrinsic.
1395 // If we found a non-intrinsic call, we can't remove the stack
1397 CannotRemove = true;
1403 // If the stack restore is in a return, resume, or unwind block and if there
1404 // are no allocas or calls between the restore and the return, nuke the
1406 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1407 return EraseInstFromFunction(CI);
1410 case Intrinsic::assume: {
1411 // Canonicalize assume(a && b) -> assume(a); assume(b);
1412 // Note: New assumption intrinsics created here are registered by
1413 // the InstCombineIRInserter object.
1414 Value *IIOperand = II->getArgOperand(0), *A, *B,
1415 *AssumeIntrinsic = II->getCalledValue();
1416 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1417 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1418 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1419 return EraseInstFromFunction(*II);
1421 // assume(!(a || b)) -> assume(!a); assume(!b);
1422 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1423 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1425 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1427 return EraseInstFromFunction(*II);
1430 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1431 // (if assume is valid at the load)
1432 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1433 Value *LHS = ICmp->getOperand(0);
1434 Value *RHS = ICmp->getOperand(1);
1435 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1436 isa<LoadInst>(LHS) &&
1437 isa<Constant>(RHS) &&
1438 RHS->getType()->isPointerTy() &&
1439 cast<Constant>(RHS)->isNullValue()) {
1440 LoadInst* LI = cast<LoadInst>(LHS);
1441 if (isValidAssumeForContext(II, LI, DT)) {
1442 MDNode *MD = MDNode::get(II->getContext(), None);
1443 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1444 return EraseInstFromFunction(*II);
1447 // TODO: apply nonnull return attributes to calls and invokes
1448 // TODO: apply range metadata for range check patterns?
1450 // If there is a dominating assume with the same condition as this one,
1451 // then this one is redundant, and should be removed.
1452 APInt KnownZero(1, 0), KnownOne(1, 0);
1453 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1454 if (KnownOne.isAllOnesValue())
1455 return EraseInstFromFunction(*II);
1459 case Intrinsic::experimental_gc_relocate: {
1460 // Translate facts known about a pointer before relocating into
1461 // facts about the relocate value, while being careful to
1462 // preserve relocation semantics.
1463 GCRelocateOperands Operands(II);
1464 Value *DerivedPtr = Operands.getDerivedPtr();
1465 auto *GCRelocateType = cast<PointerType>(II->getType());
1467 // Remove the relocation if unused, note that this check is required
1468 // to prevent the cases below from looping forever.
1469 if (II->use_empty())
1470 return EraseInstFromFunction(*II);
1472 // Undef is undef, even after relocation.
1473 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1474 // most practical collectors, but there was discussion in the review thread
1475 // about whether it was legal for all possible collectors.
1476 if (isa<UndefValue>(DerivedPtr)) {
1477 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1478 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1481 // The relocation of null will be null for most any collector.
1482 // TODO: provide a hook for this in GCStrategy. There might be some weird
1483 // collector this property does not hold for.
1484 if (isa<ConstantPointerNull>(DerivedPtr)) {
1485 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1486 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1489 // isKnownNonNull -> nonnull attribute
1490 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI))
1491 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1493 // isDereferenceablePointer -> deref attribute
1494 if (isDereferenceablePointer(DerivedPtr, DL)) {
1495 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1496 uint64_t Bytes = A->getDereferenceableBytes();
1497 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1501 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1502 // Canonicalize on the type from the uses to the defs
1504 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1508 return visitCallSite(II);
1511 // InvokeInst simplification
1513 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1514 return visitCallSite(&II);
1517 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1518 /// passed through the varargs area, we can eliminate the use of the cast.
1519 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1520 const DataLayout &DL,
1521 const CastInst *const CI,
1523 if (!CI->isLosslessCast())
1526 // If this is a GC intrinsic, avoid munging types. We need types for
1527 // statepoint reconstruction in SelectionDAG.
1528 // TODO: This is probably something which should be expanded to all
1529 // intrinsics since the entire point of intrinsics is that
1530 // they are understandable by the optimizer.
1531 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1534 // The size of ByVal or InAlloca arguments is derived from the type, so we
1535 // can't change to a type with a different size. If the size were
1536 // passed explicitly we could avoid this check.
1537 if (!CS.isByValOrInAllocaArgument(ix))
1541 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1542 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1543 if (!SrcTy->isSized() || !DstTy->isSized())
1545 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1550 // Try to fold some different type of calls here.
1551 // Currently we're only working with the checking functions, memcpy_chk,
1552 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1553 // strcat_chk and strncat_chk.
1554 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1555 if (!CI->getCalledFunction()) return nullptr;
1557 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1558 ReplaceInstUsesWith(*From, With);
1560 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1561 if (Value *With = Simplifier.optimizeCall(CI)) {
1563 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1569 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1570 // Strip off at most one level of pointer casts, looking for an alloca. This
1571 // is good enough in practice and simpler than handling any number of casts.
1572 Value *Underlying = TrampMem->stripPointerCasts();
1573 if (Underlying != TrampMem &&
1574 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1576 if (!isa<AllocaInst>(Underlying))
1579 IntrinsicInst *InitTrampoline = nullptr;
1580 for (User *U : TrampMem->users()) {
1581 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1584 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1586 // More than one init_trampoline writes to this value. Give up.
1588 InitTrampoline = II;
1591 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1592 // Allow any number of calls to adjust.trampoline.
1597 // No call to init.trampoline found.
1598 if (!InitTrampoline)
1601 // Check that the alloca is being used in the expected way.
1602 if (InitTrampoline->getOperand(0) != TrampMem)
1605 return InitTrampoline;
1608 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1610 // Visit all the previous instructions in the basic block, and try to find a
1611 // init.trampoline which has a direct path to the adjust.trampoline.
1612 for (BasicBlock::iterator I = AdjustTramp,
1613 E = AdjustTramp->getParent()->begin(); I != E; ) {
1614 Instruction *Inst = --I;
1615 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1616 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1617 II->getOperand(0) == TrampMem)
1619 if (Inst->mayWriteToMemory())
1625 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1626 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1627 // to a direct call to a function. Otherwise return NULL.
1629 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1630 Callee = Callee->stripPointerCasts();
1631 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1633 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1636 Value *TrampMem = AdjustTramp->getOperand(0);
1638 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1640 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1645 // visitCallSite - Improvements for call and invoke instructions.
1647 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1649 if (isAllocLikeFn(CS.getInstruction(), TLI))
1650 return visitAllocSite(*CS.getInstruction());
1652 bool Changed = false;
1654 // Mark any parameters that are known to be non-null with the nonnull
1655 // attribute. This is helpful for inlining calls to functions with null
1656 // checks on their arguments.
1658 for (Value *V : CS.args()) {
1659 if (V->getType()->isPointerTy() && !CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1660 isKnownNonNullAt(V, CS.getInstruction(), DT, TLI)) {
1661 AttributeSet AS = CS.getAttributes();
1662 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1663 Attribute::NonNull);
1664 CS.setAttributes(AS);
1669 assert(ArgNo == CS.arg_size() && "sanity check");
1671 // If the callee is a pointer to a function, attempt to move any casts to the
1672 // arguments of the call/invoke.
1673 Value *Callee = CS.getCalledValue();
1674 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1677 if (Function *CalleeF = dyn_cast<Function>(Callee))
1678 // If the call and callee calling conventions don't match, this call must
1679 // be unreachable, as the call is undefined.
1680 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1681 // Only do this for calls to a function with a body. A prototype may
1682 // not actually end up matching the implementation's calling conv for a
1683 // variety of reasons (e.g. it may be written in assembly).
1684 !CalleeF->isDeclaration()) {
1685 Instruction *OldCall = CS.getInstruction();
1686 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1687 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1689 // If OldCall does not return void then replaceAllUsesWith undef.
1690 // This allows ValueHandlers and custom metadata to adjust itself.
1691 if (!OldCall->getType()->isVoidTy())
1692 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1693 if (isa<CallInst>(OldCall))
1694 return EraseInstFromFunction(*OldCall);
1696 // We cannot remove an invoke, because it would change the CFG, just
1697 // change the callee to a null pointer.
1698 cast<InvokeInst>(OldCall)->setCalledFunction(
1699 Constant::getNullValue(CalleeF->getType()));
1703 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1704 // If CS does not return void then replaceAllUsesWith undef.
1705 // This allows ValueHandlers and custom metadata to adjust itself.
1706 if (!CS.getInstruction()->getType()->isVoidTy())
1707 ReplaceInstUsesWith(*CS.getInstruction(),
1708 UndefValue::get(CS.getInstruction()->getType()));
1710 if (isa<InvokeInst>(CS.getInstruction())) {
1711 // Can't remove an invoke because we cannot change the CFG.
1715 // This instruction is not reachable, just remove it. We insert a store to
1716 // undef so that we know that this code is not reachable, despite the fact
1717 // that we can't modify the CFG here.
1718 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1719 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1720 CS.getInstruction());
1722 return EraseInstFromFunction(*CS.getInstruction());
1725 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1726 return transformCallThroughTrampoline(CS, II);
1728 PointerType *PTy = cast<PointerType>(Callee->getType());
1729 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1730 if (FTy->isVarArg()) {
1731 int ix = FTy->getNumParams();
1732 // See if we can optimize any arguments passed through the varargs area of
1734 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1735 E = CS.arg_end(); I != E; ++I, ++ix) {
1736 CastInst *CI = dyn_cast<CastInst>(*I);
1737 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1738 *I = CI->getOperand(0);
1744 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1745 // Inline asm calls cannot throw - mark them 'nounwind'.
1746 CS.setDoesNotThrow();
1750 // Try to optimize the call if possible, we require DataLayout for most of
1751 // this. None of these calls are seen as possibly dead so go ahead and
1752 // delete the instruction now.
1753 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1754 Instruction *I = tryOptimizeCall(CI);
1755 // If we changed something return the result, etc. Otherwise let
1756 // the fallthrough check.
1757 if (I) return EraseInstFromFunction(*I);
1760 return Changed ? CS.getInstruction() : nullptr;
1763 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1764 // attempt to move the cast to the arguments of the call/invoke.
1766 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1768 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1771 // The prototype of thunks are a lie, don't try to directly call such
1773 if (Callee->hasFnAttribute("thunk"))
1775 Instruction *Caller = CS.getInstruction();
1776 const AttributeSet &CallerPAL = CS.getAttributes();
1778 // Okay, this is a cast from a function to a different type. Unless doing so
1779 // would cause a type conversion of one of our arguments, change this call to
1780 // be a direct call with arguments casted to the appropriate types.
1782 FunctionType *FT = Callee->getFunctionType();
1783 Type *OldRetTy = Caller->getType();
1784 Type *NewRetTy = FT->getReturnType();
1786 // Check to see if we are changing the return type...
1787 if (OldRetTy != NewRetTy) {
1789 if (NewRetTy->isStructTy())
1790 return false; // TODO: Handle multiple return values.
1792 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1793 if (Callee->isDeclaration())
1794 return false; // Cannot transform this return value.
1796 if (!Caller->use_empty() &&
1797 // void -> non-void is handled specially
1798 !NewRetTy->isVoidTy())
1799 return false; // Cannot transform this return value.
1802 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1803 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1804 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1805 return false; // Attribute not compatible with transformed value.
1808 // If the callsite is an invoke instruction, and the return value is used by
1809 // a PHI node in a successor, we cannot change the return type of the call
1810 // because there is no place to put the cast instruction (without breaking
1811 // the critical edge). Bail out in this case.
1812 if (!Caller->use_empty())
1813 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1814 for (User *U : II->users())
1815 if (PHINode *PN = dyn_cast<PHINode>(U))
1816 if (PN->getParent() == II->getNormalDest() ||
1817 PN->getParent() == II->getUnwindDest())
1821 unsigned NumActualArgs = CS.arg_size();
1822 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1824 // Prevent us turning:
1825 // declare void @takes_i32_inalloca(i32* inalloca)
1826 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1829 // call void @takes_i32_inalloca(i32* null)
1831 // Similarly, avoid folding away bitcasts of byval calls.
1832 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1833 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1836 CallSite::arg_iterator AI = CS.arg_begin();
1837 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1838 Type *ParamTy = FT->getParamType(i);
1839 Type *ActTy = (*AI)->getType();
1841 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1842 return false; // Cannot transform this parameter value.
1844 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1845 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1846 return false; // Attribute not compatible with transformed value.
1848 if (CS.isInAllocaArgument(i))
1849 return false; // Cannot transform to and from inalloca.
1851 // If the parameter is passed as a byval argument, then we have to have a
1852 // sized type and the sized type has to have the same size as the old type.
1853 if (ParamTy != ActTy &&
1854 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1855 Attribute::ByVal)) {
1856 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1857 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1860 Type *CurElTy = ActTy->getPointerElementType();
1861 if (DL.getTypeAllocSize(CurElTy) !=
1862 DL.getTypeAllocSize(ParamPTy->getElementType()))
1867 if (Callee->isDeclaration()) {
1868 // Do not delete arguments unless we have a function body.
1869 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1872 // If the callee is just a declaration, don't change the varargsness of the
1873 // call. We don't want to introduce a varargs call where one doesn't
1875 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1876 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1879 // If both the callee and the cast type are varargs, we still have to make
1880 // sure the number of fixed parameters are the same or we have the same
1881 // ABI issues as if we introduce a varargs call.
1882 if (FT->isVarArg() &&
1883 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1884 FT->getNumParams() !=
1885 cast<FunctionType>(APTy->getElementType())->getNumParams())
1889 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1890 !CallerPAL.isEmpty())
1891 // In this case we have more arguments than the new function type, but we
1892 // won't be dropping them. Check that these extra arguments have attributes
1893 // that are compatible with being a vararg call argument.
1894 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1895 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1896 if (Index <= FT->getNumParams())
1899 // Check if it has an attribute that's incompatible with varargs.
1900 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1901 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1906 // Okay, we decided that this is a safe thing to do: go ahead and start
1907 // inserting cast instructions as necessary.
1908 std::vector<Value*> Args;
1909 Args.reserve(NumActualArgs);
1910 SmallVector<AttributeSet, 8> attrVec;
1911 attrVec.reserve(NumCommonArgs);
1913 // Get any return attributes.
1914 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1916 // If the return value is not being used, the type may not be compatible
1917 // with the existing attributes. Wipe out any problematic attributes.
1918 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1920 // Add the new return attributes.
1921 if (RAttrs.hasAttributes())
1922 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1923 AttributeSet::ReturnIndex, RAttrs));
1925 AI = CS.arg_begin();
1926 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1927 Type *ParamTy = FT->getParamType(i);
1929 if ((*AI)->getType() == ParamTy) {
1930 Args.push_back(*AI);
1932 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1935 // Add any parameter attributes.
1936 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1937 if (PAttrs.hasAttributes())
1938 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1942 // If the function takes more arguments than the call was taking, add them
1944 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1945 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1947 // If we are removing arguments to the function, emit an obnoxious warning.
1948 if (FT->getNumParams() < NumActualArgs) {
1949 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1950 if (FT->isVarArg()) {
1951 // Add all of the arguments in their promoted form to the arg list.
1952 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1953 Type *PTy = getPromotedType((*AI)->getType());
1954 if (PTy != (*AI)->getType()) {
1955 // Must promote to pass through va_arg area!
1956 Instruction::CastOps opcode =
1957 CastInst::getCastOpcode(*AI, false, PTy, false);
1958 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1960 Args.push_back(*AI);
1963 // Add any parameter attributes.
1964 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1965 if (PAttrs.hasAttributes())
1966 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1972 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1973 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1974 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1976 if (NewRetTy->isVoidTy())
1977 Caller->setName(""); // Void type should not have a name.
1979 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1983 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1984 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1985 II->getUnwindDest(), Args);
1987 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1988 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1990 CallInst *CI = cast<CallInst>(Caller);
1991 NC = Builder->CreateCall(Callee, Args);
1993 if (CI->isTailCall())
1994 cast<CallInst>(NC)->setTailCall();
1995 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1996 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1999 // Insert a cast of the return type as necessary.
2001 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
2002 if (!NV->getType()->isVoidTy()) {
2003 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
2004 NC->setDebugLoc(Caller->getDebugLoc());
2006 // If this is an invoke instruction, we should insert it after the first
2007 // non-phi, instruction in the normal successor block.
2008 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2009 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
2010 InsertNewInstBefore(NC, *I);
2012 // Otherwise, it's a call, just insert cast right after the call.
2013 InsertNewInstBefore(NC, *Caller);
2015 Worklist.AddUsersToWorkList(*Caller);
2017 NV = UndefValue::get(Caller->getType());
2021 if (!Caller->use_empty())
2022 ReplaceInstUsesWith(*Caller, NV);
2023 else if (Caller->hasValueHandle()) {
2024 if (OldRetTy == NV->getType())
2025 ValueHandleBase::ValueIsRAUWd(Caller, NV);
2027 // We cannot call ValueIsRAUWd with a different type, and the
2028 // actual tracked value will disappear.
2029 ValueHandleBase::ValueIsDeleted(Caller);
2032 EraseInstFromFunction(*Caller);
2036 // transformCallThroughTrampoline - Turn a call to a function created by
2037 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
2038 // underlying function.
2041 InstCombiner::transformCallThroughTrampoline(CallSite CS,
2042 IntrinsicInst *Tramp) {
2043 Value *Callee = CS.getCalledValue();
2044 PointerType *PTy = cast<PointerType>(Callee->getType());
2045 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
2046 const AttributeSet &Attrs = CS.getAttributes();
2048 // If the call already has the 'nest' attribute somewhere then give up -
2049 // otherwise 'nest' would occur twice after splicing in the chain.
2050 if (Attrs.hasAttrSomewhere(Attribute::Nest))
2054 "transformCallThroughTrampoline called with incorrect CallSite.");
2056 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
2057 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
2058 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
2060 const AttributeSet &NestAttrs = NestF->getAttributes();
2061 if (!NestAttrs.isEmpty()) {
2062 unsigned NestIdx = 1;
2063 Type *NestTy = nullptr;
2064 AttributeSet NestAttr;
2066 // Look for a parameter marked with the 'nest' attribute.
2067 for (FunctionType::param_iterator I = NestFTy->param_begin(),
2068 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
2069 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
2070 // Record the parameter type and any other attributes.
2072 NestAttr = NestAttrs.getParamAttributes(NestIdx);
2077 Instruction *Caller = CS.getInstruction();
2078 std::vector<Value*> NewArgs;
2079 NewArgs.reserve(CS.arg_size() + 1);
2081 SmallVector<AttributeSet, 8> NewAttrs;
2082 NewAttrs.reserve(Attrs.getNumSlots() + 1);
2084 // Insert the nest argument into the call argument list, which may
2085 // mean appending it. Likewise for attributes.
2087 // Add any result attributes.
2088 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
2089 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2090 Attrs.getRetAttributes()));
2094 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
2096 if (Idx == NestIdx) {
2097 // Add the chain argument and attributes.
2098 Value *NestVal = Tramp->getArgOperand(2);
2099 if (NestVal->getType() != NestTy)
2100 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
2101 NewArgs.push_back(NestVal);
2102 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2109 // Add the original argument and attributes.
2110 NewArgs.push_back(*I);
2111 AttributeSet Attr = Attrs.getParamAttributes(Idx);
2112 if (Attr.hasAttributes(Idx)) {
2113 AttrBuilder B(Attr, Idx);
2114 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2115 Idx + (Idx >= NestIdx), B));
2122 // Add any function attributes.
2123 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
2124 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
2125 Attrs.getFnAttributes()));
2127 // The trampoline may have been bitcast to a bogus type (FTy).
2128 // Handle this by synthesizing a new function type, equal to FTy
2129 // with the chain parameter inserted.
2131 std::vector<Type*> NewTypes;
2132 NewTypes.reserve(FTy->getNumParams()+1);
2134 // Insert the chain's type into the list of parameter types, which may
2135 // mean appending it.
2138 FunctionType::param_iterator I = FTy->param_begin(),
2139 E = FTy->param_end();
2143 // Add the chain's type.
2144 NewTypes.push_back(NestTy);
2149 // Add the original type.
2150 NewTypes.push_back(*I);
2156 // Replace the trampoline call with a direct call. Let the generic
2157 // code sort out any function type mismatches.
2158 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
2160 Constant *NewCallee =
2161 NestF->getType() == PointerType::getUnqual(NewFTy) ?
2162 NestF : ConstantExpr::getBitCast(NestF,
2163 PointerType::getUnqual(NewFTy));
2164 const AttributeSet &NewPAL =
2165 AttributeSet::get(FTy->getContext(), NewAttrs);
2167 Instruction *NewCaller;
2168 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2169 NewCaller = InvokeInst::Create(NewCallee,
2170 II->getNormalDest(), II->getUnwindDest(),
2172 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
2173 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
2175 NewCaller = CallInst::Create(NewCallee, NewArgs);
2176 if (cast<CallInst>(Caller)->isTailCall())
2177 cast<CallInst>(NewCaller)->setTailCall();
2178 cast<CallInst>(NewCaller)->
2179 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
2180 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
2187 // Replace the trampoline call with a direct call. Since there is no 'nest'
2188 // parameter, there is no need to adjust the argument list. Let the generic
2189 // code sort out any function type mismatches.
2190 Constant *NewCallee =
2191 NestF->getType() == PTy ? NestF :
2192 ConstantExpr::getBitCast(NestF, PTy);
2193 CS.setCalledFunction(NewCallee);
2194 return CS.getInstruction();