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 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) {
858 II->setArgOperand(0, V);
864 case Intrinsic::x86_sse_cvtss2si:
865 case Intrinsic::x86_sse_cvtss2si64:
866 case Intrinsic::x86_sse_cvttss2si:
867 case Intrinsic::x86_sse_cvttss2si64:
868 case Intrinsic::x86_sse2_cvtsd2si:
869 case Intrinsic::x86_sse2_cvtsd2si64:
870 case Intrinsic::x86_sse2_cvttsd2si:
871 case Intrinsic::x86_sse2_cvttsd2si64: {
872 // These intrinsics only demand the 0th element of their input vectors. If
873 // we can simplify the input based on that, do so now.
874 Value *Arg = II->getArgOperand(0);
875 unsigned VWidth = Arg->getType()->getVectorNumElements();
876 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
877 II->setArgOperand(0, V);
883 // Constant fold ashr( <A x Bi>, Ci ).
884 // Constant fold lshr( <A x Bi>, Ci ).
885 // Constant fold shl( <A x Bi>, Ci ).
886 case Intrinsic::x86_sse2_psrai_d:
887 case Intrinsic::x86_sse2_psrai_w:
888 case Intrinsic::x86_avx2_psrai_d:
889 case Intrinsic::x86_avx2_psrai_w:
890 case Intrinsic::x86_sse2_psrli_d:
891 case Intrinsic::x86_sse2_psrli_q:
892 case Intrinsic::x86_sse2_psrli_w:
893 case Intrinsic::x86_avx2_psrli_d:
894 case Intrinsic::x86_avx2_psrli_q:
895 case Intrinsic::x86_avx2_psrli_w:
896 case Intrinsic::x86_sse2_pslli_d:
897 case Intrinsic::x86_sse2_pslli_q:
898 case Intrinsic::x86_sse2_pslli_w:
899 case Intrinsic::x86_avx2_pslli_d:
900 case Intrinsic::x86_avx2_pslli_q:
901 case Intrinsic::x86_avx2_pslli_w:
902 if (Value *V = SimplifyX86immshift(*II, *Builder))
903 return ReplaceInstUsesWith(*II, V);
906 case Intrinsic::x86_sse2_psra_d:
907 case Intrinsic::x86_sse2_psra_w:
908 case Intrinsic::x86_avx2_psra_d:
909 case Intrinsic::x86_avx2_psra_w:
910 case Intrinsic::x86_sse2_psrl_d:
911 case Intrinsic::x86_sse2_psrl_q:
912 case Intrinsic::x86_sse2_psrl_w:
913 case Intrinsic::x86_avx2_psrl_d:
914 case Intrinsic::x86_avx2_psrl_q:
915 case Intrinsic::x86_avx2_psrl_w:
916 case Intrinsic::x86_sse2_psll_d:
917 case Intrinsic::x86_sse2_psll_q:
918 case Intrinsic::x86_sse2_psll_w:
919 case Intrinsic::x86_avx2_psll_d:
920 case Intrinsic::x86_avx2_psll_q:
921 case Intrinsic::x86_avx2_psll_w: {
922 if (Value *V = SimplifyX86immshift(*II, *Builder))
923 return ReplaceInstUsesWith(*II, V);
925 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
926 // operand to compute the shift amount.
927 Value *Arg1 = II->getArgOperand(1);
928 assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 &&
929 "Unexpected packed shift size");
930 unsigned VWidth = Arg1->getType()->getVectorNumElements();
932 if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) {
933 II->setArgOperand(1, V);
939 case Intrinsic::x86_sse41_pmovsxbd:
940 case Intrinsic::x86_sse41_pmovsxbq:
941 case Intrinsic::x86_sse41_pmovsxbw:
942 case Intrinsic::x86_sse41_pmovsxdq:
943 case Intrinsic::x86_sse41_pmovsxwd:
944 case Intrinsic::x86_sse41_pmovsxwq:
945 case Intrinsic::x86_avx2_pmovsxbd:
946 case Intrinsic::x86_avx2_pmovsxbq:
947 case Intrinsic::x86_avx2_pmovsxbw:
948 case Intrinsic::x86_avx2_pmovsxdq:
949 case Intrinsic::x86_avx2_pmovsxwd:
950 case Intrinsic::x86_avx2_pmovsxwq:
951 if (Value *V = SimplifyX86extend(*II, *Builder, true))
952 return ReplaceInstUsesWith(*II, V);
955 case Intrinsic::x86_sse41_pmovzxbd:
956 case Intrinsic::x86_sse41_pmovzxbq:
957 case Intrinsic::x86_sse41_pmovzxbw:
958 case Intrinsic::x86_sse41_pmovzxdq:
959 case Intrinsic::x86_sse41_pmovzxwd:
960 case Intrinsic::x86_sse41_pmovzxwq:
961 case Intrinsic::x86_avx2_pmovzxbd:
962 case Intrinsic::x86_avx2_pmovzxbq:
963 case Intrinsic::x86_avx2_pmovzxbw:
964 case Intrinsic::x86_avx2_pmovzxdq:
965 case Intrinsic::x86_avx2_pmovzxwd:
966 case Intrinsic::x86_avx2_pmovzxwq:
967 if (Value *V = SimplifyX86extend(*II, *Builder, false))
968 return ReplaceInstUsesWith(*II, V);
971 case Intrinsic::x86_sse41_insertps:
972 if (Value *V = SimplifyX86insertps(*II, *Builder))
973 return ReplaceInstUsesWith(*II, V);
976 case Intrinsic::x86_sse4a_extrq: {
977 // EXTRQ uses only the lowest 64-bits of the first 128-bit vector
978 // operands and the lowest 16-bits of the second.
979 Value *Op0 = II->getArgOperand(0);
980 Value *Op1 = II->getArgOperand(1);
981 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
982 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
983 assert(VWidth0 == 2 && VWidth1 == 16 && "Unexpected operand sizes");
985 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
986 II->setArgOperand(0, V);
989 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
990 II->setArgOperand(1, V);
996 case Intrinsic::x86_sse4a_extrqi: {
997 // EXTRQI uses only the lowest 64-bits of the first 128-bit vector
999 Value *Op = II->getArgOperand(0);
1000 unsigned VWidth = Op->getType()->getVectorNumElements();
1001 assert(VWidth == 2 && "Unexpected operand size");
1003 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1004 II->setArgOperand(0, V);
1010 case Intrinsic::x86_sse4a_insertq: {
1011 // INSERTQ uses only the lowest 64-bits of the first 128-bit vector
1013 Value *Op = II->getArgOperand(0);
1014 unsigned VWidth = Op->getType()->getVectorNumElements();
1015 assert(VWidth == 2 && "Unexpected operand size");
1017 if (Value *V = SimplifyDemandedVectorEltsLow(Op, VWidth, 1)) {
1018 II->setArgOperand(0, V);
1024 case Intrinsic::x86_sse4a_insertqi: {
1025 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
1027 // TODO: eventually we should lower this intrinsic to IR
1028 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
1029 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
1030 unsigned Index = CIIndex->getZExtValue();
1031 // From AMD documentation: "a value of zero in the field length is
1032 // defined as length of 64".
1033 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
1035 // From AMD documentation: "If the sum of the bit index + length field
1036 // is greater than 64, the results are undefined".
1037 unsigned End = Index + Length;
1039 // Note that both field index and field length are 8-bit quantities.
1040 // Since variables 'Index' and 'Length' are unsigned values
1041 // obtained from zero-extending field index and field length
1042 // respectively, their sum should never wrap around.
1044 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
1046 if (Length == 64 && Index == 0) {
1047 Value *Vec = II->getArgOperand(1);
1048 Value *Undef = UndefValue::get(Vec->getType());
1049 const uint32_t Mask[] = { 0, 2 };
1050 return ReplaceInstUsesWith(
1052 Builder->CreateShuffleVector(
1053 Vec, Undef, ConstantDataVector::get(
1054 II->getContext(), makeArrayRef(Mask))));
1055 } else if (auto Source =
1056 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1057 if (Source->hasOneUse() &&
1058 Source->getArgOperand(1) == II->getArgOperand(1)) {
1059 // If the source of the insert has only one use and it's another
1060 // insert (and they're both inserting from the same vector), try to
1061 // bundle both together.
1062 auto CISourceLength =
1063 dyn_cast<ConstantInt>(Source->getArgOperand(2));
1064 auto CISourceIndex =
1065 dyn_cast<ConstantInt>(Source->getArgOperand(3));
1066 if (CISourceIndex && CISourceLength) {
1067 unsigned SourceIndex = CISourceIndex->getZExtValue();
1068 unsigned SourceLength = CISourceLength->getZExtValue();
1069 unsigned SourceEnd = SourceIndex + SourceLength;
1070 unsigned NewIndex, NewLength;
1071 bool ShouldReplace = false;
1072 if (Index <= SourceIndex && SourceIndex <= End) {
1074 NewLength = std::max(End, SourceEnd) - NewIndex;
1075 ShouldReplace = true;
1076 } else if (SourceIndex <= Index && Index <= SourceEnd) {
1077 NewIndex = SourceIndex;
1078 NewLength = std::max(SourceEnd, End) - NewIndex;
1079 ShouldReplace = true;
1082 if (ShouldReplace) {
1083 Constant *ConstantLength = ConstantInt::get(
1084 II->getArgOperand(2)->getType(), NewLength, false);
1085 Constant *ConstantIndex = ConstantInt::get(
1086 II->getArgOperand(3)->getType(), NewIndex, false);
1087 Value *Args[4] = { Source->getArgOperand(0),
1088 II->getArgOperand(1), ConstantLength,
1090 Module *M = CI.getParent()->getParent()->getParent();
1092 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1093 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
1101 // INSERTQI uses only the lowest 64-bits of the first two 128-bit vector
1103 Value *Op0 = II->getArgOperand(0);
1104 Value *Op1 = II->getArgOperand(1);
1105 unsigned VWidth0 = Op0->getType()->getVectorNumElements();
1106 unsigned VWidth1 = Op1->getType()->getVectorNumElements();
1107 assert(VWidth0 == 2 && VWidth1 == 2 && "Unexpected operand sizes");
1109 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
1110 II->setArgOperand(0, V);
1114 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
1115 II->setArgOperand(1, V);
1121 case Intrinsic::x86_sse41_pblendvb:
1122 case Intrinsic::x86_sse41_blendvps:
1123 case Intrinsic::x86_sse41_blendvpd:
1124 case Intrinsic::x86_avx_blendv_ps_256:
1125 case Intrinsic::x86_avx_blendv_pd_256:
1126 case Intrinsic::x86_avx2_pblendvb: {
1127 // Convert blendv* to vector selects if the mask is constant.
1128 // This optimization is convoluted because the intrinsic is defined as
1129 // getting a vector of floats or doubles for the ps and pd versions.
1130 // FIXME: That should be changed.
1132 Value *Op0 = II->getArgOperand(0);
1133 Value *Op1 = II->getArgOperand(1);
1134 Value *Mask = II->getArgOperand(2);
1136 // fold (blend A, A, Mask) -> A
1138 return ReplaceInstUsesWith(CI, Op0);
1140 // Zero Mask - select 1st argument.
1141 if (isa<ConstantAggregateZero>(Mask))
1142 return ReplaceInstUsesWith(CI, Op0);
1144 // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
1145 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
1146 auto Tyi1 = Builder->getInt1Ty();
1147 auto SelectorType = cast<VectorType>(Mask->getType());
1148 auto EltTy = SelectorType->getElementType();
1149 unsigned Size = SelectorType->getNumElements();
1153 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
1154 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
1155 "Wrong arguments for variable blend intrinsic");
1156 SmallVector<Constant *, 32> Selectors;
1157 for (unsigned I = 0; I < Size; ++I) {
1158 // The intrinsics only read the top bit
1161 Selector = C->getElementAsInteger(I);
1163 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
1164 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
1166 auto NewSelector = ConstantVector::get(Selectors);
1167 return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
1172 case Intrinsic::x86_avx_vpermilvar_ps:
1173 case Intrinsic::x86_avx_vpermilvar_ps_256:
1174 case Intrinsic::x86_avx_vpermilvar_pd:
1175 case Intrinsic::x86_avx_vpermilvar_pd_256: {
1176 // Convert vpermil* to shufflevector if the mask is constant.
1177 Value *V = II->getArgOperand(1);
1178 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1179 assert(Size == 8 || Size == 4 || Size == 2);
1180 uint32_t Indexes[8];
1181 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1182 // The intrinsics only read one or two bits, clear the rest.
1183 for (unsigned I = 0; I < Size; ++I) {
1184 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1185 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1186 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1190 } else if (isa<ConstantAggregateZero>(V)) {
1191 for (unsigned I = 0; I < Size; ++I)
1196 // The _256 variants are a bit trickier since the mask bits always index
1197 // into the corresponding 128 half. In order to convert to a generic
1198 // shuffle, we have to make that explicit.
1199 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1200 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1201 for (unsigned I = Size / 2; I < Size; ++I)
1202 Indexes[I] += Size / 2;
1205 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1206 auto V1 = II->getArgOperand(0);
1207 auto V2 = UndefValue::get(V1->getType());
1208 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1209 return ReplaceInstUsesWith(CI, Shuffle);
1212 case Intrinsic::x86_avx_vperm2f128_pd_256:
1213 case Intrinsic::x86_avx_vperm2f128_ps_256:
1214 case Intrinsic::x86_avx_vperm2f128_si_256:
1215 case Intrinsic::x86_avx2_vperm2i128:
1216 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1217 return ReplaceInstUsesWith(*II, V);
1220 case Intrinsic::ppc_altivec_vperm:
1221 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1222 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1223 // a vectorshuffle for little endian, we must undo the transformation
1224 // performed on vec_perm in altivec.h. That is, we must complement
1225 // the permutation mask with respect to 31 and reverse the order of
1227 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1228 assert(Mask->getType()->getVectorNumElements() == 16 &&
1229 "Bad type for intrinsic!");
1231 // Check that all of the elements are integer constants or undefs.
1232 bool AllEltsOk = true;
1233 for (unsigned i = 0; i != 16; ++i) {
1234 Constant *Elt = Mask->getAggregateElement(i);
1235 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1242 // Cast the input vectors to byte vectors.
1243 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1245 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1247 Value *Result = UndefValue::get(Op0->getType());
1249 // Only extract each element once.
1250 Value *ExtractedElts[32];
1251 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1253 for (unsigned i = 0; i != 16; ++i) {
1254 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1257 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1258 Idx &= 31; // Match the hardware behavior.
1259 if (DL.isLittleEndian())
1262 if (!ExtractedElts[Idx]) {
1263 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1264 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1265 ExtractedElts[Idx] =
1266 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1267 Builder->getInt32(Idx&15));
1270 // Insert this value into the result vector.
1271 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1272 Builder->getInt32(i));
1274 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1279 case Intrinsic::arm_neon_vld1:
1280 case Intrinsic::arm_neon_vld2:
1281 case Intrinsic::arm_neon_vld3:
1282 case Intrinsic::arm_neon_vld4:
1283 case Intrinsic::arm_neon_vld2lane:
1284 case Intrinsic::arm_neon_vld3lane:
1285 case Intrinsic::arm_neon_vld4lane:
1286 case Intrinsic::arm_neon_vst1:
1287 case Intrinsic::arm_neon_vst2:
1288 case Intrinsic::arm_neon_vst3:
1289 case Intrinsic::arm_neon_vst4:
1290 case Intrinsic::arm_neon_vst2lane:
1291 case Intrinsic::arm_neon_vst3lane:
1292 case Intrinsic::arm_neon_vst4lane: {
1293 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1294 unsigned AlignArg = II->getNumArgOperands() - 1;
1295 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1296 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1297 II->setArgOperand(AlignArg,
1298 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1305 case Intrinsic::arm_neon_vmulls:
1306 case Intrinsic::arm_neon_vmullu:
1307 case Intrinsic::aarch64_neon_smull:
1308 case Intrinsic::aarch64_neon_umull: {
1309 Value *Arg0 = II->getArgOperand(0);
1310 Value *Arg1 = II->getArgOperand(1);
1312 // Handle mul by zero first:
1313 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1314 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1317 // Check for constant LHS & RHS - in this case we just simplify.
1318 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1319 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1320 VectorType *NewVT = cast<VectorType>(II->getType());
1321 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1322 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1323 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1324 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1326 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1329 // Couldn't simplify - canonicalize constant to the RHS.
1330 std::swap(Arg0, Arg1);
1333 // Handle mul by one:
1334 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1335 if (ConstantInt *Splat =
1336 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1338 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1339 /*isSigned=*/!Zext);
1344 case Intrinsic::AMDGPU_rcp: {
1345 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1346 const APFloat &ArgVal = C->getValueAPF();
1347 APFloat Val(ArgVal.getSemantics(), 1.0);
1348 APFloat::opStatus Status = Val.divide(ArgVal,
1349 APFloat::rmNearestTiesToEven);
1350 // Only do this if it was exact and therefore not dependent on the
1352 if (Status == APFloat::opOK)
1353 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1358 case Intrinsic::stackrestore: {
1359 // If the save is right next to the restore, remove the restore. This can
1360 // happen when variable allocas are DCE'd.
1361 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1362 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1363 BasicBlock::iterator BI = SS;
1365 return EraseInstFromFunction(CI);
1369 // Scan down this block to see if there is another stack restore in the
1370 // same block without an intervening call/alloca.
1371 BasicBlock::iterator BI = II;
1372 TerminatorInst *TI = II->getParent()->getTerminator();
1373 bool CannotRemove = false;
1374 for (++BI; &*BI != TI; ++BI) {
1375 if (isa<AllocaInst>(BI)) {
1376 CannotRemove = true;
1379 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1380 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1381 // If there is a stackrestore below this one, remove this one.
1382 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1383 return EraseInstFromFunction(CI);
1384 // Otherwise, ignore the intrinsic.
1386 // If we found a non-intrinsic call, we can't remove the stack
1388 CannotRemove = true;
1394 // If the stack restore is in a return, resume, or unwind block and if there
1395 // are no allocas or calls between the restore and the return, nuke the
1397 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1398 return EraseInstFromFunction(CI);
1401 case Intrinsic::assume: {
1402 // Canonicalize assume(a && b) -> assume(a); assume(b);
1403 // Note: New assumption intrinsics created here are registered by
1404 // the InstCombineIRInserter object.
1405 Value *IIOperand = II->getArgOperand(0), *A, *B,
1406 *AssumeIntrinsic = II->getCalledValue();
1407 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1408 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1409 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1410 return EraseInstFromFunction(*II);
1412 // assume(!(a || b)) -> assume(!a); assume(!b);
1413 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1414 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1416 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1418 return EraseInstFromFunction(*II);
1421 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1422 // (if assume is valid at the load)
1423 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1424 Value *LHS = ICmp->getOperand(0);
1425 Value *RHS = ICmp->getOperand(1);
1426 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1427 isa<LoadInst>(LHS) &&
1428 isa<Constant>(RHS) &&
1429 RHS->getType()->isPointerTy() &&
1430 cast<Constant>(RHS)->isNullValue()) {
1431 LoadInst* LI = cast<LoadInst>(LHS);
1432 if (isValidAssumeForContext(II, LI, DT)) {
1433 MDNode *MD = MDNode::get(II->getContext(), None);
1434 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1435 return EraseInstFromFunction(*II);
1438 // TODO: apply nonnull return attributes to calls and invokes
1439 // TODO: apply range metadata for range check patterns?
1441 // If there is a dominating assume with the same condition as this one,
1442 // then this one is redundant, and should be removed.
1443 APInt KnownZero(1, 0), KnownOne(1, 0);
1444 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1445 if (KnownOne.isAllOnesValue())
1446 return EraseInstFromFunction(*II);
1450 case Intrinsic::experimental_gc_relocate: {
1451 // Translate facts known about a pointer before relocating into
1452 // facts about the relocate value, while being careful to
1453 // preserve relocation semantics.
1454 GCRelocateOperands Operands(II);
1455 Value *DerivedPtr = Operands.getDerivedPtr();
1456 auto *GCRelocateType = cast<PointerType>(II->getType());
1458 // Remove the relocation if unused, note that this check is required
1459 // to prevent the cases below from looping forever.
1460 if (II->use_empty())
1461 return EraseInstFromFunction(*II);
1463 // Undef is undef, even after relocation.
1464 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1465 // most practical collectors, but there was discussion in the review thread
1466 // about whether it was legal for all possible collectors.
1467 if (isa<UndefValue>(DerivedPtr)) {
1468 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1469 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1472 // The relocation of null will be null for most any collector.
1473 // TODO: provide a hook for this in GCStrategy. There might be some weird
1474 // collector this property does not hold for.
1475 if (isa<ConstantPointerNull>(DerivedPtr)) {
1476 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1477 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1480 // isKnownNonNull -> nonnull attribute
1481 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI))
1482 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1484 // isDereferenceablePointer -> deref attribute
1485 if (isDereferenceablePointer(DerivedPtr, DL)) {
1486 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1487 uint64_t Bytes = A->getDereferenceableBytes();
1488 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1492 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1493 // Canonicalize on the type from the uses to the defs
1495 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1499 return visitCallSite(II);
1502 // InvokeInst simplification
1504 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1505 return visitCallSite(&II);
1508 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1509 /// passed through the varargs area, we can eliminate the use of the cast.
1510 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1511 const DataLayout &DL,
1512 const CastInst *const CI,
1514 if (!CI->isLosslessCast())
1517 // If this is a GC intrinsic, avoid munging types. We need types for
1518 // statepoint reconstruction in SelectionDAG.
1519 // TODO: This is probably something which should be expanded to all
1520 // intrinsics since the entire point of intrinsics is that
1521 // they are understandable by the optimizer.
1522 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1525 // The size of ByVal or InAlloca arguments is derived from the type, so we
1526 // can't change to a type with a different size. If the size were
1527 // passed explicitly we could avoid this check.
1528 if (!CS.isByValOrInAllocaArgument(ix))
1532 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1533 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1534 if (!SrcTy->isSized() || !DstTy->isSized())
1536 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1541 // Try to fold some different type of calls here.
1542 // Currently we're only working with the checking functions, memcpy_chk,
1543 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1544 // strcat_chk and strncat_chk.
1545 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1546 if (!CI->getCalledFunction()) return nullptr;
1548 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1549 ReplaceInstUsesWith(*From, With);
1551 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1552 if (Value *With = Simplifier.optimizeCall(CI)) {
1554 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1560 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1561 // Strip off at most one level of pointer casts, looking for an alloca. This
1562 // is good enough in practice and simpler than handling any number of casts.
1563 Value *Underlying = TrampMem->stripPointerCasts();
1564 if (Underlying != TrampMem &&
1565 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1567 if (!isa<AllocaInst>(Underlying))
1570 IntrinsicInst *InitTrampoline = nullptr;
1571 for (User *U : TrampMem->users()) {
1572 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1575 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1577 // More than one init_trampoline writes to this value. Give up.
1579 InitTrampoline = II;
1582 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1583 // Allow any number of calls to adjust.trampoline.
1588 // No call to init.trampoline found.
1589 if (!InitTrampoline)
1592 // Check that the alloca is being used in the expected way.
1593 if (InitTrampoline->getOperand(0) != TrampMem)
1596 return InitTrampoline;
1599 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1601 // Visit all the previous instructions in the basic block, and try to find a
1602 // init.trampoline which has a direct path to the adjust.trampoline.
1603 for (BasicBlock::iterator I = AdjustTramp,
1604 E = AdjustTramp->getParent()->begin(); I != E; ) {
1605 Instruction *Inst = --I;
1606 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1607 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1608 II->getOperand(0) == TrampMem)
1610 if (Inst->mayWriteToMemory())
1616 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1617 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1618 // to a direct call to a function. Otherwise return NULL.
1620 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1621 Callee = Callee->stripPointerCasts();
1622 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1624 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1627 Value *TrampMem = AdjustTramp->getOperand(0);
1629 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1631 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1636 // visitCallSite - Improvements for call and invoke instructions.
1638 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1640 if (isAllocLikeFn(CS.getInstruction(), TLI))
1641 return visitAllocSite(*CS.getInstruction());
1643 bool Changed = false;
1645 // Mark any parameters that are known to be non-null with the nonnull
1646 // attribute. This is helpful for inlining calls to functions with null
1647 // checks on their arguments.
1649 for (Value *V : CS.args()) {
1650 if (V->getType()->isPointerTy() && !CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1651 isKnownNonNullAt(V, CS.getInstruction(), DT, TLI)) {
1652 AttributeSet AS = CS.getAttributes();
1653 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1654 Attribute::NonNull);
1655 CS.setAttributes(AS);
1660 assert(ArgNo == CS.arg_size() && "sanity check");
1662 // If the callee is a pointer to a function, attempt to move any casts to the
1663 // arguments of the call/invoke.
1664 Value *Callee = CS.getCalledValue();
1665 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1668 if (Function *CalleeF = dyn_cast<Function>(Callee))
1669 // If the call and callee calling conventions don't match, this call must
1670 // be unreachable, as the call is undefined.
1671 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1672 // Only do this for calls to a function with a body. A prototype may
1673 // not actually end up matching the implementation's calling conv for a
1674 // variety of reasons (e.g. it may be written in assembly).
1675 !CalleeF->isDeclaration()) {
1676 Instruction *OldCall = CS.getInstruction();
1677 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1678 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1680 // If OldCall does not return void then replaceAllUsesWith undef.
1681 // This allows ValueHandlers and custom metadata to adjust itself.
1682 if (!OldCall->getType()->isVoidTy())
1683 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1684 if (isa<CallInst>(OldCall))
1685 return EraseInstFromFunction(*OldCall);
1687 // We cannot remove an invoke, because it would change the CFG, just
1688 // change the callee to a null pointer.
1689 cast<InvokeInst>(OldCall)->setCalledFunction(
1690 Constant::getNullValue(CalleeF->getType()));
1694 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1695 // If CS does not return void then replaceAllUsesWith undef.
1696 // This allows ValueHandlers and custom metadata to adjust itself.
1697 if (!CS.getInstruction()->getType()->isVoidTy())
1698 ReplaceInstUsesWith(*CS.getInstruction(),
1699 UndefValue::get(CS.getInstruction()->getType()));
1701 if (isa<InvokeInst>(CS.getInstruction())) {
1702 // Can't remove an invoke because we cannot change the CFG.
1706 // This instruction is not reachable, just remove it. We insert a store to
1707 // undef so that we know that this code is not reachable, despite the fact
1708 // that we can't modify the CFG here.
1709 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1710 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1711 CS.getInstruction());
1713 return EraseInstFromFunction(*CS.getInstruction());
1716 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1717 return transformCallThroughTrampoline(CS, II);
1719 PointerType *PTy = cast<PointerType>(Callee->getType());
1720 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1721 if (FTy->isVarArg()) {
1722 int ix = FTy->getNumParams();
1723 // See if we can optimize any arguments passed through the varargs area of
1725 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1726 E = CS.arg_end(); I != E; ++I, ++ix) {
1727 CastInst *CI = dyn_cast<CastInst>(*I);
1728 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1729 *I = CI->getOperand(0);
1735 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1736 // Inline asm calls cannot throw - mark them 'nounwind'.
1737 CS.setDoesNotThrow();
1741 // Try to optimize the call if possible, we require DataLayout for most of
1742 // this. None of these calls are seen as possibly dead so go ahead and
1743 // delete the instruction now.
1744 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1745 Instruction *I = tryOptimizeCall(CI);
1746 // If we changed something return the result, etc. Otherwise let
1747 // the fallthrough check.
1748 if (I) return EraseInstFromFunction(*I);
1751 return Changed ? CS.getInstruction() : nullptr;
1754 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1755 // attempt to move the cast to the arguments of the call/invoke.
1757 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1759 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1762 // The prototype of thunks are a lie, don't try to directly call such
1764 if (Callee->hasFnAttribute("thunk"))
1766 Instruction *Caller = CS.getInstruction();
1767 const AttributeSet &CallerPAL = CS.getAttributes();
1769 // Okay, this is a cast from a function to a different type. Unless doing so
1770 // would cause a type conversion of one of our arguments, change this call to
1771 // be a direct call with arguments casted to the appropriate types.
1773 FunctionType *FT = Callee->getFunctionType();
1774 Type *OldRetTy = Caller->getType();
1775 Type *NewRetTy = FT->getReturnType();
1777 // Check to see if we are changing the return type...
1778 if (OldRetTy != NewRetTy) {
1780 if (NewRetTy->isStructTy())
1781 return false; // TODO: Handle multiple return values.
1783 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1784 if (Callee->isDeclaration())
1785 return false; // Cannot transform this return value.
1787 if (!Caller->use_empty() &&
1788 // void -> non-void is handled specially
1789 !NewRetTy->isVoidTy())
1790 return false; // Cannot transform this return value.
1793 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1794 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1795 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1796 return false; // Attribute not compatible with transformed value.
1799 // If the callsite is an invoke instruction, and the return value is used by
1800 // a PHI node in a successor, we cannot change the return type of the call
1801 // because there is no place to put the cast instruction (without breaking
1802 // the critical edge). Bail out in this case.
1803 if (!Caller->use_empty())
1804 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1805 for (User *U : II->users())
1806 if (PHINode *PN = dyn_cast<PHINode>(U))
1807 if (PN->getParent() == II->getNormalDest() ||
1808 PN->getParent() == II->getUnwindDest())
1812 unsigned NumActualArgs = CS.arg_size();
1813 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1815 // Prevent us turning:
1816 // declare void @takes_i32_inalloca(i32* inalloca)
1817 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1820 // call void @takes_i32_inalloca(i32* null)
1822 // Similarly, avoid folding away bitcasts of byval calls.
1823 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1824 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1827 CallSite::arg_iterator AI = CS.arg_begin();
1828 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1829 Type *ParamTy = FT->getParamType(i);
1830 Type *ActTy = (*AI)->getType();
1832 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1833 return false; // Cannot transform this parameter value.
1835 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1836 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1837 return false; // Attribute not compatible with transformed value.
1839 if (CS.isInAllocaArgument(i))
1840 return false; // Cannot transform to and from inalloca.
1842 // If the parameter is passed as a byval argument, then we have to have a
1843 // sized type and the sized type has to have the same size as the old type.
1844 if (ParamTy != ActTy &&
1845 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1846 Attribute::ByVal)) {
1847 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1848 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1851 Type *CurElTy = ActTy->getPointerElementType();
1852 if (DL.getTypeAllocSize(CurElTy) !=
1853 DL.getTypeAllocSize(ParamPTy->getElementType()))
1858 if (Callee->isDeclaration()) {
1859 // Do not delete arguments unless we have a function body.
1860 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1863 // If the callee is just a declaration, don't change the varargsness of the
1864 // call. We don't want to introduce a varargs call where one doesn't
1866 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1867 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1870 // If both the callee and the cast type are varargs, we still have to make
1871 // sure the number of fixed parameters are the same or we have the same
1872 // ABI issues as if we introduce a varargs call.
1873 if (FT->isVarArg() &&
1874 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1875 FT->getNumParams() !=
1876 cast<FunctionType>(APTy->getElementType())->getNumParams())
1880 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1881 !CallerPAL.isEmpty())
1882 // In this case we have more arguments than the new function type, but we
1883 // won't be dropping them. Check that these extra arguments have attributes
1884 // that are compatible with being a vararg call argument.
1885 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1886 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1887 if (Index <= FT->getNumParams())
1890 // Check if it has an attribute that's incompatible with varargs.
1891 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1892 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1897 // Okay, we decided that this is a safe thing to do: go ahead and start
1898 // inserting cast instructions as necessary.
1899 std::vector<Value*> Args;
1900 Args.reserve(NumActualArgs);
1901 SmallVector<AttributeSet, 8> attrVec;
1902 attrVec.reserve(NumCommonArgs);
1904 // Get any return attributes.
1905 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1907 // If the return value is not being used, the type may not be compatible
1908 // with the existing attributes. Wipe out any problematic attributes.
1909 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1911 // Add the new return attributes.
1912 if (RAttrs.hasAttributes())
1913 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1914 AttributeSet::ReturnIndex, RAttrs));
1916 AI = CS.arg_begin();
1917 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1918 Type *ParamTy = FT->getParamType(i);
1920 if ((*AI)->getType() == ParamTy) {
1921 Args.push_back(*AI);
1923 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1926 // Add any parameter attributes.
1927 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1928 if (PAttrs.hasAttributes())
1929 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1933 // If the function takes more arguments than the call was taking, add them
1935 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1936 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1938 // If we are removing arguments to the function, emit an obnoxious warning.
1939 if (FT->getNumParams() < NumActualArgs) {
1940 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1941 if (FT->isVarArg()) {
1942 // Add all of the arguments in their promoted form to the arg list.
1943 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1944 Type *PTy = getPromotedType((*AI)->getType());
1945 if (PTy != (*AI)->getType()) {
1946 // Must promote to pass through va_arg area!
1947 Instruction::CastOps opcode =
1948 CastInst::getCastOpcode(*AI, false, PTy, false);
1949 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1951 Args.push_back(*AI);
1954 // Add any parameter attributes.
1955 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1956 if (PAttrs.hasAttributes())
1957 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1963 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1964 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1965 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1967 if (NewRetTy->isVoidTy())
1968 Caller->setName(""); // Void type should not have a name.
1970 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1974 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1975 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1976 II->getUnwindDest(), Args);
1978 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1979 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1981 CallInst *CI = cast<CallInst>(Caller);
1982 NC = Builder->CreateCall(Callee, Args);
1984 if (CI->isTailCall())
1985 cast<CallInst>(NC)->setTailCall();
1986 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1987 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1990 // Insert a cast of the return type as necessary.
1992 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1993 if (!NV->getType()->isVoidTy()) {
1994 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1995 NC->setDebugLoc(Caller->getDebugLoc());
1997 // If this is an invoke instruction, we should insert it after the first
1998 // non-phi, instruction in the normal successor block.
1999 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2000 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
2001 InsertNewInstBefore(NC, *I);
2003 // Otherwise, it's a call, just insert cast right after the call.
2004 InsertNewInstBefore(NC, *Caller);
2006 Worklist.AddUsersToWorkList(*Caller);
2008 NV = UndefValue::get(Caller->getType());
2012 if (!Caller->use_empty())
2013 ReplaceInstUsesWith(*Caller, NV);
2014 else if (Caller->hasValueHandle()) {
2015 if (OldRetTy == NV->getType())
2016 ValueHandleBase::ValueIsRAUWd(Caller, NV);
2018 // We cannot call ValueIsRAUWd with a different type, and the
2019 // actual tracked value will disappear.
2020 ValueHandleBase::ValueIsDeleted(Caller);
2023 EraseInstFromFunction(*Caller);
2027 // transformCallThroughTrampoline - Turn a call to a function created by
2028 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
2029 // underlying function.
2032 InstCombiner::transformCallThroughTrampoline(CallSite CS,
2033 IntrinsicInst *Tramp) {
2034 Value *Callee = CS.getCalledValue();
2035 PointerType *PTy = cast<PointerType>(Callee->getType());
2036 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
2037 const AttributeSet &Attrs = CS.getAttributes();
2039 // If the call already has the 'nest' attribute somewhere then give up -
2040 // otherwise 'nest' would occur twice after splicing in the chain.
2041 if (Attrs.hasAttrSomewhere(Attribute::Nest))
2045 "transformCallThroughTrampoline called with incorrect CallSite.");
2047 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
2048 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
2049 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
2051 const AttributeSet &NestAttrs = NestF->getAttributes();
2052 if (!NestAttrs.isEmpty()) {
2053 unsigned NestIdx = 1;
2054 Type *NestTy = nullptr;
2055 AttributeSet NestAttr;
2057 // Look for a parameter marked with the 'nest' attribute.
2058 for (FunctionType::param_iterator I = NestFTy->param_begin(),
2059 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
2060 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
2061 // Record the parameter type and any other attributes.
2063 NestAttr = NestAttrs.getParamAttributes(NestIdx);
2068 Instruction *Caller = CS.getInstruction();
2069 std::vector<Value*> NewArgs;
2070 NewArgs.reserve(CS.arg_size() + 1);
2072 SmallVector<AttributeSet, 8> NewAttrs;
2073 NewAttrs.reserve(Attrs.getNumSlots() + 1);
2075 // Insert the nest argument into the call argument list, which may
2076 // mean appending it. Likewise for attributes.
2078 // Add any result attributes.
2079 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
2080 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2081 Attrs.getRetAttributes()));
2085 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
2087 if (Idx == NestIdx) {
2088 // Add the chain argument and attributes.
2089 Value *NestVal = Tramp->getArgOperand(2);
2090 if (NestVal->getType() != NestTy)
2091 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
2092 NewArgs.push_back(NestVal);
2093 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2100 // Add the original argument and attributes.
2101 NewArgs.push_back(*I);
2102 AttributeSet Attr = Attrs.getParamAttributes(Idx);
2103 if (Attr.hasAttributes(Idx)) {
2104 AttrBuilder B(Attr, Idx);
2105 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2106 Idx + (Idx >= NestIdx), B));
2113 // Add any function attributes.
2114 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
2115 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
2116 Attrs.getFnAttributes()));
2118 // The trampoline may have been bitcast to a bogus type (FTy).
2119 // Handle this by synthesizing a new function type, equal to FTy
2120 // with the chain parameter inserted.
2122 std::vector<Type*> NewTypes;
2123 NewTypes.reserve(FTy->getNumParams()+1);
2125 // Insert the chain's type into the list of parameter types, which may
2126 // mean appending it.
2129 FunctionType::param_iterator I = FTy->param_begin(),
2130 E = FTy->param_end();
2134 // Add the chain's type.
2135 NewTypes.push_back(NestTy);
2140 // Add the original type.
2141 NewTypes.push_back(*I);
2147 // Replace the trampoline call with a direct call. Let the generic
2148 // code sort out any function type mismatches.
2149 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
2151 Constant *NewCallee =
2152 NestF->getType() == PointerType::getUnqual(NewFTy) ?
2153 NestF : ConstantExpr::getBitCast(NestF,
2154 PointerType::getUnqual(NewFTy));
2155 const AttributeSet &NewPAL =
2156 AttributeSet::get(FTy->getContext(), NewAttrs);
2158 Instruction *NewCaller;
2159 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2160 NewCaller = InvokeInst::Create(NewCallee,
2161 II->getNormalDest(), II->getUnwindDest(),
2163 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
2164 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
2166 NewCaller = CallInst::Create(NewCallee, NewArgs);
2167 if (cast<CallInst>(Caller)->isTailCall())
2168 cast<CallInst>(NewCaller)->setTailCall();
2169 cast<CallInst>(NewCaller)->
2170 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
2171 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
2178 // Replace the trampoline call with a direct call. Since there is no 'nest'
2179 // parameter, there is no need to adjust the argument list. Let the generic
2180 // code sort out any function type mismatches.
2181 Constant *NewCallee =
2182 NestF->getType() == PTy ? NestF :
2183 ConstantExpr::getBitCast(NestF, PTy);
2184 CS.setCalledFunction(NewCallee);
2185 return CS.getInstruction();