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 switch (II->getIntrinsicID()) {
532 case Intrinsic::objectsize: {
534 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
535 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
538 case Intrinsic::bswap: {
539 Value *IIOperand = II->getArgOperand(0);
542 // bswap(bswap(x)) -> x
543 if (match(IIOperand, m_BSwap(m_Value(X))))
544 return ReplaceInstUsesWith(CI, X);
546 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
547 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
548 unsigned C = X->getType()->getPrimitiveSizeInBits() -
549 IIOperand->getType()->getPrimitiveSizeInBits();
550 Value *CV = ConstantInt::get(X->getType(), C);
551 Value *V = Builder->CreateLShr(X, CV);
552 return new TruncInst(V, IIOperand->getType());
557 case Intrinsic::powi:
558 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
561 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
564 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
565 // powi(x, -1) -> 1/x
566 if (Power->isAllOnesValue())
567 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
568 II->getArgOperand(0));
571 case Intrinsic::cttz: {
572 // If all bits below the first known one are known zero,
573 // this value is constant.
574 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
575 // FIXME: Try to simplify vectors of integers.
577 uint32_t BitWidth = IT->getBitWidth();
578 APInt KnownZero(BitWidth, 0);
579 APInt KnownOne(BitWidth, 0);
580 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
581 unsigned TrailingZeros = KnownOne.countTrailingZeros();
582 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
583 if ((Mask & KnownZero) == Mask)
584 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
585 APInt(BitWidth, TrailingZeros)));
589 case Intrinsic::ctlz: {
590 // If all bits above the first known one are known zero,
591 // this value is constant.
592 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
593 // FIXME: Try to simplify vectors of integers.
595 uint32_t BitWidth = IT->getBitWidth();
596 APInt KnownZero(BitWidth, 0);
597 APInt KnownOne(BitWidth, 0);
598 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
599 unsigned LeadingZeros = KnownOne.countLeadingZeros();
600 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
601 if ((Mask & KnownZero) == Mask)
602 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
603 APInt(BitWidth, LeadingZeros)));
608 case Intrinsic::uadd_with_overflow:
609 case Intrinsic::sadd_with_overflow:
610 case Intrinsic::umul_with_overflow:
611 case Intrinsic::smul_with_overflow:
612 if (isa<Constant>(II->getArgOperand(0)) &&
613 !isa<Constant>(II->getArgOperand(1))) {
614 // Canonicalize constants into the RHS.
615 Value *LHS = II->getArgOperand(0);
616 II->setArgOperand(0, II->getArgOperand(1));
617 II->setArgOperand(1, LHS);
622 case Intrinsic::usub_with_overflow:
623 case Intrinsic::ssub_with_overflow: {
624 OverflowCheckFlavor OCF =
625 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
626 assert(OCF != OCF_INVALID && "unexpected!");
628 Value *OperationResult = nullptr;
629 Constant *OverflowResult = nullptr;
630 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
631 *II, OperationResult, OverflowResult))
632 return CreateOverflowTuple(II, OperationResult, OverflowResult);
637 case Intrinsic::minnum:
638 case Intrinsic::maxnum: {
639 Value *Arg0 = II->getArgOperand(0);
640 Value *Arg1 = II->getArgOperand(1);
644 return ReplaceInstUsesWith(CI, Arg0);
646 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
647 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
649 // Canonicalize constants into the RHS.
651 II->setArgOperand(0, Arg1);
652 II->setArgOperand(1, Arg0);
657 if (C1 && C1->isNaN())
658 return ReplaceInstUsesWith(CI, Arg0);
660 // This is the value because if undef were NaN, we would return the other
661 // value and cannot return a NaN unless both operands are.
663 // fmin(undef, x) -> x
664 if (isa<UndefValue>(Arg0))
665 return ReplaceInstUsesWith(CI, Arg1);
667 // fmin(x, undef) -> x
668 if (isa<UndefValue>(Arg1))
669 return ReplaceInstUsesWith(CI, Arg0);
673 if (II->getIntrinsicID() == Intrinsic::minnum) {
674 // fmin(x, fmin(x, y)) -> fmin(x, y)
675 // fmin(y, fmin(x, y)) -> fmin(x, y)
676 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
677 if (Arg0 == X || Arg0 == Y)
678 return ReplaceInstUsesWith(CI, Arg1);
681 // fmin(fmin(x, y), x) -> fmin(x, y)
682 // fmin(fmin(x, y), y) -> fmin(x, y)
683 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
684 if (Arg1 == X || Arg1 == Y)
685 return ReplaceInstUsesWith(CI, Arg0);
688 // TODO: fmin(nnan x, inf) -> x
689 // TODO: fmin(nnan ninf x, flt_max) -> x
690 if (C1 && C1->isInfinity()) {
691 // fmin(x, -inf) -> -inf
692 if (C1->isNegative())
693 return ReplaceInstUsesWith(CI, Arg1);
696 assert(II->getIntrinsicID() == Intrinsic::maxnum);
697 // fmax(x, fmax(x, y)) -> fmax(x, y)
698 // fmax(y, fmax(x, y)) -> fmax(x, y)
699 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
700 if (Arg0 == X || Arg0 == Y)
701 return ReplaceInstUsesWith(CI, Arg1);
704 // fmax(fmax(x, y), x) -> fmax(x, y)
705 // fmax(fmax(x, y), y) -> fmax(x, y)
706 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
707 if (Arg1 == X || Arg1 == Y)
708 return ReplaceInstUsesWith(CI, Arg0);
711 // TODO: fmax(nnan x, -inf) -> x
712 // TODO: fmax(nnan ninf x, -flt_max) -> x
713 if (C1 && C1->isInfinity()) {
714 // fmax(x, inf) -> inf
715 if (!C1->isNegative())
716 return ReplaceInstUsesWith(CI, Arg1);
721 case Intrinsic::ppc_altivec_lvx:
722 case Intrinsic::ppc_altivec_lvxl:
723 // Turn PPC lvx -> load if the pointer is known aligned.
724 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
726 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
727 PointerType::getUnqual(II->getType()));
728 return new LoadInst(Ptr);
731 case Intrinsic::ppc_vsx_lxvw4x:
732 case Intrinsic::ppc_vsx_lxvd2x: {
733 // Turn PPC VSX loads into normal loads.
734 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
735 PointerType::getUnqual(II->getType()));
736 return new LoadInst(Ptr, Twine(""), false, 1);
738 case Intrinsic::ppc_altivec_stvx:
739 case Intrinsic::ppc_altivec_stvxl:
740 // Turn stvx -> store if the pointer is known aligned.
741 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
744 PointerType::getUnqual(II->getArgOperand(0)->getType());
745 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
746 return new StoreInst(II->getArgOperand(0), Ptr);
749 case Intrinsic::ppc_vsx_stxvw4x:
750 case Intrinsic::ppc_vsx_stxvd2x: {
751 // Turn PPC VSX stores into normal stores.
752 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
753 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
754 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
756 case Intrinsic::ppc_qpx_qvlfs:
757 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
758 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
760 Type *VTy = VectorType::get(Builder->getFloatTy(),
761 II->getType()->getVectorNumElements());
762 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
763 PointerType::getUnqual(VTy));
764 Value *Load = Builder->CreateLoad(Ptr);
765 return new FPExtInst(Load, II->getType());
768 case Intrinsic::ppc_qpx_qvlfd:
769 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
770 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
772 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
773 PointerType::getUnqual(II->getType()));
774 return new LoadInst(Ptr);
777 case Intrinsic::ppc_qpx_qvstfs:
778 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
779 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
781 Type *VTy = VectorType::get(Builder->getFloatTy(),
782 II->getArgOperand(0)->getType()->getVectorNumElements());
783 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
784 Type *OpPtrTy = PointerType::getUnqual(VTy);
785 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
786 return new StoreInst(TOp, Ptr);
789 case Intrinsic::ppc_qpx_qvstfd:
790 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
791 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
794 PointerType::getUnqual(II->getArgOperand(0)->getType());
795 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
796 return new StoreInst(II->getArgOperand(0), Ptr);
800 case Intrinsic::x86_sse_storeu_ps:
801 case Intrinsic::x86_sse2_storeu_pd:
802 case Intrinsic::x86_sse2_storeu_dq:
803 // Turn X86 storeu -> store if the pointer is known aligned.
804 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
807 PointerType::getUnqual(II->getArgOperand(1)->getType());
808 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
809 return new StoreInst(II->getArgOperand(1), Ptr);
813 case Intrinsic::x86_vcvtph2ps_128:
814 case Intrinsic::x86_vcvtph2ps_256: {
815 auto Arg = II->getArgOperand(0);
816 auto ArgType = cast<VectorType>(Arg->getType());
817 auto RetType = cast<VectorType>(II->getType());
818 unsigned ArgWidth = ArgType->getNumElements();
819 unsigned RetWidth = RetType->getNumElements();
820 assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
821 assert(ArgType->isIntOrIntVectorTy() &&
822 ArgType->getScalarSizeInBits() == 16 &&
823 "CVTPH2PS input type should be 16-bit integer vector");
824 assert(RetType->getScalarType()->isFloatTy() &&
825 "CVTPH2PS output type should be 32-bit float vector");
827 // Constant folding: Convert to generic half to single conversion.
828 if (isa<ConstantAggregateZero>(Arg))
829 return ReplaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
831 if (isa<ConstantDataVector>(Arg)) {
832 auto VectorHalfAsShorts = Arg;
833 if (RetWidth < ArgWidth) {
834 SmallVector<int, 8> SubVecMask;
835 for (unsigned i = 0; i != RetWidth; ++i)
836 SubVecMask.push_back((int)i);
837 VectorHalfAsShorts = Builder->CreateShuffleVector(
838 Arg, UndefValue::get(ArgType), SubVecMask);
841 auto VectorHalfType =
842 VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
844 Builder->CreateBitCast(VectorHalfAsShorts, VectorHalfType);
845 auto VectorFloats = Builder->CreateFPExt(VectorHalfs, RetType);
846 return ReplaceInstUsesWith(*II, VectorFloats);
849 // We only use the lowest lanes of the argument.
850 APInt DemandedElts = APInt::getLowBitsSet(ArgWidth, RetWidth);
851 APInt UndefElts(ArgWidth, 0);
852 if (Value *V = SimplifyDemandedVectorElts(Arg, DemandedElts, UndefElts)) {
853 II->setArgOperand(0, V);
859 case Intrinsic::x86_sse_cvtss2si:
860 case Intrinsic::x86_sse_cvtss2si64:
861 case Intrinsic::x86_sse_cvttss2si:
862 case Intrinsic::x86_sse_cvttss2si64:
863 case Intrinsic::x86_sse2_cvtsd2si:
864 case Intrinsic::x86_sse2_cvtsd2si64:
865 case Intrinsic::x86_sse2_cvttsd2si:
866 case Intrinsic::x86_sse2_cvttsd2si64: {
867 // These intrinsics only demand the 0th element of their input vectors. If
868 // we can simplify the input based on that, do so now.
870 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
871 APInt DemandedElts(VWidth, 1);
872 APInt UndefElts(VWidth, 0);
873 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
874 DemandedElts, UndefElts)) {
875 II->setArgOperand(0, V);
881 // Constant fold ashr( <A x Bi>, Ci ).
882 // Constant fold lshr( <A x Bi>, Ci ).
883 // Constant fold shl( <A x Bi>, Ci ).
884 case Intrinsic::x86_sse2_psrai_d:
885 case Intrinsic::x86_sse2_psrai_w:
886 case Intrinsic::x86_avx2_psrai_d:
887 case Intrinsic::x86_avx2_psrai_w:
888 case Intrinsic::x86_sse2_psrli_d:
889 case Intrinsic::x86_sse2_psrli_q:
890 case Intrinsic::x86_sse2_psrli_w:
891 case Intrinsic::x86_avx2_psrli_d:
892 case Intrinsic::x86_avx2_psrli_q:
893 case Intrinsic::x86_avx2_psrli_w:
894 case Intrinsic::x86_sse2_pslli_d:
895 case Intrinsic::x86_sse2_pslli_q:
896 case Intrinsic::x86_sse2_pslli_w:
897 case Intrinsic::x86_avx2_pslli_d:
898 case Intrinsic::x86_avx2_pslli_q:
899 case Intrinsic::x86_avx2_pslli_w:
900 if (Value *V = SimplifyX86immshift(*II, *Builder))
901 return ReplaceInstUsesWith(*II, V);
904 case Intrinsic::x86_sse2_psra_d:
905 case Intrinsic::x86_sse2_psra_w:
906 case Intrinsic::x86_avx2_psra_d:
907 case Intrinsic::x86_avx2_psra_w:
908 case Intrinsic::x86_sse2_psrl_d:
909 case Intrinsic::x86_sse2_psrl_q:
910 case Intrinsic::x86_sse2_psrl_w:
911 case Intrinsic::x86_avx2_psrl_d:
912 case Intrinsic::x86_avx2_psrl_q:
913 case Intrinsic::x86_avx2_psrl_w:
914 case Intrinsic::x86_sse2_psll_d:
915 case Intrinsic::x86_sse2_psll_q:
916 case Intrinsic::x86_sse2_psll_w:
917 case Intrinsic::x86_avx2_psll_d:
918 case Intrinsic::x86_avx2_psll_q:
919 case Intrinsic::x86_avx2_psll_w: {
920 if (Value *V = SimplifyX86immshift(*II, *Builder))
921 return ReplaceInstUsesWith(*II, V);
923 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
924 // operand to compute the shift amount.
925 auto ShiftAmt = II->getArgOperand(1);
926 auto ShiftType = cast<VectorType>(ShiftAmt->getType());
927 assert(ShiftType->getPrimitiveSizeInBits() == 128 &&
928 "Unexpected packed shift size");
929 unsigned VWidth = ShiftType->getNumElements();
931 APInt DemandedElts = APInt::getLowBitsSet(VWidth, VWidth / 2);
932 APInt UndefElts(VWidth, 0);
934 SimplifyDemandedVectorElts(ShiftAmt, DemandedElts, UndefElts)) {
935 II->setArgOperand(1, V);
941 case Intrinsic::x86_sse41_pmovsxbd:
942 case Intrinsic::x86_sse41_pmovsxbq:
943 case Intrinsic::x86_sse41_pmovsxbw:
944 case Intrinsic::x86_sse41_pmovsxdq:
945 case Intrinsic::x86_sse41_pmovsxwd:
946 case Intrinsic::x86_sse41_pmovsxwq:
947 case Intrinsic::x86_avx2_pmovsxbd:
948 case Intrinsic::x86_avx2_pmovsxbq:
949 case Intrinsic::x86_avx2_pmovsxbw:
950 case Intrinsic::x86_avx2_pmovsxdq:
951 case Intrinsic::x86_avx2_pmovsxwd:
952 case Intrinsic::x86_avx2_pmovsxwq:
953 if (Value *V = SimplifyX86extend(*II, *Builder, true))
954 return ReplaceInstUsesWith(*II, V);
957 case Intrinsic::x86_sse41_pmovzxbd:
958 case Intrinsic::x86_sse41_pmovzxbq:
959 case Intrinsic::x86_sse41_pmovzxbw:
960 case Intrinsic::x86_sse41_pmovzxdq:
961 case Intrinsic::x86_sse41_pmovzxwd:
962 case Intrinsic::x86_sse41_pmovzxwq:
963 case Intrinsic::x86_avx2_pmovzxbd:
964 case Intrinsic::x86_avx2_pmovzxbq:
965 case Intrinsic::x86_avx2_pmovzxbw:
966 case Intrinsic::x86_avx2_pmovzxdq:
967 case Intrinsic::x86_avx2_pmovzxwd:
968 case Intrinsic::x86_avx2_pmovzxwq:
969 if (Value *V = SimplifyX86extend(*II, *Builder, false))
970 return ReplaceInstUsesWith(*II, V);
973 case Intrinsic::x86_sse41_insertps:
974 if (Value *V = SimplifyX86insertps(*II, *Builder))
975 return ReplaceInstUsesWith(*II, V);
978 case Intrinsic::x86_sse4a_insertqi: {
979 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
981 // TODO: eventually we should lower this intrinsic to IR
982 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
983 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
984 unsigned Index = CIIndex->getZExtValue();
985 // From AMD documentation: "a value of zero in the field length is
986 // defined as length of 64".
987 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
989 // From AMD documentation: "If the sum of the bit index + length field
990 // is greater than 64, the results are undefined".
991 unsigned End = Index + Length;
993 // Note that both field index and field length are 8-bit quantities.
994 // Since variables 'Index' and 'Length' are unsigned values
995 // obtained from zero-extending field index and field length
996 // respectively, their sum should never wrap around.
998 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
1000 if (Length == 64 && Index == 0) {
1001 Value *Vec = II->getArgOperand(1);
1002 Value *Undef = UndefValue::get(Vec->getType());
1003 const uint32_t Mask[] = { 0, 2 };
1004 return ReplaceInstUsesWith(
1006 Builder->CreateShuffleVector(
1007 Vec, Undef, ConstantDataVector::get(
1008 II->getContext(), makeArrayRef(Mask))));
1009 } else if (auto Source =
1010 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1011 if (Source->hasOneUse() &&
1012 Source->getArgOperand(1) == II->getArgOperand(1)) {
1013 // If the source of the insert has only one use and it's another
1014 // insert (and they're both inserting from the same vector), try to
1015 // bundle both together.
1016 auto CISourceLength =
1017 dyn_cast<ConstantInt>(Source->getArgOperand(2));
1018 auto CISourceIndex =
1019 dyn_cast<ConstantInt>(Source->getArgOperand(3));
1020 if (CISourceIndex && CISourceLength) {
1021 unsigned SourceIndex = CISourceIndex->getZExtValue();
1022 unsigned SourceLength = CISourceLength->getZExtValue();
1023 unsigned SourceEnd = SourceIndex + SourceLength;
1024 unsigned NewIndex, NewLength;
1025 bool ShouldReplace = false;
1026 if (Index <= SourceIndex && SourceIndex <= End) {
1028 NewLength = std::max(End, SourceEnd) - NewIndex;
1029 ShouldReplace = true;
1030 } else if (SourceIndex <= Index && Index <= SourceEnd) {
1031 NewIndex = SourceIndex;
1032 NewLength = std::max(SourceEnd, End) - NewIndex;
1033 ShouldReplace = true;
1036 if (ShouldReplace) {
1037 Constant *ConstantLength = ConstantInt::get(
1038 II->getArgOperand(2)->getType(), NewLength, false);
1039 Constant *ConstantIndex = ConstantInt::get(
1040 II->getArgOperand(3)->getType(), NewIndex, false);
1041 Value *Args[4] = { Source->getArgOperand(0),
1042 II->getArgOperand(1), ConstantLength,
1044 Module *M = CI.getParent()->getParent()->getParent();
1046 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1047 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
1057 case Intrinsic::x86_sse41_pblendvb:
1058 case Intrinsic::x86_sse41_blendvps:
1059 case Intrinsic::x86_sse41_blendvpd:
1060 case Intrinsic::x86_avx_blendv_ps_256:
1061 case Intrinsic::x86_avx_blendv_pd_256:
1062 case Intrinsic::x86_avx2_pblendvb: {
1063 // Convert blendv* to vector selects if the mask is constant.
1064 // This optimization is convoluted because the intrinsic is defined as
1065 // getting a vector of floats or doubles for the ps and pd versions.
1066 // FIXME: That should be changed.
1068 Value *Op0 = II->getArgOperand(0);
1069 Value *Op1 = II->getArgOperand(1);
1070 Value *Mask = II->getArgOperand(2);
1072 // fold (blend A, A, Mask) -> A
1074 return ReplaceInstUsesWith(CI, Op0);
1076 // Zero Mask - select 1st argument.
1077 if (isa<ConstantAggregateZero>(Mask))
1078 return ReplaceInstUsesWith(CI, Op0);
1080 // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
1081 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
1082 auto Tyi1 = Builder->getInt1Ty();
1083 auto SelectorType = cast<VectorType>(Mask->getType());
1084 auto EltTy = SelectorType->getElementType();
1085 unsigned Size = SelectorType->getNumElements();
1089 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
1090 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
1091 "Wrong arguments for variable blend intrinsic");
1092 SmallVector<Constant *, 32> Selectors;
1093 for (unsigned I = 0; I < Size; ++I) {
1094 // The intrinsics only read the top bit
1097 Selector = C->getElementAsInteger(I);
1099 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
1100 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
1102 auto NewSelector = ConstantVector::get(Selectors);
1103 return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
1108 case Intrinsic::x86_avx_vpermilvar_ps:
1109 case Intrinsic::x86_avx_vpermilvar_ps_256:
1110 case Intrinsic::x86_avx_vpermilvar_pd:
1111 case Intrinsic::x86_avx_vpermilvar_pd_256: {
1112 // Convert vpermil* to shufflevector if the mask is constant.
1113 Value *V = II->getArgOperand(1);
1114 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1115 assert(Size == 8 || Size == 4 || Size == 2);
1116 uint32_t Indexes[8];
1117 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1118 // The intrinsics only read one or two bits, clear the rest.
1119 for (unsigned I = 0; I < Size; ++I) {
1120 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1121 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1122 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1126 } else if (isa<ConstantAggregateZero>(V)) {
1127 for (unsigned I = 0; I < Size; ++I)
1132 // The _256 variants are a bit trickier since the mask bits always index
1133 // into the corresponding 128 half. In order to convert to a generic
1134 // shuffle, we have to make that explicit.
1135 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1136 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1137 for (unsigned I = Size / 2; I < Size; ++I)
1138 Indexes[I] += Size / 2;
1141 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1142 auto V1 = II->getArgOperand(0);
1143 auto V2 = UndefValue::get(V1->getType());
1144 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1145 return ReplaceInstUsesWith(CI, Shuffle);
1148 case Intrinsic::x86_avx_vperm2f128_pd_256:
1149 case Intrinsic::x86_avx_vperm2f128_ps_256:
1150 case Intrinsic::x86_avx_vperm2f128_si_256:
1151 case Intrinsic::x86_avx2_vperm2i128:
1152 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1153 return ReplaceInstUsesWith(*II, V);
1156 case Intrinsic::ppc_altivec_vperm:
1157 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1158 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1159 // a vectorshuffle for little endian, we must undo the transformation
1160 // performed on vec_perm in altivec.h. That is, we must complement
1161 // the permutation mask with respect to 31 and reverse the order of
1163 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1164 assert(Mask->getType()->getVectorNumElements() == 16 &&
1165 "Bad type for intrinsic!");
1167 // Check that all of the elements are integer constants or undefs.
1168 bool AllEltsOk = true;
1169 for (unsigned i = 0; i != 16; ++i) {
1170 Constant *Elt = Mask->getAggregateElement(i);
1171 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1178 // Cast the input vectors to byte vectors.
1179 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1181 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1183 Value *Result = UndefValue::get(Op0->getType());
1185 // Only extract each element once.
1186 Value *ExtractedElts[32];
1187 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1189 for (unsigned i = 0; i != 16; ++i) {
1190 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1193 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1194 Idx &= 31; // Match the hardware behavior.
1195 if (DL.isLittleEndian())
1198 if (!ExtractedElts[Idx]) {
1199 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1200 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1201 ExtractedElts[Idx] =
1202 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1203 Builder->getInt32(Idx&15));
1206 // Insert this value into the result vector.
1207 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1208 Builder->getInt32(i));
1210 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1215 case Intrinsic::arm_neon_vld1:
1216 case Intrinsic::arm_neon_vld2:
1217 case Intrinsic::arm_neon_vld3:
1218 case Intrinsic::arm_neon_vld4:
1219 case Intrinsic::arm_neon_vld2lane:
1220 case Intrinsic::arm_neon_vld3lane:
1221 case Intrinsic::arm_neon_vld4lane:
1222 case Intrinsic::arm_neon_vst1:
1223 case Intrinsic::arm_neon_vst2:
1224 case Intrinsic::arm_neon_vst3:
1225 case Intrinsic::arm_neon_vst4:
1226 case Intrinsic::arm_neon_vst2lane:
1227 case Intrinsic::arm_neon_vst3lane:
1228 case Intrinsic::arm_neon_vst4lane: {
1229 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1230 unsigned AlignArg = II->getNumArgOperands() - 1;
1231 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1232 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1233 II->setArgOperand(AlignArg,
1234 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1241 case Intrinsic::arm_neon_vmulls:
1242 case Intrinsic::arm_neon_vmullu:
1243 case Intrinsic::aarch64_neon_smull:
1244 case Intrinsic::aarch64_neon_umull: {
1245 Value *Arg0 = II->getArgOperand(0);
1246 Value *Arg1 = II->getArgOperand(1);
1248 // Handle mul by zero first:
1249 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1250 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1253 // Check for constant LHS & RHS - in this case we just simplify.
1254 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1255 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1256 VectorType *NewVT = cast<VectorType>(II->getType());
1257 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1258 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1259 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1260 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1262 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1265 // Couldn't simplify - canonicalize constant to the RHS.
1266 std::swap(Arg0, Arg1);
1269 // Handle mul by one:
1270 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1271 if (ConstantInt *Splat =
1272 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1274 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1275 /*isSigned=*/!Zext);
1280 case Intrinsic::AMDGPU_rcp: {
1281 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1282 const APFloat &ArgVal = C->getValueAPF();
1283 APFloat Val(ArgVal.getSemantics(), 1.0);
1284 APFloat::opStatus Status = Val.divide(ArgVal,
1285 APFloat::rmNearestTiesToEven);
1286 // Only do this if it was exact and therefore not dependent on the
1288 if (Status == APFloat::opOK)
1289 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1294 case Intrinsic::stackrestore: {
1295 // If the save is right next to the restore, remove the restore. This can
1296 // happen when variable allocas are DCE'd.
1297 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1298 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1299 BasicBlock::iterator BI = SS;
1301 return EraseInstFromFunction(CI);
1305 // Scan down this block to see if there is another stack restore in the
1306 // same block without an intervening call/alloca.
1307 BasicBlock::iterator BI = II;
1308 TerminatorInst *TI = II->getParent()->getTerminator();
1309 bool CannotRemove = false;
1310 for (++BI; &*BI != TI; ++BI) {
1311 if (isa<AllocaInst>(BI)) {
1312 CannotRemove = true;
1315 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1316 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1317 // If there is a stackrestore below this one, remove this one.
1318 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1319 return EraseInstFromFunction(CI);
1320 // Otherwise, ignore the intrinsic.
1322 // If we found a non-intrinsic call, we can't remove the stack
1324 CannotRemove = true;
1330 // If the stack restore is in a return, resume, or unwind block and if there
1331 // are no allocas or calls between the restore and the return, nuke the
1333 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1334 return EraseInstFromFunction(CI);
1337 case Intrinsic::assume: {
1338 // Canonicalize assume(a && b) -> assume(a); assume(b);
1339 // Note: New assumption intrinsics created here are registered by
1340 // the InstCombineIRInserter object.
1341 Value *IIOperand = II->getArgOperand(0), *A, *B,
1342 *AssumeIntrinsic = II->getCalledValue();
1343 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1344 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1345 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1346 return EraseInstFromFunction(*II);
1348 // assume(!(a || b)) -> assume(!a); assume(!b);
1349 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1350 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1352 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1354 return EraseInstFromFunction(*II);
1357 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1358 // (if assume is valid at the load)
1359 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1360 Value *LHS = ICmp->getOperand(0);
1361 Value *RHS = ICmp->getOperand(1);
1362 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1363 isa<LoadInst>(LHS) &&
1364 isa<Constant>(RHS) &&
1365 RHS->getType()->isPointerTy() &&
1366 cast<Constant>(RHS)->isNullValue()) {
1367 LoadInst* LI = cast<LoadInst>(LHS);
1368 if (isValidAssumeForContext(II, LI, DT)) {
1369 MDNode *MD = MDNode::get(II->getContext(), None);
1370 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1371 return EraseInstFromFunction(*II);
1374 // TODO: apply nonnull return attributes to calls and invokes
1375 // TODO: apply range metadata for range check patterns?
1377 // If there is a dominating assume with the same condition as this one,
1378 // then this one is redundant, and should be removed.
1379 APInt KnownZero(1, 0), KnownOne(1, 0);
1380 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1381 if (KnownOne.isAllOnesValue())
1382 return EraseInstFromFunction(*II);
1386 case Intrinsic::experimental_gc_relocate: {
1387 // Translate facts known about a pointer before relocating into
1388 // facts about the relocate value, while being careful to
1389 // preserve relocation semantics.
1390 GCRelocateOperands Operands(II);
1391 Value *DerivedPtr = Operands.getDerivedPtr();
1392 auto *GCRelocateType = cast<PointerType>(II->getType());
1394 // Remove the relocation if unused, note that this check is required
1395 // to prevent the cases below from looping forever.
1396 if (II->use_empty())
1397 return EraseInstFromFunction(*II);
1399 // Undef is undef, even after relocation.
1400 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1401 // most practical collectors, but there was discussion in the review thread
1402 // about whether it was legal for all possible collectors.
1403 if (isa<UndefValue>(DerivedPtr)) {
1404 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1405 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1408 // The relocation of null will be null for most any collector.
1409 // TODO: provide a hook for this in GCStrategy. There might be some weird
1410 // collector this property does not hold for.
1411 if (isa<ConstantPointerNull>(DerivedPtr)) {
1412 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1413 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1416 // isKnownNonNull -> nonnull attribute
1417 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI))
1418 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1420 // isDereferenceablePointer -> deref attribute
1421 if (isDereferenceablePointer(DerivedPtr, DL)) {
1422 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1423 uint64_t Bytes = A->getDereferenceableBytes();
1424 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1428 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1429 // Canonicalize on the type from the uses to the defs
1431 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1435 return visitCallSite(II);
1438 // InvokeInst simplification
1440 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1441 return visitCallSite(&II);
1444 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1445 /// passed through the varargs area, we can eliminate the use of the cast.
1446 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1447 const DataLayout &DL,
1448 const CastInst *const CI,
1450 if (!CI->isLosslessCast())
1453 // If this is a GC intrinsic, avoid munging types. We need types for
1454 // statepoint reconstruction in SelectionDAG.
1455 // TODO: This is probably something which should be expanded to all
1456 // intrinsics since the entire point of intrinsics is that
1457 // they are understandable by the optimizer.
1458 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1461 // The size of ByVal or InAlloca arguments is derived from the type, so we
1462 // can't change to a type with a different size. If the size were
1463 // passed explicitly we could avoid this check.
1464 if (!CS.isByValOrInAllocaArgument(ix))
1468 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1469 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1470 if (!SrcTy->isSized() || !DstTy->isSized())
1472 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1477 // Try to fold some different type of calls here.
1478 // Currently we're only working with the checking functions, memcpy_chk,
1479 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1480 // strcat_chk and strncat_chk.
1481 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1482 if (!CI->getCalledFunction()) return nullptr;
1484 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1485 ReplaceInstUsesWith(*From, With);
1487 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1488 if (Value *With = Simplifier.optimizeCall(CI)) {
1490 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1496 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1497 // Strip off at most one level of pointer casts, looking for an alloca. This
1498 // is good enough in practice and simpler than handling any number of casts.
1499 Value *Underlying = TrampMem->stripPointerCasts();
1500 if (Underlying != TrampMem &&
1501 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1503 if (!isa<AllocaInst>(Underlying))
1506 IntrinsicInst *InitTrampoline = nullptr;
1507 for (User *U : TrampMem->users()) {
1508 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1511 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1513 // More than one init_trampoline writes to this value. Give up.
1515 InitTrampoline = II;
1518 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1519 // Allow any number of calls to adjust.trampoline.
1524 // No call to init.trampoline found.
1525 if (!InitTrampoline)
1528 // Check that the alloca is being used in the expected way.
1529 if (InitTrampoline->getOperand(0) != TrampMem)
1532 return InitTrampoline;
1535 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1537 // Visit all the previous instructions in the basic block, and try to find a
1538 // init.trampoline which has a direct path to the adjust.trampoline.
1539 for (BasicBlock::iterator I = AdjustTramp,
1540 E = AdjustTramp->getParent()->begin(); I != E; ) {
1541 Instruction *Inst = --I;
1542 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1543 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1544 II->getOperand(0) == TrampMem)
1546 if (Inst->mayWriteToMemory())
1552 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1553 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1554 // to a direct call to a function. Otherwise return NULL.
1556 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1557 Callee = Callee->stripPointerCasts();
1558 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1560 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1563 Value *TrampMem = AdjustTramp->getOperand(0);
1565 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1567 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1572 // visitCallSite - Improvements for call and invoke instructions.
1574 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1576 if (isAllocLikeFn(CS.getInstruction(), TLI))
1577 return visitAllocSite(*CS.getInstruction());
1579 bool Changed = false;
1581 // Mark any parameters that are known to be non-null with the nonnull
1582 // attribute. This is helpful for inlining calls to functions with null
1583 // checks on their arguments.
1585 for (Value *V : CS.args()) {
1586 if (V->getType()->isPointerTy() && !CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1587 isKnownNonNullAt(V, CS.getInstruction(), DT, TLI)) {
1588 AttributeSet AS = CS.getAttributes();
1589 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1590 Attribute::NonNull);
1591 CS.setAttributes(AS);
1596 assert(ArgNo == CS.arg_size() && "sanity check");
1598 // If the callee is a pointer to a function, attempt to move any casts to the
1599 // arguments of the call/invoke.
1600 Value *Callee = CS.getCalledValue();
1601 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1604 if (Function *CalleeF = dyn_cast<Function>(Callee))
1605 // If the call and callee calling conventions don't match, this call must
1606 // be unreachable, as the call is undefined.
1607 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1608 // Only do this for calls to a function with a body. A prototype may
1609 // not actually end up matching the implementation's calling conv for a
1610 // variety of reasons (e.g. it may be written in assembly).
1611 !CalleeF->isDeclaration()) {
1612 Instruction *OldCall = CS.getInstruction();
1613 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1614 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1616 // If OldCall does not return void then replaceAllUsesWith undef.
1617 // This allows ValueHandlers and custom metadata to adjust itself.
1618 if (!OldCall->getType()->isVoidTy())
1619 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1620 if (isa<CallInst>(OldCall))
1621 return EraseInstFromFunction(*OldCall);
1623 // We cannot remove an invoke, because it would change the CFG, just
1624 // change the callee to a null pointer.
1625 cast<InvokeInst>(OldCall)->setCalledFunction(
1626 Constant::getNullValue(CalleeF->getType()));
1630 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1631 // If CS does not return void then replaceAllUsesWith undef.
1632 // This allows ValueHandlers and custom metadata to adjust itself.
1633 if (!CS.getInstruction()->getType()->isVoidTy())
1634 ReplaceInstUsesWith(*CS.getInstruction(),
1635 UndefValue::get(CS.getInstruction()->getType()));
1637 if (isa<InvokeInst>(CS.getInstruction())) {
1638 // Can't remove an invoke because we cannot change the CFG.
1642 // This instruction is not reachable, just remove it. We insert a store to
1643 // undef so that we know that this code is not reachable, despite the fact
1644 // that we can't modify the CFG here.
1645 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1646 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1647 CS.getInstruction());
1649 return EraseInstFromFunction(*CS.getInstruction());
1652 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1653 return transformCallThroughTrampoline(CS, II);
1655 PointerType *PTy = cast<PointerType>(Callee->getType());
1656 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1657 if (FTy->isVarArg()) {
1658 int ix = FTy->getNumParams();
1659 // See if we can optimize any arguments passed through the varargs area of
1661 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1662 E = CS.arg_end(); I != E; ++I, ++ix) {
1663 CastInst *CI = dyn_cast<CastInst>(*I);
1664 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1665 *I = CI->getOperand(0);
1671 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1672 // Inline asm calls cannot throw - mark them 'nounwind'.
1673 CS.setDoesNotThrow();
1677 // Try to optimize the call if possible, we require DataLayout for most of
1678 // this. None of these calls are seen as possibly dead so go ahead and
1679 // delete the instruction now.
1680 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1681 Instruction *I = tryOptimizeCall(CI);
1682 // If we changed something return the result, etc. Otherwise let
1683 // the fallthrough check.
1684 if (I) return EraseInstFromFunction(*I);
1687 return Changed ? CS.getInstruction() : nullptr;
1690 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1691 // attempt to move the cast to the arguments of the call/invoke.
1693 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1695 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1698 // The prototype of thunks are a lie, don't try to directly call such
1700 if (Callee->hasFnAttribute("thunk"))
1702 Instruction *Caller = CS.getInstruction();
1703 const AttributeSet &CallerPAL = CS.getAttributes();
1705 // Okay, this is a cast from a function to a different type. Unless doing so
1706 // would cause a type conversion of one of our arguments, change this call to
1707 // be a direct call with arguments casted to the appropriate types.
1709 FunctionType *FT = Callee->getFunctionType();
1710 Type *OldRetTy = Caller->getType();
1711 Type *NewRetTy = FT->getReturnType();
1713 // Check to see if we are changing the return type...
1714 if (OldRetTy != NewRetTy) {
1716 if (NewRetTy->isStructTy())
1717 return false; // TODO: Handle multiple return values.
1719 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1720 if (Callee->isDeclaration())
1721 return false; // Cannot transform this return value.
1723 if (!Caller->use_empty() &&
1724 // void -> non-void is handled specially
1725 !NewRetTy->isVoidTy())
1726 return false; // Cannot transform this return value.
1729 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1730 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1731 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1732 return false; // Attribute not compatible with transformed value.
1735 // If the callsite is an invoke instruction, and the return value is used by
1736 // a PHI node in a successor, we cannot change the return type of the call
1737 // because there is no place to put the cast instruction (without breaking
1738 // the critical edge). Bail out in this case.
1739 if (!Caller->use_empty())
1740 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1741 for (User *U : II->users())
1742 if (PHINode *PN = dyn_cast<PHINode>(U))
1743 if (PN->getParent() == II->getNormalDest() ||
1744 PN->getParent() == II->getUnwindDest())
1748 unsigned NumActualArgs = CS.arg_size();
1749 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1751 // Prevent us turning:
1752 // declare void @takes_i32_inalloca(i32* inalloca)
1753 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1756 // call void @takes_i32_inalloca(i32* null)
1758 // Similarly, avoid folding away bitcasts of byval calls.
1759 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1760 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1763 CallSite::arg_iterator AI = CS.arg_begin();
1764 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1765 Type *ParamTy = FT->getParamType(i);
1766 Type *ActTy = (*AI)->getType();
1768 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1769 return false; // Cannot transform this parameter value.
1771 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1772 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1773 return false; // Attribute not compatible with transformed value.
1775 if (CS.isInAllocaArgument(i))
1776 return false; // Cannot transform to and from inalloca.
1778 // If the parameter is passed as a byval argument, then we have to have a
1779 // sized type and the sized type has to have the same size as the old type.
1780 if (ParamTy != ActTy &&
1781 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1782 Attribute::ByVal)) {
1783 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1784 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1787 Type *CurElTy = ActTy->getPointerElementType();
1788 if (DL.getTypeAllocSize(CurElTy) !=
1789 DL.getTypeAllocSize(ParamPTy->getElementType()))
1794 if (Callee->isDeclaration()) {
1795 // Do not delete arguments unless we have a function body.
1796 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1799 // If the callee is just a declaration, don't change the varargsness of the
1800 // call. We don't want to introduce a varargs call where one doesn't
1802 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1803 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1806 // If both the callee and the cast type are varargs, we still have to make
1807 // sure the number of fixed parameters are the same or we have the same
1808 // ABI issues as if we introduce a varargs call.
1809 if (FT->isVarArg() &&
1810 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1811 FT->getNumParams() !=
1812 cast<FunctionType>(APTy->getElementType())->getNumParams())
1816 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1817 !CallerPAL.isEmpty())
1818 // In this case we have more arguments than the new function type, but we
1819 // won't be dropping them. Check that these extra arguments have attributes
1820 // that are compatible with being a vararg call argument.
1821 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1822 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1823 if (Index <= FT->getNumParams())
1826 // Check if it has an attribute that's incompatible with varargs.
1827 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1828 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1833 // Okay, we decided that this is a safe thing to do: go ahead and start
1834 // inserting cast instructions as necessary.
1835 std::vector<Value*> Args;
1836 Args.reserve(NumActualArgs);
1837 SmallVector<AttributeSet, 8> attrVec;
1838 attrVec.reserve(NumCommonArgs);
1840 // Get any return attributes.
1841 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1843 // If the return value is not being used, the type may not be compatible
1844 // with the existing attributes. Wipe out any problematic attributes.
1845 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1847 // Add the new return attributes.
1848 if (RAttrs.hasAttributes())
1849 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1850 AttributeSet::ReturnIndex, RAttrs));
1852 AI = CS.arg_begin();
1853 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1854 Type *ParamTy = FT->getParamType(i);
1856 if ((*AI)->getType() == ParamTy) {
1857 Args.push_back(*AI);
1859 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1862 // Add any parameter attributes.
1863 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1864 if (PAttrs.hasAttributes())
1865 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1869 // If the function takes more arguments than the call was taking, add them
1871 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1872 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1874 // If we are removing arguments to the function, emit an obnoxious warning.
1875 if (FT->getNumParams() < NumActualArgs) {
1876 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1877 if (FT->isVarArg()) {
1878 // Add all of the arguments in their promoted form to the arg list.
1879 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1880 Type *PTy = getPromotedType((*AI)->getType());
1881 if (PTy != (*AI)->getType()) {
1882 // Must promote to pass through va_arg area!
1883 Instruction::CastOps opcode =
1884 CastInst::getCastOpcode(*AI, false, PTy, false);
1885 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1887 Args.push_back(*AI);
1890 // Add any parameter attributes.
1891 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1892 if (PAttrs.hasAttributes())
1893 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1899 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1900 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1901 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1903 if (NewRetTy->isVoidTy())
1904 Caller->setName(""); // Void type should not have a name.
1906 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1910 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1911 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1912 II->getUnwindDest(), Args);
1914 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1915 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1917 CallInst *CI = cast<CallInst>(Caller);
1918 NC = Builder->CreateCall(Callee, Args);
1920 if (CI->isTailCall())
1921 cast<CallInst>(NC)->setTailCall();
1922 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1923 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1926 // Insert a cast of the return type as necessary.
1928 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1929 if (!NV->getType()->isVoidTy()) {
1930 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1931 NC->setDebugLoc(Caller->getDebugLoc());
1933 // If this is an invoke instruction, we should insert it after the first
1934 // non-phi, instruction in the normal successor block.
1935 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1936 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1937 InsertNewInstBefore(NC, *I);
1939 // Otherwise, it's a call, just insert cast right after the call.
1940 InsertNewInstBefore(NC, *Caller);
1942 Worklist.AddUsersToWorkList(*Caller);
1944 NV = UndefValue::get(Caller->getType());
1948 if (!Caller->use_empty())
1949 ReplaceInstUsesWith(*Caller, NV);
1950 else if (Caller->hasValueHandle()) {
1951 if (OldRetTy == NV->getType())
1952 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1954 // We cannot call ValueIsRAUWd with a different type, and the
1955 // actual tracked value will disappear.
1956 ValueHandleBase::ValueIsDeleted(Caller);
1959 EraseInstFromFunction(*Caller);
1963 // transformCallThroughTrampoline - Turn a call to a function created by
1964 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1965 // underlying function.
1968 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1969 IntrinsicInst *Tramp) {
1970 Value *Callee = CS.getCalledValue();
1971 PointerType *PTy = cast<PointerType>(Callee->getType());
1972 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1973 const AttributeSet &Attrs = CS.getAttributes();
1975 // If the call already has the 'nest' attribute somewhere then give up -
1976 // otherwise 'nest' would occur twice after splicing in the chain.
1977 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1981 "transformCallThroughTrampoline called with incorrect CallSite.");
1983 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1984 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1985 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1987 const AttributeSet &NestAttrs = NestF->getAttributes();
1988 if (!NestAttrs.isEmpty()) {
1989 unsigned NestIdx = 1;
1990 Type *NestTy = nullptr;
1991 AttributeSet NestAttr;
1993 // Look for a parameter marked with the 'nest' attribute.
1994 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1995 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1996 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1997 // Record the parameter type and any other attributes.
1999 NestAttr = NestAttrs.getParamAttributes(NestIdx);
2004 Instruction *Caller = CS.getInstruction();
2005 std::vector<Value*> NewArgs;
2006 NewArgs.reserve(CS.arg_size() + 1);
2008 SmallVector<AttributeSet, 8> NewAttrs;
2009 NewAttrs.reserve(Attrs.getNumSlots() + 1);
2011 // Insert the nest argument into the call argument list, which may
2012 // mean appending it. Likewise for attributes.
2014 // Add any result attributes.
2015 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
2016 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2017 Attrs.getRetAttributes()));
2021 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
2023 if (Idx == NestIdx) {
2024 // Add the chain argument and attributes.
2025 Value *NestVal = Tramp->getArgOperand(2);
2026 if (NestVal->getType() != NestTy)
2027 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
2028 NewArgs.push_back(NestVal);
2029 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2036 // Add the original argument and attributes.
2037 NewArgs.push_back(*I);
2038 AttributeSet Attr = Attrs.getParamAttributes(Idx);
2039 if (Attr.hasAttributes(Idx)) {
2040 AttrBuilder B(Attr, Idx);
2041 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
2042 Idx + (Idx >= NestIdx), B));
2049 // Add any function attributes.
2050 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
2051 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
2052 Attrs.getFnAttributes()));
2054 // The trampoline may have been bitcast to a bogus type (FTy).
2055 // Handle this by synthesizing a new function type, equal to FTy
2056 // with the chain parameter inserted.
2058 std::vector<Type*> NewTypes;
2059 NewTypes.reserve(FTy->getNumParams()+1);
2061 // Insert the chain's type into the list of parameter types, which may
2062 // mean appending it.
2065 FunctionType::param_iterator I = FTy->param_begin(),
2066 E = FTy->param_end();
2070 // Add the chain's type.
2071 NewTypes.push_back(NestTy);
2076 // Add the original type.
2077 NewTypes.push_back(*I);
2083 // Replace the trampoline call with a direct call. Let the generic
2084 // code sort out any function type mismatches.
2085 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
2087 Constant *NewCallee =
2088 NestF->getType() == PointerType::getUnqual(NewFTy) ?
2089 NestF : ConstantExpr::getBitCast(NestF,
2090 PointerType::getUnqual(NewFTy));
2091 const AttributeSet &NewPAL =
2092 AttributeSet::get(FTy->getContext(), NewAttrs);
2094 Instruction *NewCaller;
2095 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2096 NewCaller = InvokeInst::Create(NewCallee,
2097 II->getNormalDest(), II->getUnwindDest(),
2099 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
2100 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
2102 NewCaller = CallInst::Create(NewCallee, NewArgs);
2103 if (cast<CallInst>(Caller)->isTailCall())
2104 cast<CallInst>(NewCaller)->setTailCall();
2105 cast<CallInst>(NewCaller)->
2106 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
2107 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
2114 // Replace the trampoline call with a direct call. Since there is no 'nest'
2115 // parameter, there is no need to adjust the argument list. Let the generic
2116 // code sort out any function type mismatches.
2117 Constant *NewCallee =
2118 NestF->getType() == PTy ? NestF :
2119 ConstantExpr::getBitCast(NestF, PTy);
2120 CS.setCalledFunction(NewCallee);
2121 return CS.getInstruction();