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);
799 case Intrinsic::x86_sse_storeu_ps:
800 case Intrinsic::x86_sse2_storeu_pd:
801 case Intrinsic::x86_sse2_storeu_dq:
802 // Turn X86 storeu -> store if the pointer is known aligned.
803 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
806 PointerType::getUnqual(II->getArgOperand(1)->getType());
807 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
808 return new StoreInst(II->getArgOperand(1), Ptr);
812 case Intrinsic::x86_sse_cvtss2si:
813 case Intrinsic::x86_sse_cvtss2si64:
814 case Intrinsic::x86_sse_cvttss2si:
815 case Intrinsic::x86_sse_cvttss2si64:
816 case Intrinsic::x86_sse2_cvtsd2si:
817 case Intrinsic::x86_sse2_cvtsd2si64:
818 case Intrinsic::x86_sse2_cvttsd2si:
819 case Intrinsic::x86_sse2_cvttsd2si64: {
820 // These intrinsics only demand the 0th element of their input vectors. If
821 // we can simplify the input based on that, do so now.
823 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
824 APInt DemandedElts(VWidth, 1);
825 APInt UndefElts(VWidth, 0);
826 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
827 DemandedElts, UndefElts)) {
828 II->setArgOperand(0, V);
834 // Constant fold ashr( <A x Bi>, Ci ).
835 // Constant fold lshr( <A x Bi>, Ci ).
836 // Constant fold shl( <A x Bi>, Ci ).
837 case Intrinsic::x86_sse2_psrai_d:
838 case Intrinsic::x86_sse2_psrai_w:
839 case Intrinsic::x86_avx2_psrai_d:
840 case Intrinsic::x86_avx2_psrai_w:
841 case Intrinsic::x86_sse2_psrli_d:
842 case Intrinsic::x86_sse2_psrli_q:
843 case Intrinsic::x86_sse2_psrli_w:
844 case Intrinsic::x86_avx2_psrli_d:
845 case Intrinsic::x86_avx2_psrli_q:
846 case Intrinsic::x86_avx2_psrli_w:
847 case Intrinsic::x86_sse2_pslli_d:
848 case Intrinsic::x86_sse2_pslli_q:
849 case Intrinsic::x86_sse2_pslli_w:
850 case Intrinsic::x86_avx2_pslli_d:
851 case Intrinsic::x86_avx2_pslli_q:
852 case Intrinsic::x86_avx2_pslli_w:
853 if (Value *V = SimplifyX86immshift(*II, *Builder))
854 return ReplaceInstUsesWith(*II, V);
857 case Intrinsic::x86_sse2_psra_d:
858 case Intrinsic::x86_sse2_psra_w:
859 case Intrinsic::x86_avx2_psra_d:
860 case Intrinsic::x86_avx2_psra_w:
861 case Intrinsic::x86_sse2_psrl_d:
862 case Intrinsic::x86_sse2_psrl_q:
863 case Intrinsic::x86_sse2_psrl_w:
864 case Intrinsic::x86_avx2_psrl_d:
865 case Intrinsic::x86_avx2_psrl_q:
866 case Intrinsic::x86_avx2_psrl_w:
867 case Intrinsic::x86_sse2_psll_d:
868 case Intrinsic::x86_sse2_psll_q:
869 case Intrinsic::x86_sse2_psll_w:
870 case Intrinsic::x86_avx2_psll_d:
871 case Intrinsic::x86_avx2_psll_q:
872 case Intrinsic::x86_avx2_psll_w: {
873 if (Value *V = SimplifyX86immshift(*II, *Builder))
874 return ReplaceInstUsesWith(*II, V);
876 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
877 // operand to compute the shift amount.
878 auto ShiftAmt = II->getArgOperand(1);
879 auto ShiftType = cast<VectorType>(ShiftAmt->getType());
880 assert(ShiftType->getPrimitiveSizeInBits() == 128 &&
881 "Unexpected packed shift size");
882 unsigned VWidth = ShiftType->getNumElements();
884 APInt DemandedElts = APInt::getLowBitsSet(VWidth, VWidth / 2);
885 APInt UndefElts(VWidth, 0);
887 SimplifyDemandedVectorElts(ShiftAmt, DemandedElts, UndefElts)) {
888 II->setArgOperand(1, V);
894 case Intrinsic::x86_sse41_pmovsxbd:
895 case Intrinsic::x86_sse41_pmovsxbq:
896 case Intrinsic::x86_sse41_pmovsxbw:
897 case Intrinsic::x86_sse41_pmovsxdq:
898 case Intrinsic::x86_sse41_pmovsxwd:
899 case Intrinsic::x86_sse41_pmovsxwq:
900 case Intrinsic::x86_avx2_pmovsxbd:
901 case Intrinsic::x86_avx2_pmovsxbq:
902 case Intrinsic::x86_avx2_pmovsxbw:
903 case Intrinsic::x86_avx2_pmovsxdq:
904 case Intrinsic::x86_avx2_pmovsxwd:
905 case Intrinsic::x86_avx2_pmovsxwq:
906 if (Value *V = SimplifyX86extend(*II, *Builder, true))
907 return ReplaceInstUsesWith(*II, V);
910 case Intrinsic::x86_sse41_pmovzxbd:
911 case Intrinsic::x86_sse41_pmovzxbq:
912 case Intrinsic::x86_sse41_pmovzxbw:
913 case Intrinsic::x86_sse41_pmovzxdq:
914 case Intrinsic::x86_sse41_pmovzxwd:
915 case Intrinsic::x86_sse41_pmovzxwq:
916 case Intrinsic::x86_avx2_pmovzxbd:
917 case Intrinsic::x86_avx2_pmovzxbq:
918 case Intrinsic::x86_avx2_pmovzxbw:
919 case Intrinsic::x86_avx2_pmovzxdq:
920 case Intrinsic::x86_avx2_pmovzxwd:
921 case Intrinsic::x86_avx2_pmovzxwq:
922 if (Value *V = SimplifyX86extend(*II, *Builder, false))
923 return ReplaceInstUsesWith(*II, V);
926 case Intrinsic::x86_sse41_insertps:
927 if (Value *V = SimplifyX86insertps(*II, *Builder))
928 return ReplaceInstUsesWith(*II, V);
931 case Intrinsic::x86_sse4a_insertqi: {
932 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
934 // TODO: eventually we should lower this intrinsic to IR
935 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
936 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
937 unsigned Index = CIIndex->getZExtValue();
938 // From AMD documentation: "a value of zero in the field length is
939 // defined as length of 64".
940 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
942 // From AMD documentation: "If the sum of the bit index + length field
943 // is greater than 64, the results are undefined".
944 unsigned End = Index + Length;
946 // Note that both field index and field length are 8-bit quantities.
947 // Since variables 'Index' and 'Length' are unsigned values
948 // obtained from zero-extending field index and field length
949 // respectively, their sum should never wrap around.
951 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
953 if (Length == 64 && Index == 0) {
954 Value *Vec = II->getArgOperand(1);
955 Value *Undef = UndefValue::get(Vec->getType());
956 const uint32_t Mask[] = { 0, 2 };
957 return ReplaceInstUsesWith(
959 Builder->CreateShuffleVector(
960 Vec, Undef, ConstantDataVector::get(
961 II->getContext(), makeArrayRef(Mask))));
962 } else if (auto Source =
963 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
964 if (Source->hasOneUse() &&
965 Source->getArgOperand(1) == II->getArgOperand(1)) {
966 // If the source of the insert has only one use and it's another
967 // insert (and they're both inserting from the same vector), try to
968 // bundle both together.
969 auto CISourceLength =
970 dyn_cast<ConstantInt>(Source->getArgOperand(2));
972 dyn_cast<ConstantInt>(Source->getArgOperand(3));
973 if (CISourceIndex && CISourceLength) {
974 unsigned SourceIndex = CISourceIndex->getZExtValue();
975 unsigned SourceLength = CISourceLength->getZExtValue();
976 unsigned SourceEnd = SourceIndex + SourceLength;
977 unsigned NewIndex, NewLength;
978 bool ShouldReplace = false;
979 if (Index <= SourceIndex && SourceIndex <= End) {
981 NewLength = std::max(End, SourceEnd) - NewIndex;
982 ShouldReplace = true;
983 } else if (SourceIndex <= Index && Index <= SourceEnd) {
984 NewIndex = SourceIndex;
985 NewLength = std::max(SourceEnd, End) - NewIndex;
986 ShouldReplace = true;
990 Constant *ConstantLength = ConstantInt::get(
991 II->getArgOperand(2)->getType(), NewLength, false);
992 Constant *ConstantIndex = ConstantInt::get(
993 II->getArgOperand(3)->getType(), NewIndex, false);
994 Value *Args[4] = { Source->getArgOperand(0),
995 II->getArgOperand(1), ConstantLength,
997 Module *M = CI.getParent()->getParent()->getParent();
999 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1000 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
1010 case Intrinsic::x86_sse41_pblendvb:
1011 case Intrinsic::x86_sse41_blendvps:
1012 case Intrinsic::x86_sse41_blendvpd:
1013 case Intrinsic::x86_avx_blendv_ps_256:
1014 case Intrinsic::x86_avx_blendv_pd_256:
1015 case Intrinsic::x86_avx2_pblendvb: {
1016 // Convert blendv* to vector selects if the mask is constant.
1017 // This optimization is convoluted because the intrinsic is defined as
1018 // getting a vector of floats or doubles for the ps and pd versions.
1019 // FIXME: That should be changed.
1021 Value *Op0 = II->getArgOperand(0);
1022 Value *Op1 = II->getArgOperand(1);
1023 Value *Mask = II->getArgOperand(2);
1025 // fold (blend A, A, Mask) -> A
1027 return ReplaceInstUsesWith(CI, Op0);
1029 // Zero Mask - select 1st argument.
1030 if (isa<ConstantAggregateZero>(Mask))
1031 return ReplaceInstUsesWith(CI, Op0);
1033 // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
1034 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
1035 auto Tyi1 = Builder->getInt1Ty();
1036 auto SelectorType = cast<VectorType>(Mask->getType());
1037 auto EltTy = SelectorType->getElementType();
1038 unsigned Size = SelectorType->getNumElements();
1042 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
1043 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
1044 "Wrong arguments for variable blend intrinsic");
1045 SmallVector<Constant *, 32> Selectors;
1046 for (unsigned I = 0; I < Size; ++I) {
1047 // The intrinsics only read the top bit
1050 Selector = C->getElementAsInteger(I);
1052 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
1053 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
1055 auto NewSelector = ConstantVector::get(Selectors);
1056 return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
1061 case Intrinsic::x86_avx_vpermilvar_ps:
1062 case Intrinsic::x86_avx_vpermilvar_ps_256:
1063 case Intrinsic::x86_avx_vpermilvar_pd:
1064 case Intrinsic::x86_avx_vpermilvar_pd_256: {
1065 // Convert vpermil* to shufflevector if the mask is constant.
1066 Value *V = II->getArgOperand(1);
1067 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1068 assert(Size == 8 || Size == 4 || Size == 2);
1069 uint32_t Indexes[8];
1070 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1071 // The intrinsics only read one or two bits, clear the rest.
1072 for (unsigned I = 0; I < Size; ++I) {
1073 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1074 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1075 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1079 } else if (isa<ConstantAggregateZero>(V)) {
1080 for (unsigned I = 0; I < Size; ++I)
1085 // The _256 variants are a bit trickier since the mask bits always index
1086 // into the corresponding 128 half. In order to convert to a generic
1087 // shuffle, we have to make that explicit.
1088 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1089 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1090 for (unsigned I = Size / 2; I < Size; ++I)
1091 Indexes[I] += Size / 2;
1094 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1095 auto V1 = II->getArgOperand(0);
1096 auto V2 = UndefValue::get(V1->getType());
1097 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1098 return ReplaceInstUsesWith(CI, Shuffle);
1101 case Intrinsic::x86_avx_vperm2f128_pd_256:
1102 case Intrinsic::x86_avx_vperm2f128_ps_256:
1103 case Intrinsic::x86_avx_vperm2f128_si_256:
1104 case Intrinsic::x86_avx2_vperm2i128:
1105 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1106 return ReplaceInstUsesWith(*II, V);
1109 case Intrinsic::ppc_altivec_vperm:
1110 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1111 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1112 // a vectorshuffle for little endian, we must undo the transformation
1113 // performed on vec_perm in altivec.h. That is, we must complement
1114 // the permutation mask with respect to 31 and reverse the order of
1116 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1117 assert(Mask->getType()->getVectorNumElements() == 16 &&
1118 "Bad type for intrinsic!");
1120 // Check that all of the elements are integer constants or undefs.
1121 bool AllEltsOk = true;
1122 for (unsigned i = 0; i != 16; ++i) {
1123 Constant *Elt = Mask->getAggregateElement(i);
1124 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1131 // Cast the input vectors to byte vectors.
1132 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1134 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1136 Value *Result = UndefValue::get(Op0->getType());
1138 // Only extract each element once.
1139 Value *ExtractedElts[32];
1140 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1142 for (unsigned i = 0; i != 16; ++i) {
1143 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1146 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1147 Idx &= 31; // Match the hardware behavior.
1148 if (DL.isLittleEndian())
1151 if (!ExtractedElts[Idx]) {
1152 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1153 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1154 ExtractedElts[Idx] =
1155 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1156 Builder->getInt32(Idx&15));
1159 // Insert this value into the result vector.
1160 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1161 Builder->getInt32(i));
1163 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1168 case Intrinsic::arm_neon_vld1:
1169 case Intrinsic::arm_neon_vld2:
1170 case Intrinsic::arm_neon_vld3:
1171 case Intrinsic::arm_neon_vld4:
1172 case Intrinsic::arm_neon_vld2lane:
1173 case Intrinsic::arm_neon_vld3lane:
1174 case Intrinsic::arm_neon_vld4lane:
1175 case Intrinsic::arm_neon_vst1:
1176 case Intrinsic::arm_neon_vst2:
1177 case Intrinsic::arm_neon_vst3:
1178 case Intrinsic::arm_neon_vst4:
1179 case Intrinsic::arm_neon_vst2lane:
1180 case Intrinsic::arm_neon_vst3lane:
1181 case Intrinsic::arm_neon_vst4lane: {
1182 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1183 unsigned AlignArg = II->getNumArgOperands() - 1;
1184 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1185 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1186 II->setArgOperand(AlignArg,
1187 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1194 case Intrinsic::arm_neon_vmulls:
1195 case Intrinsic::arm_neon_vmullu:
1196 case Intrinsic::aarch64_neon_smull:
1197 case Intrinsic::aarch64_neon_umull: {
1198 Value *Arg0 = II->getArgOperand(0);
1199 Value *Arg1 = II->getArgOperand(1);
1201 // Handle mul by zero first:
1202 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1203 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1206 // Check for constant LHS & RHS - in this case we just simplify.
1207 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1208 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1209 VectorType *NewVT = cast<VectorType>(II->getType());
1210 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1211 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1212 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1213 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1215 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1218 // Couldn't simplify - canonicalize constant to the RHS.
1219 std::swap(Arg0, Arg1);
1222 // Handle mul by one:
1223 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1224 if (ConstantInt *Splat =
1225 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1227 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1228 /*isSigned=*/!Zext);
1233 case Intrinsic::AMDGPU_rcp: {
1234 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1235 const APFloat &ArgVal = C->getValueAPF();
1236 APFloat Val(ArgVal.getSemantics(), 1.0);
1237 APFloat::opStatus Status = Val.divide(ArgVal,
1238 APFloat::rmNearestTiesToEven);
1239 // Only do this if it was exact and therefore not dependent on the
1241 if (Status == APFloat::opOK)
1242 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1247 case Intrinsic::stackrestore: {
1248 // If the save is right next to the restore, remove the restore. This can
1249 // happen when variable allocas are DCE'd.
1250 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1251 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1252 BasicBlock::iterator BI = SS;
1254 return EraseInstFromFunction(CI);
1258 // Scan down this block to see if there is another stack restore in the
1259 // same block without an intervening call/alloca.
1260 BasicBlock::iterator BI = II;
1261 TerminatorInst *TI = II->getParent()->getTerminator();
1262 bool CannotRemove = false;
1263 for (++BI; &*BI != TI; ++BI) {
1264 if (isa<AllocaInst>(BI)) {
1265 CannotRemove = true;
1268 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1269 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1270 // If there is a stackrestore below this one, remove this one.
1271 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1272 return EraseInstFromFunction(CI);
1273 // Otherwise, ignore the intrinsic.
1275 // If we found a non-intrinsic call, we can't remove the stack
1277 CannotRemove = true;
1283 // If the stack restore is in a return, resume, or unwind block and if there
1284 // are no allocas or calls between the restore and the return, nuke the
1286 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1287 return EraseInstFromFunction(CI);
1290 case Intrinsic::assume: {
1291 // Canonicalize assume(a && b) -> assume(a); assume(b);
1292 // Note: New assumption intrinsics created here are registered by
1293 // the InstCombineIRInserter object.
1294 Value *IIOperand = II->getArgOperand(0), *A, *B,
1295 *AssumeIntrinsic = II->getCalledValue();
1296 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1297 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1298 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1299 return EraseInstFromFunction(*II);
1301 // assume(!(a || b)) -> assume(!a); assume(!b);
1302 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1303 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1305 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1307 return EraseInstFromFunction(*II);
1310 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1311 // (if assume is valid at the load)
1312 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1313 Value *LHS = ICmp->getOperand(0);
1314 Value *RHS = ICmp->getOperand(1);
1315 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1316 isa<LoadInst>(LHS) &&
1317 isa<Constant>(RHS) &&
1318 RHS->getType()->isPointerTy() &&
1319 cast<Constant>(RHS)->isNullValue()) {
1320 LoadInst* LI = cast<LoadInst>(LHS);
1321 if (isValidAssumeForContext(II, LI, DT)) {
1322 MDNode *MD = MDNode::get(II->getContext(), None);
1323 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1324 return EraseInstFromFunction(*II);
1327 // TODO: apply nonnull return attributes to calls and invokes
1328 // TODO: apply range metadata for range check patterns?
1330 // If there is a dominating assume with the same condition as this one,
1331 // then this one is redundant, and should be removed.
1332 APInt KnownZero(1, 0), KnownOne(1, 0);
1333 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1334 if (KnownOne.isAllOnesValue())
1335 return EraseInstFromFunction(*II);
1339 case Intrinsic::experimental_gc_relocate: {
1340 // Translate facts known about a pointer before relocating into
1341 // facts about the relocate value, while being careful to
1342 // preserve relocation semantics.
1343 GCRelocateOperands Operands(II);
1344 Value *DerivedPtr = Operands.getDerivedPtr();
1345 auto *GCRelocateType = cast<PointerType>(II->getType());
1347 // Remove the relocation if unused, note that this check is required
1348 // to prevent the cases below from looping forever.
1349 if (II->use_empty())
1350 return EraseInstFromFunction(*II);
1352 // Undef is undef, even after relocation.
1353 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1354 // most practical collectors, but there was discussion in the review thread
1355 // about whether it was legal for all possible collectors.
1356 if (isa<UndefValue>(DerivedPtr)) {
1357 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1358 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1361 // The relocation of null will be null for most any collector.
1362 // TODO: provide a hook for this in GCStrategy. There might be some weird
1363 // collector this property does not hold for.
1364 if (isa<ConstantPointerNull>(DerivedPtr)) {
1365 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1366 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1369 // isKnownNonNull -> nonnull attribute
1370 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI))
1371 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1373 // isDereferenceablePointer -> deref attribute
1374 if (isDereferenceablePointer(DerivedPtr, DL)) {
1375 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1376 uint64_t Bytes = A->getDereferenceableBytes();
1377 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1381 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1382 // Canonicalize on the type from the uses to the defs
1384 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1388 return visitCallSite(II);
1391 // InvokeInst simplification
1393 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1394 return visitCallSite(&II);
1397 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1398 /// passed through the varargs area, we can eliminate the use of the cast.
1399 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1400 const DataLayout &DL,
1401 const CastInst *const CI,
1403 if (!CI->isLosslessCast())
1406 // If this is a GC intrinsic, avoid munging types. We need types for
1407 // statepoint reconstruction in SelectionDAG.
1408 // TODO: This is probably something which should be expanded to all
1409 // intrinsics since the entire point of intrinsics is that
1410 // they are understandable by the optimizer.
1411 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1414 // The size of ByVal or InAlloca arguments is derived from the type, so we
1415 // can't change to a type with a different size. If the size were
1416 // passed explicitly we could avoid this check.
1417 if (!CS.isByValOrInAllocaArgument(ix))
1421 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1422 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1423 if (!SrcTy->isSized() || !DstTy->isSized())
1425 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1430 // Try to fold some different type of calls here.
1431 // Currently we're only working with the checking functions, memcpy_chk,
1432 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1433 // strcat_chk and strncat_chk.
1434 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1435 if (!CI->getCalledFunction()) return nullptr;
1437 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1438 ReplaceInstUsesWith(*From, With);
1440 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1441 if (Value *With = Simplifier.optimizeCall(CI)) {
1443 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1449 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1450 // Strip off at most one level of pointer casts, looking for an alloca. This
1451 // is good enough in practice and simpler than handling any number of casts.
1452 Value *Underlying = TrampMem->stripPointerCasts();
1453 if (Underlying != TrampMem &&
1454 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1456 if (!isa<AllocaInst>(Underlying))
1459 IntrinsicInst *InitTrampoline = nullptr;
1460 for (User *U : TrampMem->users()) {
1461 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1464 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1466 // More than one init_trampoline writes to this value. Give up.
1468 InitTrampoline = II;
1471 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1472 // Allow any number of calls to adjust.trampoline.
1477 // No call to init.trampoline found.
1478 if (!InitTrampoline)
1481 // Check that the alloca is being used in the expected way.
1482 if (InitTrampoline->getOperand(0) != TrampMem)
1485 return InitTrampoline;
1488 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1490 // Visit all the previous instructions in the basic block, and try to find a
1491 // init.trampoline which has a direct path to the adjust.trampoline.
1492 for (BasicBlock::iterator I = AdjustTramp,
1493 E = AdjustTramp->getParent()->begin(); I != E; ) {
1494 Instruction *Inst = --I;
1495 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1496 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1497 II->getOperand(0) == TrampMem)
1499 if (Inst->mayWriteToMemory())
1505 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1506 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1507 // to a direct call to a function. Otherwise return NULL.
1509 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1510 Callee = Callee->stripPointerCasts();
1511 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1513 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1516 Value *TrampMem = AdjustTramp->getOperand(0);
1518 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1520 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1525 // visitCallSite - Improvements for call and invoke instructions.
1527 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1529 if (isAllocLikeFn(CS.getInstruction(), TLI))
1530 return visitAllocSite(*CS.getInstruction());
1532 bool Changed = false;
1534 // Mark any parameters that are known to be non-null with the nonnull
1535 // attribute. This is helpful for inlining calls to functions with null
1536 // checks on their arguments.
1538 for (Value *V : CS.args()) {
1539 if (!CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1540 isKnownNonNull(V)) {
1541 AttributeSet AS = CS.getAttributes();
1542 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1543 Attribute::NonNull);
1544 CS.setAttributes(AS);
1549 assert(ArgNo == CS.arg_size() && "sanity check");
1551 // If the callee is a pointer to a function, attempt to move any casts to the
1552 // arguments of the call/invoke.
1553 Value *Callee = CS.getCalledValue();
1554 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1557 if (Function *CalleeF = dyn_cast<Function>(Callee))
1558 // If the call and callee calling conventions don't match, this call must
1559 // be unreachable, as the call is undefined.
1560 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1561 // Only do this for calls to a function with a body. A prototype may
1562 // not actually end up matching the implementation's calling conv for a
1563 // variety of reasons (e.g. it may be written in assembly).
1564 !CalleeF->isDeclaration()) {
1565 Instruction *OldCall = CS.getInstruction();
1566 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1567 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1569 // If OldCall does not return void then replaceAllUsesWith undef.
1570 // This allows ValueHandlers and custom metadata to adjust itself.
1571 if (!OldCall->getType()->isVoidTy())
1572 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1573 if (isa<CallInst>(OldCall))
1574 return EraseInstFromFunction(*OldCall);
1576 // We cannot remove an invoke, because it would change the CFG, just
1577 // change the callee to a null pointer.
1578 cast<InvokeInst>(OldCall)->setCalledFunction(
1579 Constant::getNullValue(CalleeF->getType()));
1583 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1584 // If CS does not return void then replaceAllUsesWith undef.
1585 // This allows ValueHandlers and custom metadata to adjust itself.
1586 if (!CS.getInstruction()->getType()->isVoidTy())
1587 ReplaceInstUsesWith(*CS.getInstruction(),
1588 UndefValue::get(CS.getInstruction()->getType()));
1590 if (isa<InvokeInst>(CS.getInstruction())) {
1591 // Can't remove an invoke because we cannot change the CFG.
1595 // This instruction is not reachable, just remove it. We insert a store to
1596 // undef so that we know that this code is not reachable, despite the fact
1597 // that we can't modify the CFG here.
1598 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1599 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1600 CS.getInstruction());
1602 return EraseInstFromFunction(*CS.getInstruction());
1605 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1606 return transformCallThroughTrampoline(CS, II);
1608 PointerType *PTy = cast<PointerType>(Callee->getType());
1609 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1610 if (FTy->isVarArg()) {
1611 int ix = FTy->getNumParams();
1612 // See if we can optimize any arguments passed through the varargs area of
1614 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1615 E = CS.arg_end(); I != E; ++I, ++ix) {
1616 CastInst *CI = dyn_cast<CastInst>(*I);
1617 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1618 *I = CI->getOperand(0);
1624 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1625 // Inline asm calls cannot throw - mark them 'nounwind'.
1626 CS.setDoesNotThrow();
1630 // Try to optimize the call if possible, we require DataLayout for most of
1631 // this. None of these calls are seen as possibly dead so go ahead and
1632 // delete the instruction now.
1633 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1634 Instruction *I = tryOptimizeCall(CI);
1635 // If we changed something return the result, etc. Otherwise let
1636 // the fallthrough check.
1637 if (I) return EraseInstFromFunction(*I);
1640 return Changed ? CS.getInstruction() : nullptr;
1643 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1644 // attempt to move the cast to the arguments of the call/invoke.
1646 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1648 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1651 // The prototype of thunks are a lie, don't try to directly call such
1653 if (Callee->hasFnAttribute("thunk"))
1655 Instruction *Caller = CS.getInstruction();
1656 const AttributeSet &CallerPAL = CS.getAttributes();
1658 // Okay, this is a cast from a function to a different type. Unless doing so
1659 // would cause a type conversion of one of our arguments, change this call to
1660 // be a direct call with arguments casted to the appropriate types.
1662 FunctionType *FT = Callee->getFunctionType();
1663 Type *OldRetTy = Caller->getType();
1664 Type *NewRetTy = FT->getReturnType();
1666 // Check to see if we are changing the return type...
1667 if (OldRetTy != NewRetTy) {
1669 if (NewRetTy->isStructTy())
1670 return false; // TODO: Handle multiple return values.
1672 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1673 if (Callee->isDeclaration())
1674 return false; // Cannot transform this return value.
1676 if (!Caller->use_empty() &&
1677 // void -> non-void is handled specially
1678 !NewRetTy->isVoidTy())
1679 return false; // Cannot transform this return value.
1682 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1683 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1684 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1685 return false; // Attribute not compatible with transformed value.
1688 // If the callsite is an invoke instruction, and the return value is used by
1689 // a PHI node in a successor, we cannot change the return type of the call
1690 // because there is no place to put the cast instruction (without breaking
1691 // the critical edge). Bail out in this case.
1692 if (!Caller->use_empty())
1693 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1694 for (User *U : II->users())
1695 if (PHINode *PN = dyn_cast<PHINode>(U))
1696 if (PN->getParent() == II->getNormalDest() ||
1697 PN->getParent() == II->getUnwindDest())
1701 unsigned NumActualArgs = CS.arg_size();
1702 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1704 // Prevent us turning:
1705 // declare void @takes_i32_inalloca(i32* inalloca)
1706 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1709 // call void @takes_i32_inalloca(i32* null)
1711 // Similarly, avoid folding away bitcasts of byval calls.
1712 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1713 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1716 CallSite::arg_iterator AI = CS.arg_begin();
1717 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1718 Type *ParamTy = FT->getParamType(i);
1719 Type *ActTy = (*AI)->getType();
1721 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1722 return false; // Cannot transform this parameter value.
1724 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1725 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1726 return false; // Attribute not compatible with transformed value.
1728 if (CS.isInAllocaArgument(i))
1729 return false; // Cannot transform to and from inalloca.
1731 // If the parameter is passed as a byval argument, then we have to have a
1732 // sized type and the sized type has to have the same size as the old type.
1733 if (ParamTy != ActTy &&
1734 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1735 Attribute::ByVal)) {
1736 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1737 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1740 Type *CurElTy = ActTy->getPointerElementType();
1741 if (DL.getTypeAllocSize(CurElTy) !=
1742 DL.getTypeAllocSize(ParamPTy->getElementType()))
1747 if (Callee->isDeclaration()) {
1748 // Do not delete arguments unless we have a function body.
1749 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1752 // If the callee is just a declaration, don't change the varargsness of the
1753 // call. We don't want to introduce a varargs call where one doesn't
1755 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1756 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1759 // If both the callee and the cast type are varargs, we still have to make
1760 // sure the number of fixed parameters are the same or we have the same
1761 // ABI issues as if we introduce a varargs call.
1762 if (FT->isVarArg() &&
1763 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1764 FT->getNumParams() !=
1765 cast<FunctionType>(APTy->getElementType())->getNumParams())
1769 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1770 !CallerPAL.isEmpty())
1771 // In this case we have more arguments than the new function type, but we
1772 // won't be dropping them. Check that these extra arguments have attributes
1773 // that are compatible with being a vararg call argument.
1774 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1775 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1776 if (Index <= FT->getNumParams())
1779 // Check if it has an attribute that's incompatible with varargs.
1780 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1781 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1786 // Okay, we decided that this is a safe thing to do: go ahead and start
1787 // inserting cast instructions as necessary.
1788 std::vector<Value*> Args;
1789 Args.reserve(NumActualArgs);
1790 SmallVector<AttributeSet, 8> attrVec;
1791 attrVec.reserve(NumCommonArgs);
1793 // Get any return attributes.
1794 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1796 // If the return value is not being used, the type may not be compatible
1797 // with the existing attributes. Wipe out any problematic attributes.
1798 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1800 // Add the new return attributes.
1801 if (RAttrs.hasAttributes())
1802 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1803 AttributeSet::ReturnIndex, RAttrs));
1805 AI = CS.arg_begin();
1806 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1807 Type *ParamTy = FT->getParamType(i);
1809 if ((*AI)->getType() == ParamTy) {
1810 Args.push_back(*AI);
1812 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1815 // Add any parameter attributes.
1816 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1817 if (PAttrs.hasAttributes())
1818 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1822 // If the function takes more arguments than the call was taking, add them
1824 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1825 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1827 // If we are removing arguments to the function, emit an obnoxious warning.
1828 if (FT->getNumParams() < NumActualArgs) {
1829 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1830 if (FT->isVarArg()) {
1831 // Add all of the arguments in their promoted form to the arg list.
1832 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1833 Type *PTy = getPromotedType((*AI)->getType());
1834 if (PTy != (*AI)->getType()) {
1835 // Must promote to pass through va_arg area!
1836 Instruction::CastOps opcode =
1837 CastInst::getCastOpcode(*AI, false, PTy, false);
1838 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1840 Args.push_back(*AI);
1843 // Add any parameter attributes.
1844 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1845 if (PAttrs.hasAttributes())
1846 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1852 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1853 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1854 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1856 if (NewRetTy->isVoidTy())
1857 Caller->setName(""); // Void type should not have a name.
1859 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1863 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1864 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1865 II->getUnwindDest(), Args);
1867 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1868 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1870 CallInst *CI = cast<CallInst>(Caller);
1871 NC = Builder->CreateCall(Callee, Args);
1873 if (CI->isTailCall())
1874 cast<CallInst>(NC)->setTailCall();
1875 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1876 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1879 // Insert a cast of the return type as necessary.
1881 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1882 if (!NV->getType()->isVoidTy()) {
1883 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1884 NC->setDebugLoc(Caller->getDebugLoc());
1886 // If this is an invoke instruction, we should insert it after the first
1887 // non-phi, instruction in the normal successor block.
1888 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1889 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1890 InsertNewInstBefore(NC, *I);
1892 // Otherwise, it's a call, just insert cast right after the call.
1893 InsertNewInstBefore(NC, *Caller);
1895 Worklist.AddUsersToWorkList(*Caller);
1897 NV = UndefValue::get(Caller->getType());
1901 if (!Caller->use_empty())
1902 ReplaceInstUsesWith(*Caller, NV);
1903 else if (Caller->hasValueHandle()) {
1904 if (OldRetTy == NV->getType())
1905 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1907 // We cannot call ValueIsRAUWd with a different type, and the
1908 // actual tracked value will disappear.
1909 ValueHandleBase::ValueIsDeleted(Caller);
1912 EraseInstFromFunction(*Caller);
1916 // transformCallThroughTrampoline - Turn a call to a function created by
1917 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1918 // underlying function.
1921 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1922 IntrinsicInst *Tramp) {
1923 Value *Callee = CS.getCalledValue();
1924 PointerType *PTy = cast<PointerType>(Callee->getType());
1925 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1926 const AttributeSet &Attrs = CS.getAttributes();
1928 // If the call already has the 'nest' attribute somewhere then give up -
1929 // otherwise 'nest' would occur twice after splicing in the chain.
1930 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1934 "transformCallThroughTrampoline called with incorrect CallSite.");
1936 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1937 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1938 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1940 const AttributeSet &NestAttrs = NestF->getAttributes();
1941 if (!NestAttrs.isEmpty()) {
1942 unsigned NestIdx = 1;
1943 Type *NestTy = nullptr;
1944 AttributeSet NestAttr;
1946 // Look for a parameter marked with the 'nest' attribute.
1947 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1948 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1949 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1950 // Record the parameter type and any other attributes.
1952 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1957 Instruction *Caller = CS.getInstruction();
1958 std::vector<Value*> NewArgs;
1959 NewArgs.reserve(CS.arg_size() + 1);
1961 SmallVector<AttributeSet, 8> NewAttrs;
1962 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1964 // Insert the nest argument into the call argument list, which may
1965 // mean appending it. Likewise for attributes.
1967 // Add any result attributes.
1968 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1969 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1970 Attrs.getRetAttributes()));
1974 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1976 if (Idx == NestIdx) {
1977 // Add the chain argument and attributes.
1978 Value *NestVal = Tramp->getArgOperand(2);
1979 if (NestVal->getType() != NestTy)
1980 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1981 NewArgs.push_back(NestVal);
1982 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1989 // Add the original argument and attributes.
1990 NewArgs.push_back(*I);
1991 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1992 if (Attr.hasAttributes(Idx)) {
1993 AttrBuilder B(Attr, Idx);
1994 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1995 Idx + (Idx >= NestIdx), B));
2002 // Add any function attributes.
2003 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
2004 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
2005 Attrs.getFnAttributes()));
2007 // The trampoline may have been bitcast to a bogus type (FTy).
2008 // Handle this by synthesizing a new function type, equal to FTy
2009 // with the chain parameter inserted.
2011 std::vector<Type*> NewTypes;
2012 NewTypes.reserve(FTy->getNumParams()+1);
2014 // Insert the chain's type into the list of parameter types, which may
2015 // mean appending it.
2018 FunctionType::param_iterator I = FTy->param_begin(),
2019 E = FTy->param_end();
2023 // Add the chain's type.
2024 NewTypes.push_back(NestTy);
2029 // Add the original type.
2030 NewTypes.push_back(*I);
2036 // Replace the trampoline call with a direct call. Let the generic
2037 // code sort out any function type mismatches.
2038 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
2040 Constant *NewCallee =
2041 NestF->getType() == PointerType::getUnqual(NewFTy) ?
2042 NestF : ConstantExpr::getBitCast(NestF,
2043 PointerType::getUnqual(NewFTy));
2044 const AttributeSet &NewPAL =
2045 AttributeSet::get(FTy->getContext(), NewAttrs);
2047 Instruction *NewCaller;
2048 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
2049 NewCaller = InvokeInst::Create(NewCallee,
2050 II->getNormalDest(), II->getUnwindDest(),
2052 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
2053 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
2055 NewCaller = CallInst::Create(NewCallee, NewArgs);
2056 if (cast<CallInst>(Caller)->isTailCall())
2057 cast<CallInst>(NewCaller)->setTailCall();
2058 cast<CallInst>(NewCaller)->
2059 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
2060 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
2067 // Replace the trampoline call with a direct call. Since there is no 'nest'
2068 // parameter, there is no need to adjust the argument list. Let the generic
2069 // code sort out any function type mismatches.
2070 Constant *NewCallee =
2071 NestF->getType() == PTy ? NestF :
2072 ConstantExpr::getBitCast(NestF, PTy);
2073 CS.setCalledFunction(NewCallee);
2074 return CS.getInstruction();