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, bool ShiftLeft) {
203 assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
205 // Simplify if count is constant.
206 auto Arg1 = II.getArgOperand(1);
207 auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1);
208 auto CDV = dyn_cast<ConstantDataVector>(Arg1);
209 auto CInt = dyn_cast<ConstantInt>(Arg1);
210 if (!CAZ && !CDV && !CInt)
215 // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
216 // operand to compute the shift amount.
217 auto VT = cast<VectorType>(CDV->getType());
218 unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits();
219 assert((64 % BitWidth) == 0 && "Unexpected packed shift size");
220 unsigned NumSubElts = 64 / BitWidth;
222 // Concatenate the sub-elements to create the 64-bit value.
223 for (unsigned i = 0; i != NumSubElts; ++i) {
224 unsigned SubEltIdx = (NumSubElts - 1) - i;
225 auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
226 Count = Count.shl(BitWidth);
227 Count |= SubElt->getValue().zextOrTrunc(64);
231 Count = CInt->getValue();
233 auto Vec = II.getArgOperand(0);
234 auto VT = cast<VectorType>(Vec->getType());
235 auto SVT = VT->getElementType();
236 unsigned VWidth = VT->getNumElements();
237 unsigned BitWidth = SVT->getPrimitiveSizeInBits();
239 // If shift-by-zero then just return the original value.
243 // Handle cases when Shift >= BitWidth.
244 if (Count.uge(BitWidth)) {
245 // If LogicalShift - just return zero.
247 return ConstantAggregateZero::get(VT);
249 // If ArithmeticShift - clamp Shift to (BitWidth - 1).
250 Count = APInt(64, BitWidth - 1);
253 // Get a constant vector of the same type as the first operand.
254 auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
255 auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
258 return Builder.CreateShl(Vec, ShiftVec);
261 return Builder.CreateLShr(Vec, ShiftVec);
263 return Builder.CreateAShr(Vec, ShiftVec);
266 static Value *SimplifyX86extend(const IntrinsicInst &II,
267 InstCombiner::BuilderTy &Builder,
269 VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType());
270 VectorType *DstTy = cast<VectorType>(II.getType());
271 unsigned NumDstElts = DstTy->getNumElements();
273 // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
274 SmallVector<int, 8> ShuffleMask;
275 for (int i = 0; i != (int)NumDstElts; ++i)
276 ShuffleMask.push_back(i);
278 Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0),
279 UndefValue::get(SrcTy), ShuffleMask);
280 return SignExtend ? Builder.CreateSExt(SV, DstTy)
281 : Builder.CreateZExt(SV, DstTy);
284 static Value *SimplifyX86insertps(const IntrinsicInst &II,
285 InstCombiner::BuilderTy &Builder) {
286 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
287 VectorType *VecTy = cast<VectorType>(II.getType());
288 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
290 // The immediate permute control byte looks like this:
291 // [3:0] - zero mask for each 32-bit lane
292 // [5:4] - select one 32-bit destination lane
293 // [7:6] - select one 32-bit source lane
295 uint8_t Imm = CInt->getZExtValue();
296 uint8_t ZMask = Imm & 0xf;
297 uint8_t DestLane = (Imm >> 4) & 0x3;
298 uint8_t SourceLane = (Imm >> 6) & 0x3;
300 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
302 // If all zero mask bits are set, this was just a weird way to
303 // generate a zero vector.
307 // Initialize by passing all of the first source bits through.
308 int ShuffleMask[4] = { 0, 1, 2, 3 };
310 // We may replace the second operand with the zero vector.
311 Value *V1 = II.getArgOperand(1);
314 // If the zero mask is being used with a single input or the zero mask
315 // overrides the destination lane, this is a shuffle with the zero vector.
316 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
317 (ZMask & (1 << DestLane))) {
319 // We may still move 32-bits of the first source vector from one lane
321 ShuffleMask[DestLane] = SourceLane;
322 // The zero mask may override the previous insert operation.
323 for (unsigned i = 0; i < 4; ++i)
324 if ((ZMask >> i) & 0x1)
325 ShuffleMask[i] = i + 4;
327 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
331 // Replace the selected destination lane with the selected source lane.
332 ShuffleMask[DestLane] = SourceLane + 4;
335 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
340 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
341 /// source vectors, unless a zero bit is set. If a zero bit is set,
342 /// then ignore that half of the mask and clear that half of the vector.
343 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
344 InstCombiner::BuilderTy &Builder) {
345 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
346 VectorType *VecTy = cast<VectorType>(II.getType());
347 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
349 // The immediate permute control byte looks like this:
350 // [1:0] - select 128 bits from sources for low half of destination
352 // [3] - zero low half of destination
353 // [5:4] - select 128 bits from sources for high half of destination
355 // [7] - zero high half of destination
357 uint8_t Imm = CInt->getZExtValue();
359 bool LowHalfZero = Imm & 0x08;
360 bool HighHalfZero = Imm & 0x80;
362 // If both zero mask bits are set, this was just a weird way to
363 // generate a zero vector.
364 if (LowHalfZero && HighHalfZero)
367 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
368 unsigned NumElts = VecTy->getNumElements();
369 unsigned HalfSize = NumElts / 2;
370 SmallVector<int, 8> ShuffleMask(NumElts);
372 // The high bit of the selection field chooses the 1st or 2nd operand.
373 bool LowInputSelect = Imm & 0x02;
374 bool HighInputSelect = Imm & 0x20;
376 // The low bit of the selection field chooses the low or high half
377 // of the selected operand.
378 bool LowHalfSelect = Imm & 0x01;
379 bool HighHalfSelect = Imm & 0x10;
381 // Determine which operand(s) are actually in use for this instruction.
382 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
383 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
385 // If needed, replace operands based on zero mask.
386 V0 = LowHalfZero ? ZeroVector : V0;
387 V1 = HighHalfZero ? ZeroVector : V1;
389 // Permute low half of result.
390 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
391 for (unsigned i = 0; i < HalfSize; ++i)
392 ShuffleMask[i] = StartIndex + i;
394 // Permute high half of result.
395 StartIndex = HighHalfSelect ? HalfSize : 0;
396 StartIndex += NumElts;
397 for (unsigned i = 0; i < HalfSize; ++i)
398 ShuffleMask[i + HalfSize] = StartIndex + i;
400 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
405 /// visitCallInst - CallInst simplification. This mostly only handles folding
406 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
407 /// the heavy lifting.
409 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
410 auto Args = CI.arg_operands();
411 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
413 return ReplaceInstUsesWith(CI, V);
415 if (isFreeCall(&CI, TLI))
416 return visitFree(CI);
418 // If the caller function is nounwind, mark the call as nounwind, even if the
420 if (CI.getParent()->getParent()->doesNotThrow() &&
421 !CI.doesNotThrow()) {
422 CI.setDoesNotThrow();
426 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
427 if (!II) return visitCallSite(&CI);
429 // Intrinsics cannot occur in an invoke, so handle them here instead of in
431 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
432 bool Changed = false;
434 // memmove/cpy/set of zero bytes is a noop.
435 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
436 if (NumBytes->isNullValue())
437 return EraseInstFromFunction(CI);
439 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
440 if (CI->getZExtValue() == 1) {
441 // Replace the instruction with just byte operations. We would
442 // transform other cases to loads/stores, but we don't know if
443 // alignment is sufficient.
447 // No other transformations apply to volatile transfers.
448 if (MI->isVolatile())
451 // If we have a memmove and the source operation is a constant global,
452 // then the source and dest pointers can't alias, so we can change this
453 // into a call to memcpy.
454 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
455 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
456 if (GVSrc->isConstant()) {
457 Module *M = CI.getParent()->getParent()->getParent();
458 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
459 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
460 CI.getArgOperand(1)->getType(),
461 CI.getArgOperand(2)->getType() };
462 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
467 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
468 // memmove(x,x,size) -> noop.
469 if (MTI->getSource() == MTI->getDest())
470 return EraseInstFromFunction(CI);
473 // If we can determine a pointer alignment that is bigger than currently
474 // set, update the alignment.
475 if (isa<MemTransferInst>(MI)) {
476 if (Instruction *I = SimplifyMemTransfer(MI))
478 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
479 if (Instruction *I = SimplifyMemSet(MSI))
483 if (Changed) return II;
486 switch (II->getIntrinsicID()) {
488 case Intrinsic::objectsize: {
490 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
491 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
494 case Intrinsic::bswap: {
495 Value *IIOperand = II->getArgOperand(0);
498 // bswap(bswap(x)) -> x
499 if (match(IIOperand, m_BSwap(m_Value(X))))
500 return ReplaceInstUsesWith(CI, X);
502 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
503 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
504 unsigned C = X->getType()->getPrimitiveSizeInBits() -
505 IIOperand->getType()->getPrimitiveSizeInBits();
506 Value *CV = ConstantInt::get(X->getType(), C);
507 Value *V = Builder->CreateLShr(X, CV);
508 return new TruncInst(V, IIOperand->getType());
513 case Intrinsic::powi:
514 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
517 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
520 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
521 // powi(x, -1) -> 1/x
522 if (Power->isAllOnesValue())
523 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
524 II->getArgOperand(0));
527 case Intrinsic::cttz: {
528 // If all bits below the first known one are known zero,
529 // this value is constant.
530 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
531 // FIXME: Try to simplify vectors of integers.
533 uint32_t BitWidth = IT->getBitWidth();
534 APInt KnownZero(BitWidth, 0);
535 APInt KnownOne(BitWidth, 0);
536 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
537 unsigned TrailingZeros = KnownOne.countTrailingZeros();
538 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
539 if ((Mask & KnownZero) == Mask)
540 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
541 APInt(BitWidth, TrailingZeros)));
545 case Intrinsic::ctlz: {
546 // If all bits above the first known one are known zero,
547 // this value is constant.
548 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
549 // FIXME: Try to simplify vectors of integers.
551 uint32_t BitWidth = IT->getBitWidth();
552 APInt KnownZero(BitWidth, 0);
553 APInt KnownOne(BitWidth, 0);
554 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
555 unsigned LeadingZeros = KnownOne.countLeadingZeros();
556 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
557 if ((Mask & KnownZero) == Mask)
558 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
559 APInt(BitWidth, LeadingZeros)));
564 case Intrinsic::uadd_with_overflow:
565 case Intrinsic::sadd_with_overflow:
566 case Intrinsic::umul_with_overflow:
567 case Intrinsic::smul_with_overflow:
568 if (isa<Constant>(II->getArgOperand(0)) &&
569 !isa<Constant>(II->getArgOperand(1))) {
570 // Canonicalize constants into the RHS.
571 Value *LHS = II->getArgOperand(0);
572 II->setArgOperand(0, II->getArgOperand(1));
573 II->setArgOperand(1, LHS);
578 case Intrinsic::usub_with_overflow:
579 case Intrinsic::ssub_with_overflow: {
580 OverflowCheckFlavor OCF =
581 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
582 assert(OCF != OCF_INVALID && "unexpected!");
584 Value *OperationResult = nullptr;
585 Constant *OverflowResult = nullptr;
586 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
587 *II, OperationResult, OverflowResult))
588 return CreateOverflowTuple(II, OperationResult, OverflowResult);
593 case Intrinsic::minnum:
594 case Intrinsic::maxnum: {
595 Value *Arg0 = II->getArgOperand(0);
596 Value *Arg1 = II->getArgOperand(1);
600 return ReplaceInstUsesWith(CI, Arg0);
602 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
603 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
605 // Canonicalize constants into the RHS.
607 II->setArgOperand(0, Arg1);
608 II->setArgOperand(1, Arg0);
613 if (C1 && C1->isNaN())
614 return ReplaceInstUsesWith(CI, Arg0);
616 // This is the value because if undef were NaN, we would return the other
617 // value and cannot return a NaN unless both operands are.
619 // fmin(undef, x) -> x
620 if (isa<UndefValue>(Arg0))
621 return ReplaceInstUsesWith(CI, Arg1);
623 // fmin(x, undef) -> x
624 if (isa<UndefValue>(Arg1))
625 return ReplaceInstUsesWith(CI, Arg0);
629 if (II->getIntrinsicID() == Intrinsic::minnum) {
630 // fmin(x, fmin(x, y)) -> fmin(x, y)
631 // fmin(y, fmin(x, y)) -> fmin(x, y)
632 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
633 if (Arg0 == X || Arg0 == Y)
634 return ReplaceInstUsesWith(CI, Arg1);
637 // fmin(fmin(x, y), x) -> fmin(x, y)
638 // fmin(fmin(x, y), y) -> fmin(x, y)
639 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
640 if (Arg1 == X || Arg1 == Y)
641 return ReplaceInstUsesWith(CI, Arg0);
644 // TODO: fmin(nnan x, inf) -> x
645 // TODO: fmin(nnan ninf x, flt_max) -> x
646 if (C1 && C1->isInfinity()) {
647 // fmin(x, -inf) -> -inf
648 if (C1->isNegative())
649 return ReplaceInstUsesWith(CI, Arg1);
652 assert(II->getIntrinsicID() == Intrinsic::maxnum);
653 // fmax(x, fmax(x, y)) -> fmax(x, y)
654 // fmax(y, fmax(x, y)) -> fmax(x, y)
655 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
656 if (Arg0 == X || Arg0 == Y)
657 return ReplaceInstUsesWith(CI, Arg1);
660 // fmax(fmax(x, y), x) -> fmax(x, y)
661 // fmax(fmax(x, y), y) -> fmax(x, y)
662 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
663 if (Arg1 == X || Arg1 == Y)
664 return ReplaceInstUsesWith(CI, Arg0);
667 // TODO: fmax(nnan x, -inf) -> x
668 // TODO: fmax(nnan ninf x, -flt_max) -> x
669 if (C1 && C1->isInfinity()) {
670 // fmax(x, inf) -> inf
671 if (!C1->isNegative())
672 return ReplaceInstUsesWith(CI, Arg1);
677 case Intrinsic::ppc_altivec_lvx:
678 case Intrinsic::ppc_altivec_lvxl:
679 // Turn PPC lvx -> load if the pointer is known aligned.
680 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
682 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
683 PointerType::getUnqual(II->getType()));
684 return new LoadInst(Ptr);
687 case Intrinsic::ppc_vsx_lxvw4x:
688 case Intrinsic::ppc_vsx_lxvd2x: {
689 // Turn PPC VSX loads into normal loads.
690 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
691 PointerType::getUnqual(II->getType()));
692 return new LoadInst(Ptr, Twine(""), false, 1);
694 case Intrinsic::ppc_altivec_stvx:
695 case Intrinsic::ppc_altivec_stvxl:
696 // Turn stvx -> store if the pointer is known aligned.
697 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
700 PointerType::getUnqual(II->getArgOperand(0)->getType());
701 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
702 return new StoreInst(II->getArgOperand(0), Ptr);
705 case Intrinsic::ppc_vsx_stxvw4x:
706 case Intrinsic::ppc_vsx_stxvd2x: {
707 // Turn PPC VSX stores into normal stores.
708 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
709 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
710 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
712 case Intrinsic::ppc_qpx_qvlfs:
713 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
714 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
716 Type *VTy = VectorType::get(Builder->getFloatTy(),
717 II->getType()->getVectorNumElements());
718 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
719 PointerType::getUnqual(VTy));
720 Value *Load = Builder->CreateLoad(Ptr);
721 return new FPExtInst(Load, II->getType());
724 case Intrinsic::ppc_qpx_qvlfd:
725 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
726 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
728 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
729 PointerType::getUnqual(II->getType()));
730 return new LoadInst(Ptr);
733 case Intrinsic::ppc_qpx_qvstfs:
734 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
735 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
737 Type *VTy = VectorType::get(Builder->getFloatTy(),
738 II->getArgOperand(0)->getType()->getVectorNumElements());
739 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
740 Type *OpPtrTy = PointerType::getUnqual(VTy);
741 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
742 return new StoreInst(TOp, Ptr);
745 case Intrinsic::ppc_qpx_qvstfd:
746 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
747 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
750 PointerType::getUnqual(II->getArgOperand(0)->getType());
751 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
752 return new StoreInst(II->getArgOperand(0), Ptr);
755 case Intrinsic::x86_sse_storeu_ps:
756 case Intrinsic::x86_sse2_storeu_pd:
757 case Intrinsic::x86_sse2_storeu_dq:
758 // Turn X86 storeu -> store if the pointer is known aligned.
759 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
762 PointerType::getUnqual(II->getArgOperand(1)->getType());
763 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
764 return new StoreInst(II->getArgOperand(1), Ptr);
768 case Intrinsic::x86_sse_cvtss2si:
769 case Intrinsic::x86_sse_cvtss2si64:
770 case Intrinsic::x86_sse_cvttss2si:
771 case Intrinsic::x86_sse_cvttss2si64:
772 case Intrinsic::x86_sse2_cvtsd2si:
773 case Intrinsic::x86_sse2_cvtsd2si64:
774 case Intrinsic::x86_sse2_cvttsd2si:
775 case Intrinsic::x86_sse2_cvttsd2si64: {
776 // These intrinsics only demand the 0th element of their input vectors. If
777 // we can simplify the input based on that, do so now.
779 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
780 APInt DemandedElts(VWidth, 1);
781 APInt UndefElts(VWidth, 0);
782 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
783 DemandedElts, UndefElts)) {
784 II->setArgOperand(0, V);
790 // Constant fold ashr( <A x Bi>, Ci ).
791 case Intrinsic::x86_sse2_psra_d:
792 case Intrinsic::x86_sse2_psra_w:
793 case Intrinsic::x86_sse2_psrai_d:
794 case Intrinsic::x86_sse2_psrai_w:
795 case Intrinsic::x86_avx2_psra_d:
796 case Intrinsic::x86_avx2_psra_w:
797 case Intrinsic::x86_avx2_psrai_d:
798 case Intrinsic::x86_avx2_psrai_w:
799 if (Value *V = SimplifyX86immshift(*II, *Builder, false, false))
800 return ReplaceInstUsesWith(*II, V);
803 // Constant fold lshr( <A x Bi>, Ci ).
804 case Intrinsic::x86_sse2_psrl_d:
805 case Intrinsic::x86_sse2_psrl_q:
806 case Intrinsic::x86_sse2_psrl_w:
807 case Intrinsic::x86_sse2_psrli_d:
808 case Intrinsic::x86_sse2_psrli_q:
809 case Intrinsic::x86_sse2_psrli_w:
810 case Intrinsic::x86_avx2_psrl_d:
811 case Intrinsic::x86_avx2_psrl_q:
812 case Intrinsic::x86_avx2_psrl_w:
813 case Intrinsic::x86_avx2_psrli_d:
814 case Intrinsic::x86_avx2_psrli_q:
815 case Intrinsic::x86_avx2_psrli_w:
816 if (Value *V = SimplifyX86immshift(*II, *Builder, true, false))
817 return ReplaceInstUsesWith(*II, V);
820 // Constant fold shl( <A x Bi>, Ci ).
821 case Intrinsic::x86_sse2_psll_d:
822 case Intrinsic::x86_sse2_psll_q:
823 case Intrinsic::x86_sse2_psll_w:
824 case Intrinsic::x86_sse2_pslli_d:
825 case Intrinsic::x86_sse2_pslli_q:
826 case Intrinsic::x86_sse2_pslli_w:
827 case Intrinsic::x86_avx2_psll_d:
828 case Intrinsic::x86_avx2_psll_q:
829 case Intrinsic::x86_avx2_psll_w:
830 case Intrinsic::x86_avx2_pslli_d:
831 case Intrinsic::x86_avx2_pslli_q:
832 case Intrinsic::x86_avx2_pslli_w:
833 if (Value *V = SimplifyX86immshift(*II, *Builder, true, true))
834 return ReplaceInstUsesWith(*II, V);
837 case Intrinsic::x86_sse41_pmovsxbd:
838 case Intrinsic::x86_sse41_pmovsxbq:
839 case Intrinsic::x86_sse41_pmovsxbw:
840 case Intrinsic::x86_sse41_pmovsxdq:
841 case Intrinsic::x86_sse41_pmovsxwd:
842 case Intrinsic::x86_sse41_pmovsxwq:
843 case Intrinsic::x86_avx2_pmovsxbd:
844 case Intrinsic::x86_avx2_pmovsxbq:
845 case Intrinsic::x86_avx2_pmovsxbw:
846 case Intrinsic::x86_avx2_pmovsxdq:
847 case Intrinsic::x86_avx2_pmovsxwd:
848 case Intrinsic::x86_avx2_pmovsxwq:
849 if (Value *V = SimplifyX86extend(*II, *Builder, true))
850 return ReplaceInstUsesWith(*II, V);
853 case Intrinsic::x86_sse41_pmovzxbd:
854 case Intrinsic::x86_sse41_pmovzxbq:
855 case Intrinsic::x86_sse41_pmovzxbw:
856 case Intrinsic::x86_sse41_pmovzxdq:
857 case Intrinsic::x86_sse41_pmovzxwd:
858 case Intrinsic::x86_sse41_pmovzxwq:
859 case Intrinsic::x86_avx2_pmovzxbd:
860 case Intrinsic::x86_avx2_pmovzxbq:
861 case Intrinsic::x86_avx2_pmovzxbw:
862 case Intrinsic::x86_avx2_pmovzxdq:
863 case Intrinsic::x86_avx2_pmovzxwd:
864 case Intrinsic::x86_avx2_pmovzxwq:
865 if (Value *V = SimplifyX86extend(*II, *Builder, false))
866 return ReplaceInstUsesWith(*II, V);
869 case Intrinsic::x86_sse41_insertps:
870 if (Value *V = SimplifyX86insertps(*II, *Builder))
871 return ReplaceInstUsesWith(*II, V);
874 case Intrinsic::x86_sse4a_insertqi: {
875 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
877 // TODO: eventually we should lower this intrinsic to IR
878 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
879 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
880 unsigned Index = CIIndex->getZExtValue();
881 // From AMD documentation: "a value of zero in the field length is
882 // defined as length of 64".
883 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
885 // From AMD documentation: "If the sum of the bit index + length field
886 // is greater than 64, the results are undefined".
887 unsigned End = Index + Length;
889 // Note that both field index and field length are 8-bit quantities.
890 // Since variables 'Index' and 'Length' are unsigned values
891 // obtained from zero-extending field index and field length
892 // respectively, their sum should never wrap around.
894 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
896 if (Length == 64 && Index == 0) {
897 Value *Vec = II->getArgOperand(1);
898 Value *Undef = UndefValue::get(Vec->getType());
899 const uint32_t Mask[] = { 0, 2 };
900 return ReplaceInstUsesWith(
902 Builder->CreateShuffleVector(
903 Vec, Undef, ConstantDataVector::get(
904 II->getContext(), makeArrayRef(Mask))));
905 } else if (auto Source =
906 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
907 if (Source->hasOneUse() &&
908 Source->getArgOperand(1) == II->getArgOperand(1)) {
909 // If the source of the insert has only one use and it's another
910 // insert (and they're both inserting from the same vector), try to
911 // bundle both together.
912 auto CISourceLength =
913 dyn_cast<ConstantInt>(Source->getArgOperand(2));
915 dyn_cast<ConstantInt>(Source->getArgOperand(3));
916 if (CISourceIndex && CISourceLength) {
917 unsigned SourceIndex = CISourceIndex->getZExtValue();
918 unsigned SourceLength = CISourceLength->getZExtValue();
919 unsigned SourceEnd = SourceIndex + SourceLength;
920 unsigned NewIndex, NewLength;
921 bool ShouldReplace = false;
922 if (Index <= SourceIndex && SourceIndex <= End) {
924 NewLength = std::max(End, SourceEnd) - NewIndex;
925 ShouldReplace = true;
926 } else if (SourceIndex <= Index && Index <= SourceEnd) {
927 NewIndex = SourceIndex;
928 NewLength = std::max(SourceEnd, End) - NewIndex;
929 ShouldReplace = true;
933 Constant *ConstantLength = ConstantInt::get(
934 II->getArgOperand(2)->getType(), NewLength, false);
935 Constant *ConstantIndex = ConstantInt::get(
936 II->getArgOperand(3)->getType(), NewIndex, false);
937 Value *Args[4] = { Source->getArgOperand(0),
938 II->getArgOperand(1), ConstantLength,
940 Module *M = CI.getParent()->getParent()->getParent();
942 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
943 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
953 case Intrinsic::x86_sse41_pblendvb:
954 case Intrinsic::x86_sse41_blendvps:
955 case Intrinsic::x86_sse41_blendvpd:
956 case Intrinsic::x86_avx_blendv_ps_256:
957 case Intrinsic::x86_avx_blendv_pd_256:
958 case Intrinsic::x86_avx2_pblendvb: {
959 // Convert blendv* to vector selects if the mask is constant.
960 // This optimization is convoluted because the intrinsic is defined as
961 // getting a vector of floats or doubles for the ps and pd versions.
962 // FIXME: That should be changed.
963 Value *Mask = II->getArgOperand(2);
964 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
965 auto Tyi1 = Builder->getInt1Ty();
966 auto SelectorType = cast<VectorType>(Mask->getType());
967 auto EltTy = SelectorType->getElementType();
968 unsigned Size = SelectorType->getNumElements();
972 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
973 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
974 "Wrong arguments for variable blend intrinsic");
975 SmallVector<Constant *, 32> Selectors;
976 for (unsigned I = 0; I < Size; ++I) {
977 // The intrinsics only read the top bit
980 Selector = C->getElementAsInteger(I);
982 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
983 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
985 auto NewSelector = ConstantVector::get(Selectors);
986 return SelectInst::Create(NewSelector, II->getArgOperand(1),
987 II->getArgOperand(0), "blendv");
993 case Intrinsic::x86_avx_vpermilvar_ps:
994 case Intrinsic::x86_avx_vpermilvar_ps_256:
995 case Intrinsic::x86_avx_vpermilvar_pd:
996 case Intrinsic::x86_avx_vpermilvar_pd_256: {
997 // Convert vpermil* to shufflevector if the mask is constant.
998 Value *V = II->getArgOperand(1);
999 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
1000 assert(Size == 8 || Size == 4 || Size == 2);
1001 uint32_t Indexes[8];
1002 if (auto C = dyn_cast<ConstantDataVector>(V)) {
1003 // The intrinsics only read one or two bits, clear the rest.
1004 for (unsigned I = 0; I < Size; ++I) {
1005 uint32_t Index = C->getElementAsInteger(I) & 0x3;
1006 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
1007 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
1011 } else if (isa<ConstantAggregateZero>(V)) {
1012 for (unsigned I = 0; I < Size; ++I)
1017 // The _256 variants are a bit trickier since the mask bits always index
1018 // into the corresponding 128 half. In order to convert to a generic
1019 // shuffle, we have to make that explicit.
1020 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
1021 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
1022 for (unsigned I = Size / 2; I < Size; ++I)
1023 Indexes[I] += Size / 2;
1026 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
1027 auto V1 = II->getArgOperand(0);
1028 auto V2 = UndefValue::get(V1->getType());
1029 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
1030 return ReplaceInstUsesWith(CI, Shuffle);
1033 case Intrinsic::x86_avx_vperm2f128_pd_256:
1034 case Intrinsic::x86_avx_vperm2f128_ps_256:
1035 case Intrinsic::x86_avx_vperm2f128_si_256:
1036 case Intrinsic::x86_avx2_vperm2i128:
1037 if (Value *V = SimplifyX86vperm2(*II, *Builder))
1038 return ReplaceInstUsesWith(*II, V);
1041 case Intrinsic::ppc_altivec_vperm:
1042 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1043 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1044 // a vectorshuffle for little endian, we must undo the transformation
1045 // performed on vec_perm in altivec.h. That is, we must complement
1046 // the permutation mask with respect to 31 and reverse the order of
1048 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1049 assert(Mask->getType()->getVectorNumElements() == 16 &&
1050 "Bad type for intrinsic!");
1052 // Check that all of the elements are integer constants or undefs.
1053 bool AllEltsOk = true;
1054 for (unsigned i = 0; i != 16; ++i) {
1055 Constant *Elt = Mask->getAggregateElement(i);
1056 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1063 // Cast the input vectors to byte vectors.
1064 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1066 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1068 Value *Result = UndefValue::get(Op0->getType());
1070 // Only extract each element once.
1071 Value *ExtractedElts[32];
1072 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1074 for (unsigned i = 0; i != 16; ++i) {
1075 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1078 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1079 Idx &= 31; // Match the hardware behavior.
1080 if (DL.isLittleEndian())
1083 if (!ExtractedElts[Idx]) {
1084 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1085 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1086 ExtractedElts[Idx] =
1087 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1088 Builder->getInt32(Idx&15));
1091 // Insert this value into the result vector.
1092 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1093 Builder->getInt32(i));
1095 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1100 case Intrinsic::arm_neon_vld1:
1101 case Intrinsic::arm_neon_vld2:
1102 case Intrinsic::arm_neon_vld3:
1103 case Intrinsic::arm_neon_vld4:
1104 case Intrinsic::arm_neon_vld2lane:
1105 case Intrinsic::arm_neon_vld3lane:
1106 case Intrinsic::arm_neon_vld4lane:
1107 case Intrinsic::arm_neon_vst1:
1108 case Intrinsic::arm_neon_vst2:
1109 case Intrinsic::arm_neon_vst3:
1110 case Intrinsic::arm_neon_vst4:
1111 case Intrinsic::arm_neon_vst2lane:
1112 case Intrinsic::arm_neon_vst3lane:
1113 case Intrinsic::arm_neon_vst4lane: {
1114 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1115 unsigned AlignArg = II->getNumArgOperands() - 1;
1116 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1117 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1118 II->setArgOperand(AlignArg,
1119 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1126 case Intrinsic::arm_neon_vmulls:
1127 case Intrinsic::arm_neon_vmullu:
1128 case Intrinsic::aarch64_neon_smull:
1129 case Intrinsic::aarch64_neon_umull: {
1130 Value *Arg0 = II->getArgOperand(0);
1131 Value *Arg1 = II->getArgOperand(1);
1133 // Handle mul by zero first:
1134 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1135 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1138 // Check for constant LHS & RHS - in this case we just simplify.
1139 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1140 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1141 VectorType *NewVT = cast<VectorType>(II->getType());
1142 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1143 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1144 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1145 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1147 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1150 // Couldn't simplify - canonicalize constant to the RHS.
1151 std::swap(Arg0, Arg1);
1154 // Handle mul by one:
1155 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1156 if (ConstantInt *Splat =
1157 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1159 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1160 /*isSigned=*/!Zext);
1165 case Intrinsic::AMDGPU_rcp: {
1166 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1167 const APFloat &ArgVal = C->getValueAPF();
1168 APFloat Val(ArgVal.getSemantics(), 1.0);
1169 APFloat::opStatus Status = Val.divide(ArgVal,
1170 APFloat::rmNearestTiesToEven);
1171 // Only do this if it was exact and therefore not dependent on the
1173 if (Status == APFloat::opOK)
1174 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1179 case Intrinsic::stackrestore: {
1180 // If the save is right next to the restore, remove the restore. This can
1181 // happen when variable allocas are DCE'd.
1182 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1183 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1184 BasicBlock::iterator BI = SS;
1186 return EraseInstFromFunction(CI);
1190 // Scan down this block to see if there is another stack restore in the
1191 // same block without an intervening call/alloca.
1192 BasicBlock::iterator BI = II;
1193 TerminatorInst *TI = II->getParent()->getTerminator();
1194 bool CannotRemove = false;
1195 for (++BI; &*BI != TI; ++BI) {
1196 if (isa<AllocaInst>(BI)) {
1197 CannotRemove = true;
1200 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1201 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1202 // If there is a stackrestore below this one, remove this one.
1203 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1204 return EraseInstFromFunction(CI);
1205 // Otherwise, ignore the intrinsic.
1207 // If we found a non-intrinsic call, we can't remove the stack
1209 CannotRemove = true;
1215 // If the stack restore is in a return, resume, or unwind block and if there
1216 // are no allocas or calls between the restore and the return, nuke the
1218 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1219 return EraseInstFromFunction(CI);
1222 case Intrinsic::assume: {
1223 // Canonicalize assume(a && b) -> assume(a); assume(b);
1224 // Note: New assumption intrinsics created here are registered by
1225 // the InstCombineIRInserter object.
1226 Value *IIOperand = II->getArgOperand(0), *A, *B,
1227 *AssumeIntrinsic = II->getCalledValue();
1228 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1229 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1230 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1231 return EraseInstFromFunction(*II);
1233 // assume(!(a || b)) -> assume(!a); assume(!b);
1234 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1235 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1237 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1239 return EraseInstFromFunction(*II);
1242 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1243 // (if assume is valid at the load)
1244 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1245 Value *LHS = ICmp->getOperand(0);
1246 Value *RHS = ICmp->getOperand(1);
1247 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1248 isa<LoadInst>(LHS) &&
1249 isa<Constant>(RHS) &&
1250 RHS->getType()->isPointerTy() &&
1251 cast<Constant>(RHS)->isNullValue()) {
1252 LoadInst* LI = cast<LoadInst>(LHS);
1253 if (isValidAssumeForContext(II, LI, DT)) {
1254 MDNode *MD = MDNode::get(II->getContext(), None);
1255 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1256 return EraseInstFromFunction(*II);
1259 // TODO: apply nonnull return attributes to calls and invokes
1260 // TODO: apply range metadata for range check patterns?
1262 // If there is a dominating assume with the same condition as this one,
1263 // then this one is redundant, and should be removed.
1264 APInt KnownZero(1, 0), KnownOne(1, 0);
1265 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1266 if (KnownOne.isAllOnesValue())
1267 return EraseInstFromFunction(*II);
1271 case Intrinsic::experimental_gc_relocate: {
1272 // Translate facts known about a pointer before relocating into
1273 // facts about the relocate value, while being careful to
1274 // preserve relocation semantics.
1275 GCRelocateOperands Operands(II);
1276 Value *DerivedPtr = Operands.getDerivedPtr();
1277 auto *GCRelocateType = cast<PointerType>(II->getType());
1279 // Remove the relocation if unused, note that this check is required
1280 // to prevent the cases below from looping forever.
1281 if (II->use_empty())
1282 return EraseInstFromFunction(*II);
1284 // Undef is undef, even after relocation.
1285 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1286 // most practical collectors, but there was discussion in the review thread
1287 // about whether it was legal for all possible collectors.
1288 if (isa<UndefValue>(DerivedPtr)) {
1289 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1290 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1293 // The relocation of null will be null for most any collector.
1294 // TODO: provide a hook for this in GCStrategy. There might be some weird
1295 // collector this property does not hold for.
1296 if (isa<ConstantPointerNull>(DerivedPtr)) {
1297 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1298 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1301 // isKnownNonNull -> nonnull attribute
1302 if (isKnownNonNull(DerivedPtr))
1303 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1305 // isDereferenceablePointer -> deref attribute
1306 if (isDereferenceablePointer(DerivedPtr, DL)) {
1307 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1308 uint64_t Bytes = A->getDereferenceableBytes();
1309 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1313 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1314 // Canonicalize on the type from the uses to the defs
1316 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1320 return visitCallSite(II);
1323 // InvokeInst simplification
1325 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1326 return visitCallSite(&II);
1329 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1330 /// passed through the varargs area, we can eliminate the use of the cast.
1331 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1332 const DataLayout &DL,
1333 const CastInst *const CI,
1335 if (!CI->isLosslessCast())
1338 // If this is a GC intrinsic, avoid munging types. We need types for
1339 // statepoint reconstruction in SelectionDAG.
1340 // TODO: This is probably something which should be expanded to all
1341 // intrinsics since the entire point of intrinsics is that
1342 // they are understandable by the optimizer.
1343 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1346 // The size of ByVal or InAlloca arguments is derived from the type, so we
1347 // can't change to a type with a different size. If the size were
1348 // passed explicitly we could avoid this check.
1349 if (!CS.isByValOrInAllocaArgument(ix))
1353 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1354 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1355 if (!SrcTy->isSized() || !DstTy->isSized())
1357 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1362 // Try to fold some different type of calls here.
1363 // Currently we're only working with the checking functions, memcpy_chk,
1364 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1365 // strcat_chk and strncat_chk.
1366 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1367 if (!CI->getCalledFunction()) return nullptr;
1369 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1370 ReplaceInstUsesWith(*From, With);
1372 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1373 if (Value *With = Simplifier.optimizeCall(CI)) {
1375 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1381 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1382 // Strip off at most one level of pointer casts, looking for an alloca. This
1383 // is good enough in practice and simpler than handling any number of casts.
1384 Value *Underlying = TrampMem->stripPointerCasts();
1385 if (Underlying != TrampMem &&
1386 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1388 if (!isa<AllocaInst>(Underlying))
1391 IntrinsicInst *InitTrampoline = nullptr;
1392 for (User *U : TrampMem->users()) {
1393 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1396 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1398 // More than one init_trampoline writes to this value. Give up.
1400 InitTrampoline = II;
1403 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1404 // Allow any number of calls to adjust.trampoline.
1409 // No call to init.trampoline found.
1410 if (!InitTrampoline)
1413 // Check that the alloca is being used in the expected way.
1414 if (InitTrampoline->getOperand(0) != TrampMem)
1417 return InitTrampoline;
1420 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1422 // Visit all the previous instructions in the basic block, and try to find a
1423 // init.trampoline which has a direct path to the adjust.trampoline.
1424 for (BasicBlock::iterator I = AdjustTramp,
1425 E = AdjustTramp->getParent()->begin(); I != E; ) {
1426 Instruction *Inst = --I;
1427 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1428 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1429 II->getOperand(0) == TrampMem)
1431 if (Inst->mayWriteToMemory())
1437 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1438 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1439 // to a direct call to a function. Otherwise return NULL.
1441 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1442 Callee = Callee->stripPointerCasts();
1443 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1445 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1448 Value *TrampMem = AdjustTramp->getOperand(0);
1450 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1452 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1457 // visitCallSite - Improvements for call and invoke instructions.
1459 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1461 if (isAllocLikeFn(CS.getInstruction(), TLI))
1462 return visitAllocSite(*CS.getInstruction());
1464 bool Changed = false;
1466 // Mark any parameters that are known to be non-null with the nonnull
1467 // attribute. This is helpful for inlining calls to functions with null
1468 // checks on their arguments.
1470 for (Value *V : CS.args()) {
1471 if (!CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1472 isKnownNonNull(V)) {
1473 AttributeSet AS = CS.getAttributes();
1474 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1475 Attribute::NonNull);
1476 CS.setAttributes(AS);
1481 assert(ArgNo == CS.arg_size() && "sanity check");
1483 // If the callee is a pointer to a function, attempt to move any casts to the
1484 // arguments of the call/invoke.
1485 Value *Callee = CS.getCalledValue();
1486 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1489 if (Function *CalleeF = dyn_cast<Function>(Callee))
1490 // If the call and callee calling conventions don't match, this call must
1491 // be unreachable, as the call is undefined.
1492 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1493 // Only do this for calls to a function with a body. A prototype may
1494 // not actually end up matching the implementation's calling conv for a
1495 // variety of reasons (e.g. it may be written in assembly).
1496 !CalleeF->isDeclaration()) {
1497 Instruction *OldCall = CS.getInstruction();
1498 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1499 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1501 // If OldCall does not return void then replaceAllUsesWith undef.
1502 // This allows ValueHandlers and custom metadata to adjust itself.
1503 if (!OldCall->getType()->isVoidTy())
1504 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1505 if (isa<CallInst>(OldCall))
1506 return EraseInstFromFunction(*OldCall);
1508 // We cannot remove an invoke, because it would change the CFG, just
1509 // change the callee to a null pointer.
1510 cast<InvokeInst>(OldCall)->setCalledFunction(
1511 Constant::getNullValue(CalleeF->getType()));
1515 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1516 // If CS does not return void then replaceAllUsesWith undef.
1517 // This allows ValueHandlers and custom metadata to adjust itself.
1518 if (!CS.getInstruction()->getType()->isVoidTy())
1519 ReplaceInstUsesWith(*CS.getInstruction(),
1520 UndefValue::get(CS.getInstruction()->getType()));
1522 if (isa<InvokeInst>(CS.getInstruction())) {
1523 // Can't remove an invoke because we cannot change the CFG.
1527 // This instruction is not reachable, just remove it. We insert a store to
1528 // undef so that we know that this code is not reachable, despite the fact
1529 // that we can't modify the CFG here.
1530 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1531 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1532 CS.getInstruction());
1534 return EraseInstFromFunction(*CS.getInstruction());
1537 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1538 return transformCallThroughTrampoline(CS, II);
1540 PointerType *PTy = cast<PointerType>(Callee->getType());
1541 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1542 if (FTy->isVarArg()) {
1543 int ix = FTy->getNumParams();
1544 // See if we can optimize any arguments passed through the varargs area of
1546 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1547 E = CS.arg_end(); I != E; ++I, ++ix) {
1548 CastInst *CI = dyn_cast<CastInst>(*I);
1549 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1550 *I = CI->getOperand(0);
1556 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1557 // Inline asm calls cannot throw - mark them 'nounwind'.
1558 CS.setDoesNotThrow();
1562 // Try to optimize the call if possible, we require DataLayout for most of
1563 // this. None of these calls are seen as possibly dead so go ahead and
1564 // delete the instruction now.
1565 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1566 Instruction *I = tryOptimizeCall(CI);
1567 // If we changed something return the result, etc. Otherwise let
1568 // the fallthrough check.
1569 if (I) return EraseInstFromFunction(*I);
1572 return Changed ? CS.getInstruction() : nullptr;
1575 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1576 // attempt to move the cast to the arguments of the call/invoke.
1578 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1580 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1583 // The prototype of thunks are a lie, don't try to directly call such
1585 if (Callee->hasFnAttribute("thunk"))
1587 Instruction *Caller = CS.getInstruction();
1588 const AttributeSet &CallerPAL = CS.getAttributes();
1590 // Okay, this is a cast from a function to a different type. Unless doing so
1591 // would cause a type conversion of one of our arguments, change this call to
1592 // be a direct call with arguments casted to the appropriate types.
1594 FunctionType *FT = Callee->getFunctionType();
1595 Type *OldRetTy = Caller->getType();
1596 Type *NewRetTy = FT->getReturnType();
1598 // Check to see if we are changing the return type...
1599 if (OldRetTy != NewRetTy) {
1601 if (NewRetTy->isStructTy())
1602 return false; // TODO: Handle multiple return values.
1604 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1605 if (Callee->isDeclaration())
1606 return false; // Cannot transform this return value.
1608 if (!Caller->use_empty() &&
1609 // void -> non-void is handled specially
1610 !NewRetTy->isVoidTy())
1611 return false; // Cannot transform this return value.
1614 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1615 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1616 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1617 return false; // Attribute not compatible with transformed value.
1620 // If the callsite is an invoke instruction, and the return value is used by
1621 // a PHI node in a successor, we cannot change the return type of the call
1622 // because there is no place to put the cast instruction (without breaking
1623 // the critical edge). Bail out in this case.
1624 if (!Caller->use_empty())
1625 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1626 for (User *U : II->users())
1627 if (PHINode *PN = dyn_cast<PHINode>(U))
1628 if (PN->getParent() == II->getNormalDest() ||
1629 PN->getParent() == II->getUnwindDest())
1633 unsigned NumActualArgs = CS.arg_size();
1634 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1636 // Prevent us turning:
1637 // declare void @takes_i32_inalloca(i32* inalloca)
1638 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1641 // call void @takes_i32_inalloca(i32* null)
1643 // Similarly, avoid folding away bitcasts of byval calls.
1644 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1645 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1648 CallSite::arg_iterator AI = CS.arg_begin();
1649 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1650 Type *ParamTy = FT->getParamType(i);
1651 Type *ActTy = (*AI)->getType();
1653 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1654 return false; // Cannot transform this parameter value.
1656 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1657 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1658 return false; // Attribute not compatible with transformed value.
1660 if (CS.isInAllocaArgument(i))
1661 return false; // Cannot transform to and from inalloca.
1663 // If the parameter is passed as a byval argument, then we have to have a
1664 // sized type and the sized type has to have the same size as the old type.
1665 if (ParamTy != ActTy &&
1666 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1667 Attribute::ByVal)) {
1668 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1669 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1672 Type *CurElTy = ActTy->getPointerElementType();
1673 if (DL.getTypeAllocSize(CurElTy) !=
1674 DL.getTypeAllocSize(ParamPTy->getElementType()))
1679 if (Callee->isDeclaration()) {
1680 // Do not delete arguments unless we have a function body.
1681 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1684 // If the callee is just a declaration, don't change the varargsness of the
1685 // call. We don't want to introduce a varargs call where one doesn't
1687 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1688 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1691 // If both the callee and the cast type are varargs, we still have to make
1692 // sure the number of fixed parameters are the same or we have the same
1693 // ABI issues as if we introduce a varargs call.
1694 if (FT->isVarArg() &&
1695 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1696 FT->getNumParams() !=
1697 cast<FunctionType>(APTy->getElementType())->getNumParams())
1701 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1702 !CallerPAL.isEmpty())
1703 // In this case we have more arguments than the new function type, but we
1704 // won't be dropping them. Check that these extra arguments have attributes
1705 // that are compatible with being a vararg call argument.
1706 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1707 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1708 if (Index <= FT->getNumParams())
1711 // Check if it has an attribute that's incompatible with varargs.
1712 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1713 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1718 // Okay, we decided that this is a safe thing to do: go ahead and start
1719 // inserting cast instructions as necessary.
1720 std::vector<Value*> Args;
1721 Args.reserve(NumActualArgs);
1722 SmallVector<AttributeSet, 8> attrVec;
1723 attrVec.reserve(NumCommonArgs);
1725 // Get any return attributes.
1726 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1728 // If the return value is not being used, the type may not be compatible
1729 // with the existing attributes. Wipe out any problematic attributes.
1730 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1732 // Add the new return attributes.
1733 if (RAttrs.hasAttributes())
1734 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1735 AttributeSet::ReturnIndex, RAttrs));
1737 AI = CS.arg_begin();
1738 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1739 Type *ParamTy = FT->getParamType(i);
1741 if ((*AI)->getType() == ParamTy) {
1742 Args.push_back(*AI);
1744 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1747 // Add any parameter attributes.
1748 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1749 if (PAttrs.hasAttributes())
1750 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1754 // If the function takes more arguments than the call was taking, add them
1756 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1757 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1759 // If we are removing arguments to the function, emit an obnoxious warning.
1760 if (FT->getNumParams() < NumActualArgs) {
1761 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1762 if (FT->isVarArg()) {
1763 // Add all of the arguments in their promoted form to the arg list.
1764 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1765 Type *PTy = getPromotedType((*AI)->getType());
1766 if (PTy != (*AI)->getType()) {
1767 // Must promote to pass through va_arg area!
1768 Instruction::CastOps opcode =
1769 CastInst::getCastOpcode(*AI, false, PTy, false);
1770 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1772 Args.push_back(*AI);
1775 // Add any parameter attributes.
1776 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1777 if (PAttrs.hasAttributes())
1778 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1784 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1785 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1786 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1788 if (NewRetTy->isVoidTy())
1789 Caller->setName(""); // Void type should not have a name.
1791 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1795 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1796 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1797 II->getUnwindDest(), Args);
1799 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1800 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1802 CallInst *CI = cast<CallInst>(Caller);
1803 NC = Builder->CreateCall(Callee, Args);
1805 if (CI->isTailCall())
1806 cast<CallInst>(NC)->setTailCall();
1807 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1808 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1811 // Insert a cast of the return type as necessary.
1813 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1814 if (!NV->getType()->isVoidTy()) {
1815 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1816 NC->setDebugLoc(Caller->getDebugLoc());
1818 // If this is an invoke instruction, we should insert it after the first
1819 // non-phi, instruction in the normal successor block.
1820 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1821 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1822 InsertNewInstBefore(NC, *I);
1824 // Otherwise, it's a call, just insert cast right after the call.
1825 InsertNewInstBefore(NC, *Caller);
1827 Worklist.AddUsersToWorkList(*Caller);
1829 NV = UndefValue::get(Caller->getType());
1833 if (!Caller->use_empty())
1834 ReplaceInstUsesWith(*Caller, NV);
1835 else if (Caller->hasValueHandle()) {
1836 if (OldRetTy == NV->getType())
1837 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1839 // We cannot call ValueIsRAUWd with a different type, and the
1840 // actual tracked value will disappear.
1841 ValueHandleBase::ValueIsDeleted(Caller);
1844 EraseInstFromFunction(*Caller);
1848 // transformCallThroughTrampoline - Turn a call to a function created by
1849 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1850 // underlying function.
1853 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1854 IntrinsicInst *Tramp) {
1855 Value *Callee = CS.getCalledValue();
1856 PointerType *PTy = cast<PointerType>(Callee->getType());
1857 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1858 const AttributeSet &Attrs = CS.getAttributes();
1860 // If the call already has the 'nest' attribute somewhere then give up -
1861 // otherwise 'nest' would occur twice after splicing in the chain.
1862 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1866 "transformCallThroughTrampoline called with incorrect CallSite.");
1868 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1869 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1870 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1872 const AttributeSet &NestAttrs = NestF->getAttributes();
1873 if (!NestAttrs.isEmpty()) {
1874 unsigned NestIdx = 1;
1875 Type *NestTy = nullptr;
1876 AttributeSet NestAttr;
1878 // Look for a parameter marked with the 'nest' attribute.
1879 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1880 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1881 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1882 // Record the parameter type and any other attributes.
1884 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1889 Instruction *Caller = CS.getInstruction();
1890 std::vector<Value*> NewArgs;
1891 NewArgs.reserve(CS.arg_size() + 1);
1893 SmallVector<AttributeSet, 8> NewAttrs;
1894 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1896 // Insert the nest argument into the call argument list, which may
1897 // mean appending it. Likewise for attributes.
1899 // Add any result attributes.
1900 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1901 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1902 Attrs.getRetAttributes()));
1906 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1908 if (Idx == NestIdx) {
1909 // Add the chain argument and attributes.
1910 Value *NestVal = Tramp->getArgOperand(2);
1911 if (NestVal->getType() != NestTy)
1912 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1913 NewArgs.push_back(NestVal);
1914 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1921 // Add the original argument and attributes.
1922 NewArgs.push_back(*I);
1923 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1924 if (Attr.hasAttributes(Idx)) {
1925 AttrBuilder B(Attr, Idx);
1926 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1927 Idx + (Idx >= NestIdx), B));
1934 // Add any function attributes.
1935 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1936 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1937 Attrs.getFnAttributes()));
1939 // The trampoline may have been bitcast to a bogus type (FTy).
1940 // Handle this by synthesizing a new function type, equal to FTy
1941 // with the chain parameter inserted.
1943 std::vector<Type*> NewTypes;
1944 NewTypes.reserve(FTy->getNumParams()+1);
1946 // Insert the chain's type into the list of parameter types, which may
1947 // mean appending it.
1950 FunctionType::param_iterator I = FTy->param_begin(),
1951 E = FTy->param_end();
1955 // Add the chain's type.
1956 NewTypes.push_back(NestTy);
1961 // Add the original type.
1962 NewTypes.push_back(*I);
1968 // Replace the trampoline call with a direct call. Let the generic
1969 // code sort out any function type mismatches.
1970 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1972 Constant *NewCallee =
1973 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1974 NestF : ConstantExpr::getBitCast(NestF,
1975 PointerType::getUnqual(NewFTy));
1976 const AttributeSet &NewPAL =
1977 AttributeSet::get(FTy->getContext(), NewAttrs);
1979 Instruction *NewCaller;
1980 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1981 NewCaller = InvokeInst::Create(NewCallee,
1982 II->getNormalDest(), II->getUnwindDest(),
1984 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1985 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1987 NewCaller = CallInst::Create(NewCallee, NewArgs);
1988 if (cast<CallInst>(Caller)->isTailCall())
1989 cast<CallInst>(NewCaller)->setTailCall();
1990 cast<CallInst>(NewCaller)->
1991 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1992 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1999 // Replace the trampoline call with a direct call. Since there is no 'nest'
2000 // parameter, there is no need to adjust the argument list. Let the generic
2001 // code sort out any function type mismatches.
2002 Constant *NewCallee =
2003 NestF->getType() == PTy ? NestF :
2004 ConstantExpr::getBitCast(NestF, PTy);
2005 CS.setCalledFunction(NewCallee);
2006 return CS.getInstruction();