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,
\r
201 InstCombiner::BuilderTy &Builder,
\r
203 // Simplify if count is constant. To 0 if >= BitWidth,
\r
204 // otherwise to shl/lshr.
\r
205 auto CDV = dyn_cast<ConstantDataVector>(II.getArgOperand(1));
\r
206 auto CInt = dyn_cast<ConstantInt>(II.getArgOperand(1));
\r
209 ConstantInt *Count;
\r
211 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
\r
215 auto Vec = II.getArgOperand(0);
\r
216 auto VT = cast<VectorType>(Vec->getType());
\r
217 auto SVT = VT->getElementType();
\r
218 if (Count->getZExtValue() > (SVT->getPrimitiveSizeInBits() - 1))
\r
219 return ConstantAggregateZero::get(VT);
\r
221 unsigned VWidth = VT->getNumElements();
\r
223 // Get a constant vector of the same type as the first operand.
\r
224 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
\r
227 return Builder.CreateShl(Vec, Builder.CreateVectorSplat(VWidth, VTCI));
\r
229 return Builder.CreateLShr(Vec, Builder.CreateVectorSplat(VWidth, VTCI));
\r
232 static Value *SimplifyX86extend(const IntrinsicInst &II,
\r
233 InstCombiner::BuilderTy &Builder,
\r
235 VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType());
236 VectorType *DstTy = cast<VectorType>(II.getType());
237 unsigned NumDstElts = DstTy->getNumElements();
239 // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
240 SmallVector<int, 8> ShuffleMask;
241 for (int i = 0; i != (int)NumDstElts; ++i)
242 ShuffleMask.push_back(i);
244 Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0),
245 UndefValue::get(SrcTy), ShuffleMask);
246 return SignExtend ? Builder.CreateSExt(SV, DstTy)
247 : Builder.CreateZExt(SV, DstTy);
250 static Value *SimplifyX86insertps(const IntrinsicInst &II,
251 InstCombiner::BuilderTy &Builder) {
252 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
253 VectorType *VecTy = cast<VectorType>(II.getType());
254 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
256 // The immediate permute control byte looks like this:
257 // [3:0] - zero mask for each 32-bit lane
258 // [5:4] - select one 32-bit destination lane
259 // [7:6] - select one 32-bit source lane
261 uint8_t Imm = CInt->getZExtValue();
262 uint8_t ZMask = Imm & 0xf;
263 uint8_t DestLane = (Imm >> 4) & 0x3;
264 uint8_t SourceLane = (Imm >> 6) & 0x3;
266 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
268 // If all zero mask bits are set, this was just a weird way to
269 // generate a zero vector.
273 // Initialize by passing all of the first source bits through.
274 int ShuffleMask[4] = { 0, 1, 2, 3 };
276 // We may replace the second operand with the zero vector.
277 Value *V1 = II.getArgOperand(1);
280 // If the zero mask is being used with a single input or the zero mask
281 // overrides the destination lane, this is a shuffle with the zero vector.
282 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
283 (ZMask & (1 << DestLane))) {
285 // We may still move 32-bits of the first source vector from one lane
287 ShuffleMask[DestLane] = SourceLane;
288 // The zero mask may override the previous insert operation.
289 for (unsigned i = 0; i < 4; ++i)
290 if ((ZMask >> i) & 0x1)
291 ShuffleMask[i] = i + 4;
293 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
297 // Replace the selected destination lane with the selected source lane.
298 ShuffleMask[DestLane] = SourceLane + 4;
301 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
306 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
307 /// source vectors, unless a zero bit is set. If a zero bit is set,
308 /// then ignore that half of the mask and clear that half of the vector.
309 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
310 InstCombiner::BuilderTy &Builder) {
311 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
312 VectorType *VecTy = cast<VectorType>(II.getType());
313 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
315 // The immediate permute control byte looks like this:
316 // [1:0] - select 128 bits from sources for low half of destination
318 // [3] - zero low half of destination
319 // [5:4] - select 128 bits from sources for high half of destination
321 // [7] - zero high half of destination
323 uint8_t Imm = CInt->getZExtValue();
325 bool LowHalfZero = Imm & 0x08;
326 bool HighHalfZero = Imm & 0x80;
328 // If both zero mask bits are set, this was just a weird way to
329 // generate a zero vector.
330 if (LowHalfZero && HighHalfZero)
333 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
334 unsigned NumElts = VecTy->getNumElements();
335 unsigned HalfSize = NumElts / 2;
336 SmallVector<int, 8> ShuffleMask(NumElts);
338 // The high bit of the selection field chooses the 1st or 2nd operand.
339 bool LowInputSelect = Imm & 0x02;
340 bool HighInputSelect = Imm & 0x20;
342 // The low bit of the selection field chooses the low or high half
343 // of the selected operand.
344 bool LowHalfSelect = Imm & 0x01;
345 bool HighHalfSelect = Imm & 0x10;
347 // Determine which operand(s) are actually in use for this instruction.
348 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
349 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
351 // If needed, replace operands based on zero mask.
352 V0 = LowHalfZero ? ZeroVector : V0;
353 V1 = HighHalfZero ? ZeroVector : V1;
355 // Permute low half of result.
356 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
357 for (unsigned i = 0; i < HalfSize; ++i)
358 ShuffleMask[i] = StartIndex + i;
360 // Permute high half of result.
361 StartIndex = HighHalfSelect ? HalfSize : 0;
362 StartIndex += NumElts;
363 for (unsigned i = 0; i < HalfSize; ++i)
364 ShuffleMask[i + HalfSize] = StartIndex + i;
366 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
371 /// visitCallInst - CallInst simplification. This mostly only handles folding
372 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
373 /// the heavy lifting.
375 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
376 auto Args = CI.arg_operands();
377 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
379 return ReplaceInstUsesWith(CI, V);
381 if (isFreeCall(&CI, TLI))
382 return visitFree(CI);
384 // If the caller function is nounwind, mark the call as nounwind, even if the
386 if (CI.getParent()->getParent()->doesNotThrow() &&
387 !CI.doesNotThrow()) {
388 CI.setDoesNotThrow();
392 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
393 if (!II) return visitCallSite(&CI);
395 // Intrinsics cannot occur in an invoke, so handle them here instead of in
397 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
398 bool Changed = false;
400 // memmove/cpy/set of zero bytes is a noop.
401 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
402 if (NumBytes->isNullValue())
403 return EraseInstFromFunction(CI);
405 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
406 if (CI->getZExtValue() == 1) {
407 // Replace the instruction with just byte operations. We would
408 // transform other cases to loads/stores, but we don't know if
409 // alignment is sufficient.
413 // No other transformations apply to volatile transfers.
414 if (MI->isVolatile())
417 // If we have a memmove and the source operation is a constant global,
418 // then the source and dest pointers can't alias, so we can change this
419 // into a call to memcpy.
420 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
421 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
422 if (GVSrc->isConstant()) {
423 Module *M = CI.getParent()->getParent()->getParent();
424 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
425 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
426 CI.getArgOperand(1)->getType(),
427 CI.getArgOperand(2)->getType() };
428 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
433 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
434 // memmove(x,x,size) -> noop.
435 if (MTI->getSource() == MTI->getDest())
436 return EraseInstFromFunction(CI);
439 // If we can determine a pointer alignment that is bigger than currently
440 // set, update the alignment.
441 if (isa<MemTransferInst>(MI)) {
442 if (Instruction *I = SimplifyMemTransfer(MI))
444 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
445 if (Instruction *I = SimplifyMemSet(MSI))
449 if (Changed) return II;
452 switch (II->getIntrinsicID()) {
454 case Intrinsic::objectsize: {
456 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
457 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
460 case Intrinsic::bswap: {
461 Value *IIOperand = II->getArgOperand(0);
464 // bswap(bswap(x)) -> x
465 if (match(IIOperand, m_BSwap(m_Value(X))))
466 return ReplaceInstUsesWith(CI, X);
468 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
469 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
470 unsigned C = X->getType()->getPrimitiveSizeInBits() -
471 IIOperand->getType()->getPrimitiveSizeInBits();
472 Value *CV = ConstantInt::get(X->getType(), C);
473 Value *V = Builder->CreateLShr(X, CV);
474 return new TruncInst(V, IIOperand->getType());
479 case Intrinsic::powi:
480 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
483 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
486 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
487 // powi(x, -1) -> 1/x
488 if (Power->isAllOnesValue())
489 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
490 II->getArgOperand(0));
493 case Intrinsic::cttz: {
494 // If all bits below the first known one are known zero,
495 // this value is constant.
496 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
497 // FIXME: Try to simplify vectors of integers.
499 uint32_t BitWidth = IT->getBitWidth();
500 APInt KnownZero(BitWidth, 0);
501 APInt KnownOne(BitWidth, 0);
502 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
503 unsigned TrailingZeros = KnownOne.countTrailingZeros();
504 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
505 if ((Mask & KnownZero) == Mask)
506 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
507 APInt(BitWidth, TrailingZeros)));
511 case Intrinsic::ctlz: {
512 // If all bits above the first known one are known zero,
513 // this value is constant.
514 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
515 // FIXME: Try to simplify vectors of integers.
517 uint32_t BitWidth = IT->getBitWidth();
518 APInt KnownZero(BitWidth, 0);
519 APInt KnownOne(BitWidth, 0);
520 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
521 unsigned LeadingZeros = KnownOne.countLeadingZeros();
522 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
523 if ((Mask & KnownZero) == Mask)
524 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
525 APInt(BitWidth, LeadingZeros)));
530 case Intrinsic::uadd_with_overflow:
531 case Intrinsic::sadd_with_overflow:
532 case Intrinsic::umul_with_overflow:
533 case Intrinsic::smul_with_overflow:
534 if (isa<Constant>(II->getArgOperand(0)) &&
535 !isa<Constant>(II->getArgOperand(1))) {
536 // Canonicalize constants into the RHS.
537 Value *LHS = II->getArgOperand(0);
538 II->setArgOperand(0, II->getArgOperand(1));
539 II->setArgOperand(1, LHS);
544 case Intrinsic::usub_with_overflow:
545 case Intrinsic::ssub_with_overflow: {
546 OverflowCheckFlavor OCF =
547 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
548 assert(OCF != OCF_INVALID && "unexpected!");
550 Value *OperationResult = nullptr;
551 Constant *OverflowResult = nullptr;
552 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
553 *II, OperationResult, OverflowResult))
554 return CreateOverflowTuple(II, OperationResult, OverflowResult);
559 case Intrinsic::minnum:
560 case Intrinsic::maxnum: {
561 Value *Arg0 = II->getArgOperand(0);
562 Value *Arg1 = II->getArgOperand(1);
566 return ReplaceInstUsesWith(CI, Arg0);
568 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
569 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
571 // Canonicalize constants into the RHS.
573 II->setArgOperand(0, Arg1);
574 II->setArgOperand(1, Arg0);
579 if (C1 && C1->isNaN())
580 return ReplaceInstUsesWith(CI, Arg0);
582 // This is the value because if undef were NaN, we would return the other
583 // value and cannot return a NaN unless both operands are.
585 // fmin(undef, x) -> x
586 if (isa<UndefValue>(Arg0))
587 return ReplaceInstUsesWith(CI, Arg1);
589 // fmin(x, undef) -> x
590 if (isa<UndefValue>(Arg1))
591 return ReplaceInstUsesWith(CI, Arg0);
595 if (II->getIntrinsicID() == Intrinsic::minnum) {
596 // fmin(x, fmin(x, y)) -> fmin(x, y)
597 // fmin(y, fmin(x, y)) -> fmin(x, y)
598 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
599 if (Arg0 == X || Arg0 == Y)
600 return ReplaceInstUsesWith(CI, Arg1);
603 // fmin(fmin(x, y), x) -> fmin(x, y)
604 // fmin(fmin(x, y), y) -> fmin(x, y)
605 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
606 if (Arg1 == X || Arg1 == Y)
607 return ReplaceInstUsesWith(CI, Arg0);
610 // TODO: fmin(nnan x, inf) -> x
611 // TODO: fmin(nnan ninf x, flt_max) -> x
612 if (C1 && C1->isInfinity()) {
613 // fmin(x, -inf) -> -inf
614 if (C1->isNegative())
615 return ReplaceInstUsesWith(CI, Arg1);
618 assert(II->getIntrinsicID() == Intrinsic::maxnum);
619 // fmax(x, fmax(x, y)) -> fmax(x, y)
620 // fmax(y, fmax(x, y)) -> fmax(x, y)
621 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
622 if (Arg0 == X || Arg0 == Y)
623 return ReplaceInstUsesWith(CI, Arg1);
626 // fmax(fmax(x, y), x) -> fmax(x, y)
627 // fmax(fmax(x, y), y) -> fmax(x, y)
628 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
629 if (Arg1 == X || Arg1 == Y)
630 return ReplaceInstUsesWith(CI, Arg0);
633 // TODO: fmax(nnan x, -inf) -> x
634 // TODO: fmax(nnan ninf x, -flt_max) -> x
635 if (C1 && C1->isInfinity()) {
636 // fmax(x, inf) -> inf
637 if (!C1->isNegative())
638 return ReplaceInstUsesWith(CI, Arg1);
643 case Intrinsic::ppc_altivec_lvx:
644 case Intrinsic::ppc_altivec_lvxl:
645 // Turn PPC lvx -> load if the pointer is known aligned.
646 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
648 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
649 PointerType::getUnqual(II->getType()));
650 return new LoadInst(Ptr);
653 case Intrinsic::ppc_vsx_lxvw4x:
654 case Intrinsic::ppc_vsx_lxvd2x: {
655 // Turn PPC VSX loads into normal loads.
656 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
657 PointerType::getUnqual(II->getType()));
658 return new LoadInst(Ptr, Twine(""), false, 1);
660 case Intrinsic::ppc_altivec_stvx:
661 case Intrinsic::ppc_altivec_stvxl:
662 // Turn stvx -> store if the pointer is known aligned.
663 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
666 PointerType::getUnqual(II->getArgOperand(0)->getType());
667 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
668 return new StoreInst(II->getArgOperand(0), Ptr);
671 case Intrinsic::ppc_vsx_stxvw4x:
672 case Intrinsic::ppc_vsx_stxvd2x: {
673 // Turn PPC VSX stores into normal stores.
674 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
675 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
676 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
678 case Intrinsic::ppc_qpx_qvlfs:
679 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
680 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
682 Type *VTy = VectorType::get(Builder->getFloatTy(),
683 II->getType()->getVectorNumElements());
684 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
685 PointerType::getUnqual(VTy));
686 Value *Load = Builder->CreateLoad(Ptr);
687 return new FPExtInst(Load, II->getType());
690 case Intrinsic::ppc_qpx_qvlfd:
691 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
692 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
694 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
695 PointerType::getUnqual(II->getType()));
696 return new LoadInst(Ptr);
699 case Intrinsic::ppc_qpx_qvstfs:
700 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
701 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
703 Type *VTy = VectorType::get(Builder->getFloatTy(),
704 II->getArgOperand(0)->getType()->getVectorNumElements());
705 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
706 Type *OpPtrTy = PointerType::getUnqual(VTy);
707 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
708 return new StoreInst(TOp, Ptr);
711 case Intrinsic::ppc_qpx_qvstfd:
712 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
713 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
716 PointerType::getUnqual(II->getArgOperand(0)->getType());
717 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
718 return new StoreInst(II->getArgOperand(0), Ptr);
721 case Intrinsic::x86_sse_storeu_ps:
722 case Intrinsic::x86_sse2_storeu_pd:
723 case Intrinsic::x86_sse2_storeu_dq:
724 // Turn X86 storeu -> store if the pointer is known aligned.
725 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
728 PointerType::getUnqual(II->getArgOperand(1)->getType());
729 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
730 return new StoreInst(II->getArgOperand(1), Ptr);
734 case Intrinsic::x86_sse_cvtss2si:
735 case Intrinsic::x86_sse_cvtss2si64:
736 case Intrinsic::x86_sse_cvttss2si:
737 case Intrinsic::x86_sse_cvttss2si64:
738 case Intrinsic::x86_sse2_cvtsd2si:
739 case Intrinsic::x86_sse2_cvtsd2si64:
740 case Intrinsic::x86_sse2_cvttsd2si:
741 case Intrinsic::x86_sse2_cvttsd2si64: {
742 // These intrinsics only demand the 0th element of their input vectors. If
743 // we can simplify the input based on that, do so now.
745 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
746 APInt DemandedElts(VWidth, 1);
747 APInt UndefElts(VWidth, 0);
748 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
749 DemandedElts, UndefElts)) {
750 II->setArgOperand(0, V);
756 // Constant fold lshr( <A x Bi>, Ci ).
\r
757 case Intrinsic::x86_sse2_psrl_d:
\r
758 case Intrinsic::x86_sse2_psrl_q:
\r
759 case Intrinsic::x86_sse2_psrl_w:
\r
760 case Intrinsic::x86_sse2_psrli_d:
\r
761 case Intrinsic::x86_sse2_psrli_q:
\r
762 case Intrinsic::x86_sse2_psrli_w:
\r
763 case Intrinsic::x86_avx2_psrl_d:
\r
764 case Intrinsic::x86_avx2_psrl_q:
\r
765 case Intrinsic::x86_avx2_psrl_w:
\r
766 case Intrinsic::x86_avx2_psrli_d:
\r
767 case Intrinsic::x86_avx2_psrli_q:
\r
768 case Intrinsic::x86_avx2_psrli_w:
\r
769 if (Value *V = SimplifyX86immshift(*II, *Builder, false))
\r
770 return ReplaceInstUsesWith(*II, V);
\r
773 // Constant fold shl( <A x Bi>, Ci ).
\r
774 case Intrinsic::x86_sse2_psll_d:
\r
775 case Intrinsic::x86_sse2_psll_q:
\r
776 case Intrinsic::x86_sse2_psll_w:
\r
777 case Intrinsic::x86_sse2_pslli_d:
778 case Intrinsic::x86_sse2_pslli_q:
779 case Intrinsic::x86_sse2_pslli_w:
780 case Intrinsic::x86_avx2_psll_d:
781 case Intrinsic::x86_avx2_psll_q:
782 case Intrinsic::x86_avx2_psll_w:
783 case Intrinsic::x86_avx2_pslli_d:
\r
784 case Intrinsic::x86_avx2_pslli_q:
\r
785 case Intrinsic::x86_avx2_pslli_w:
\r
786 if (Value *V = SimplifyX86immshift(*II, *Builder, true))
\r
787 return ReplaceInstUsesWith(*II, V);
\r
790 case Intrinsic::x86_sse41_pmovsxbd:
\r
791 case Intrinsic::x86_sse41_pmovsxbq:
\r
792 case Intrinsic::x86_sse41_pmovsxbw:
\r
793 case Intrinsic::x86_sse41_pmovsxdq:
794 case Intrinsic::x86_sse41_pmovsxwd:
795 case Intrinsic::x86_sse41_pmovsxwq:
796 case Intrinsic::x86_avx2_pmovsxbd:
797 case Intrinsic::x86_avx2_pmovsxbq:
798 case Intrinsic::x86_avx2_pmovsxbw:
799 case Intrinsic::x86_avx2_pmovsxdq:
800 case Intrinsic::x86_avx2_pmovsxwd:
801 case Intrinsic::x86_avx2_pmovsxwq:
802 if (Value *V = SimplifyX86extend(*II, *Builder, true))
803 return ReplaceInstUsesWith(*II, V);
806 case Intrinsic::x86_sse41_pmovzxbd:
807 case Intrinsic::x86_sse41_pmovzxbq:
808 case Intrinsic::x86_sse41_pmovzxbw:
809 case Intrinsic::x86_sse41_pmovzxdq:
810 case Intrinsic::x86_sse41_pmovzxwd:
811 case Intrinsic::x86_sse41_pmovzxwq:
812 case Intrinsic::x86_avx2_pmovzxbd:
813 case Intrinsic::x86_avx2_pmovzxbq:
814 case Intrinsic::x86_avx2_pmovzxbw:
815 case Intrinsic::x86_avx2_pmovzxdq:
816 case Intrinsic::x86_avx2_pmovzxwd:
817 case Intrinsic::x86_avx2_pmovzxwq:
818 if (Value *V = SimplifyX86extend(*II, *Builder, false))
819 return ReplaceInstUsesWith(*II, V);
822 case Intrinsic::x86_sse41_insertps:
823 if (Value *V = SimplifyX86insertps(*II, *Builder))
824 return ReplaceInstUsesWith(*II, V);
827 case Intrinsic::x86_sse4a_insertqi: {
828 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
830 // TODO: eventually we should lower this intrinsic to IR
831 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
832 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
833 unsigned Index = CIIndex->getZExtValue();
834 // From AMD documentation: "a value of zero in the field length is
835 // defined as length of 64".
836 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
838 // From AMD documentation: "If the sum of the bit index + length field
839 // is greater than 64, the results are undefined".
840 unsigned End = Index + Length;
842 // Note that both field index and field length are 8-bit quantities.
843 // Since variables 'Index' and 'Length' are unsigned values
844 // obtained from zero-extending field index and field length
845 // respectively, their sum should never wrap around.
847 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
849 if (Length == 64 && Index == 0) {
850 Value *Vec = II->getArgOperand(1);
851 Value *Undef = UndefValue::get(Vec->getType());
852 const uint32_t Mask[] = { 0, 2 };
853 return ReplaceInstUsesWith(
855 Builder->CreateShuffleVector(
856 Vec, Undef, ConstantDataVector::get(
857 II->getContext(), makeArrayRef(Mask))));
858 } else if (auto Source =
859 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
860 if (Source->hasOneUse() &&
861 Source->getArgOperand(1) == II->getArgOperand(1)) {
862 // If the source of the insert has only one use and it's another
863 // insert (and they're both inserting from the same vector), try to
864 // bundle both together.
865 auto CISourceLength =
866 dyn_cast<ConstantInt>(Source->getArgOperand(2));
868 dyn_cast<ConstantInt>(Source->getArgOperand(3));
869 if (CISourceIndex && CISourceLength) {
870 unsigned SourceIndex = CISourceIndex->getZExtValue();
871 unsigned SourceLength = CISourceLength->getZExtValue();
872 unsigned SourceEnd = SourceIndex + SourceLength;
873 unsigned NewIndex, NewLength;
874 bool ShouldReplace = false;
875 if (Index <= SourceIndex && SourceIndex <= End) {
877 NewLength = std::max(End, SourceEnd) - NewIndex;
878 ShouldReplace = true;
879 } else if (SourceIndex <= Index && Index <= SourceEnd) {
880 NewIndex = SourceIndex;
881 NewLength = std::max(SourceEnd, End) - NewIndex;
882 ShouldReplace = true;
886 Constant *ConstantLength = ConstantInt::get(
887 II->getArgOperand(2)->getType(), NewLength, false);
888 Constant *ConstantIndex = ConstantInt::get(
889 II->getArgOperand(3)->getType(), NewIndex, false);
890 Value *Args[4] = { Source->getArgOperand(0),
891 II->getArgOperand(1), ConstantLength,
893 Module *M = CI.getParent()->getParent()->getParent();
895 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
896 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
906 case Intrinsic::x86_sse41_pblendvb:
907 case Intrinsic::x86_sse41_blendvps:
908 case Intrinsic::x86_sse41_blendvpd:
909 case Intrinsic::x86_avx_blendv_ps_256:
910 case Intrinsic::x86_avx_blendv_pd_256:
911 case Intrinsic::x86_avx2_pblendvb: {
912 // Convert blendv* to vector selects if the mask is constant.
913 // This optimization is convoluted because the intrinsic is defined as
914 // getting a vector of floats or doubles for the ps and pd versions.
915 // FIXME: That should be changed.
916 Value *Mask = II->getArgOperand(2);
917 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
918 auto Tyi1 = Builder->getInt1Ty();
919 auto SelectorType = cast<VectorType>(Mask->getType());
920 auto EltTy = SelectorType->getElementType();
921 unsigned Size = SelectorType->getNumElements();
925 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
926 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
927 "Wrong arguments for variable blend intrinsic");
928 SmallVector<Constant *, 32> Selectors;
929 for (unsigned I = 0; I < Size; ++I) {
930 // The intrinsics only read the top bit
933 Selector = C->getElementAsInteger(I);
935 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
936 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
938 auto NewSelector = ConstantVector::get(Selectors);
939 return SelectInst::Create(NewSelector, II->getArgOperand(1),
940 II->getArgOperand(0), "blendv");
946 case Intrinsic::x86_avx_vpermilvar_ps:
947 case Intrinsic::x86_avx_vpermilvar_ps_256:
948 case Intrinsic::x86_avx_vpermilvar_pd:
949 case Intrinsic::x86_avx_vpermilvar_pd_256: {
950 // Convert vpermil* to shufflevector if the mask is constant.
951 Value *V = II->getArgOperand(1);
952 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
953 assert(Size == 8 || Size == 4 || Size == 2);
955 if (auto C = dyn_cast<ConstantDataVector>(V)) {
956 // The intrinsics only read one or two bits, clear the rest.
957 for (unsigned I = 0; I < Size; ++I) {
958 uint32_t Index = C->getElementAsInteger(I) & 0x3;
959 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
960 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
964 } else if (isa<ConstantAggregateZero>(V)) {
965 for (unsigned I = 0; I < Size; ++I)
970 // The _256 variants are a bit trickier since the mask bits always index
971 // into the corresponding 128 half. In order to convert to a generic
972 // shuffle, we have to make that explicit.
973 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
974 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
975 for (unsigned I = Size / 2; I < Size; ++I)
976 Indexes[I] += Size / 2;
979 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
980 auto V1 = II->getArgOperand(0);
981 auto V2 = UndefValue::get(V1->getType());
982 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
983 return ReplaceInstUsesWith(CI, Shuffle);
986 case Intrinsic::x86_avx_vperm2f128_pd_256:
987 case Intrinsic::x86_avx_vperm2f128_ps_256:
988 case Intrinsic::x86_avx_vperm2f128_si_256:
989 case Intrinsic::x86_avx2_vperm2i128:
990 if (Value *V = SimplifyX86vperm2(*II, *Builder))
991 return ReplaceInstUsesWith(*II, V);
994 case Intrinsic::ppc_altivec_vperm:
995 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
996 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
997 // a vectorshuffle for little endian, we must undo the transformation
998 // performed on vec_perm in altivec.h. That is, we must complement
999 // the permutation mask with respect to 31 and reverse the order of
1001 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1002 assert(Mask->getType()->getVectorNumElements() == 16 &&
1003 "Bad type for intrinsic!");
1005 // Check that all of the elements are integer constants or undefs.
1006 bool AllEltsOk = true;
1007 for (unsigned i = 0; i != 16; ++i) {
1008 Constant *Elt = Mask->getAggregateElement(i);
1009 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1016 // Cast the input vectors to byte vectors.
1017 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1019 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1021 Value *Result = UndefValue::get(Op0->getType());
1023 // Only extract each element once.
1024 Value *ExtractedElts[32];
1025 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1027 for (unsigned i = 0; i != 16; ++i) {
1028 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1031 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1032 Idx &= 31; // Match the hardware behavior.
1033 if (DL.isLittleEndian())
1036 if (!ExtractedElts[Idx]) {
1037 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1038 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1039 ExtractedElts[Idx] =
1040 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1041 Builder->getInt32(Idx&15));
1044 // Insert this value into the result vector.
1045 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1046 Builder->getInt32(i));
1048 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1053 case Intrinsic::arm_neon_vld1:
1054 case Intrinsic::arm_neon_vld2:
1055 case Intrinsic::arm_neon_vld3:
1056 case Intrinsic::arm_neon_vld4:
1057 case Intrinsic::arm_neon_vld2lane:
1058 case Intrinsic::arm_neon_vld3lane:
1059 case Intrinsic::arm_neon_vld4lane:
1060 case Intrinsic::arm_neon_vst1:
1061 case Intrinsic::arm_neon_vst2:
1062 case Intrinsic::arm_neon_vst3:
1063 case Intrinsic::arm_neon_vst4:
1064 case Intrinsic::arm_neon_vst2lane:
1065 case Intrinsic::arm_neon_vst3lane:
1066 case Intrinsic::arm_neon_vst4lane: {
1067 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1068 unsigned AlignArg = II->getNumArgOperands() - 1;
1069 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1070 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1071 II->setArgOperand(AlignArg,
1072 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1079 case Intrinsic::arm_neon_vmulls:
1080 case Intrinsic::arm_neon_vmullu:
1081 case Intrinsic::aarch64_neon_smull:
1082 case Intrinsic::aarch64_neon_umull: {
1083 Value *Arg0 = II->getArgOperand(0);
1084 Value *Arg1 = II->getArgOperand(1);
1086 // Handle mul by zero first:
1087 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1088 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1091 // Check for constant LHS & RHS - in this case we just simplify.
1092 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1093 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1094 VectorType *NewVT = cast<VectorType>(II->getType());
1095 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1096 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1097 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1098 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1100 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1103 // Couldn't simplify - canonicalize constant to the RHS.
1104 std::swap(Arg0, Arg1);
1107 // Handle mul by one:
1108 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1109 if (ConstantInt *Splat =
1110 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1112 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1113 /*isSigned=*/!Zext);
1118 case Intrinsic::AMDGPU_rcp: {
1119 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1120 const APFloat &ArgVal = C->getValueAPF();
1121 APFloat Val(ArgVal.getSemantics(), 1.0);
1122 APFloat::opStatus Status = Val.divide(ArgVal,
1123 APFloat::rmNearestTiesToEven);
1124 // Only do this if it was exact and therefore not dependent on the
1126 if (Status == APFloat::opOK)
1127 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1132 case Intrinsic::stackrestore: {
1133 // If the save is right next to the restore, remove the restore. This can
1134 // happen when variable allocas are DCE'd.
1135 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1136 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1137 BasicBlock::iterator BI = SS;
1139 return EraseInstFromFunction(CI);
1143 // Scan down this block to see if there is another stack restore in the
1144 // same block without an intervening call/alloca.
1145 BasicBlock::iterator BI = II;
1146 TerminatorInst *TI = II->getParent()->getTerminator();
1147 bool CannotRemove = false;
1148 for (++BI; &*BI != TI; ++BI) {
1149 if (isa<AllocaInst>(BI)) {
1150 CannotRemove = true;
1153 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1154 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1155 // If there is a stackrestore below this one, remove this one.
1156 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1157 return EraseInstFromFunction(CI);
1158 // Otherwise, ignore the intrinsic.
1160 // If we found a non-intrinsic call, we can't remove the stack
1162 CannotRemove = true;
1168 // If the stack restore is in a return, resume, or unwind block and if there
1169 // are no allocas or calls between the restore and the return, nuke the
1171 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1172 return EraseInstFromFunction(CI);
1175 case Intrinsic::assume: {
1176 // Canonicalize assume(a && b) -> assume(a); assume(b);
1177 // Note: New assumption intrinsics created here are registered by
1178 // the InstCombineIRInserter object.
1179 Value *IIOperand = II->getArgOperand(0), *A, *B,
1180 *AssumeIntrinsic = II->getCalledValue();
1181 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1182 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1183 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1184 return EraseInstFromFunction(*II);
1186 // assume(!(a || b)) -> assume(!a); assume(!b);
1187 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1188 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1190 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1192 return EraseInstFromFunction(*II);
1195 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1196 // (if assume is valid at the load)
1197 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1198 Value *LHS = ICmp->getOperand(0);
1199 Value *RHS = ICmp->getOperand(1);
1200 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1201 isa<LoadInst>(LHS) &&
1202 isa<Constant>(RHS) &&
1203 RHS->getType()->isPointerTy() &&
1204 cast<Constant>(RHS)->isNullValue()) {
1205 LoadInst* LI = cast<LoadInst>(LHS);
1206 if (isValidAssumeForContext(II, LI, DT)) {
1207 MDNode *MD = MDNode::get(II->getContext(), None);
1208 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1209 return EraseInstFromFunction(*II);
1212 // TODO: apply nonnull return attributes to calls and invokes
1213 // TODO: apply range metadata for range check patterns?
1215 // If there is a dominating assume with the same condition as this one,
1216 // then this one is redundant, and should be removed.
1217 APInt KnownZero(1, 0), KnownOne(1, 0);
1218 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1219 if (KnownOne.isAllOnesValue())
1220 return EraseInstFromFunction(*II);
1224 case Intrinsic::experimental_gc_relocate: {
1225 // Translate facts known about a pointer before relocating into
1226 // facts about the relocate value, while being careful to
1227 // preserve relocation semantics.
1228 GCRelocateOperands Operands(II);
1229 Value *DerivedPtr = Operands.getDerivedPtr();
1230 auto *GCRelocateType = cast<PointerType>(II->getType());
1232 // Remove the relocation if unused, note that this check is required
1233 // to prevent the cases below from looping forever.
1234 if (II->use_empty())
1235 return EraseInstFromFunction(*II);
1237 // Undef is undef, even after relocation.
1238 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1239 // most practical collectors, but there was discussion in the review thread
1240 // about whether it was legal for all possible collectors.
1241 if (isa<UndefValue>(DerivedPtr)) {
1242 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1243 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1246 // The relocation of null will be null for most any collector.
1247 // TODO: provide a hook for this in GCStrategy. There might be some weird
1248 // collector this property does not hold for.
1249 if (isa<ConstantPointerNull>(DerivedPtr)) {
1250 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1251 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1254 // isKnownNonNull -> nonnull attribute
1255 if (isKnownNonNull(DerivedPtr))
1256 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1258 // isDereferenceablePointer -> deref attribute
1259 if (isDereferenceablePointer(DerivedPtr, DL)) {
1260 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1261 uint64_t Bytes = A->getDereferenceableBytes();
1262 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1266 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1267 // Canonicalize on the type from the uses to the defs
1269 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1273 return visitCallSite(II);
1276 // InvokeInst simplification
1278 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1279 return visitCallSite(&II);
1282 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1283 /// passed through the varargs area, we can eliminate the use of the cast.
1284 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1285 const DataLayout &DL,
1286 const CastInst *const CI,
1288 if (!CI->isLosslessCast())
1291 // If this is a GC intrinsic, avoid munging types. We need types for
1292 // statepoint reconstruction in SelectionDAG.
1293 // TODO: This is probably something which should be expanded to all
1294 // intrinsics since the entire point of intrinsics is that
1295 // they are understandable by the optimizer.
1296 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1299 // The size of ByVal or InAlloca arguments is derived from the type, so we
1300 // can't change to a type with a different size. If the size were
1301 // passed explicitly we could avoid this check.
1302 if (!CS.isByValOrInAllocaArgument(ix))
1306 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1307 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1308 if (!SrcTy->isSized() || !DstTy->isSized())
1310 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1315 // Try to fold some different type of calls here.
1316 // Currently we're only working with the checking functions, memcpy_chk,
1317 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1318 // strcat_chk and strncat_chk.
1319 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1320 if (!CI->getCalledFunction()) return nullptr;
1322 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1323 ReplaceInstUsesWith(*From, With);
1325 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1326 if (Value *With = Simplifier.optimizeCall(CI)) {
1328 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1334 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1335 // Strip off at most one level of pointer casts, looking for an alloca. This
1336 // is good enough in practice and simpler than handling any number of casts.
1337 Value *Underlying = TrampMem->stripPointerCasts();
1338 if (Underlying != TrampMem &&
1339 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1341 if (!isa<AllocaInst>(Underlying))
1344 IntrinsicInst *InitTrampoline = nullptr;
1345 for (User *U : TrampMem->users()) {
1346 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1349 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1351 // More than one init_trampoline writes to this value. Give up.
1353 InitTrampoline = II;
1356 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1357 // Allow any number of calls to adjust.trampoline.
1362 // No call to init.trampoline found.
1363 if (!InitTrampoline)
1366 // Check that the alloca is being used in the expected way.
1367 if (InitTrampoline->getOperand(0) != TrampMem)
1370 return InitTrampoline;
1373 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1375 // Visit all the previous instructions in the basic block, and try to find a
1376 // init.trampoline which has a direct path to the adjust.trampoline.
1377 for (BasicBlock::iterator I = AdjustTramp,
1378 E = AdjustTramp->getParent()->begin(); I != E; ) {
1379 Instruction *Inst = --I;
1380 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1381 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1382 II->getOperand(0) == TrampMem)
1384 if (Inst->mayWriteToMemory())
1390 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1391 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1392 // to a direct call to a function. Otherwise return NULL.
1394 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1395 Callee = Callee->stripPointerCasts();
1396 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1398 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1401 Value *TrampMem = AdjustTramp->getOperand(0);
1403 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1405 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1410 // visitCallSite - Improvements for call and invoke instructions.
1412 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1414 if (isAllocLikeFn(CS.getInstruction(), TLI))
1415 return visitAllocSite(*CS.getInstruction());
1417 bool Changed = false;
1419 // Mark any parameters that are known to be non-null with the nonnull
1420 // attribute. This is helpful for inlining calls to functions with null
1421 // checks on their arguments.
1423 for (Value *V : CS.args()) {
1424 if (!CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1425 isKnownNonNull(V)) {
1426 AttributeSet AS = CS.getAttributes();
1427 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1428 Attribute::NonNull);
1429 CS.setAttributes(AS);
1434 assert(ArgNo == CS.arg_size() && "sanity check");
1436 // If the callee is a pointer to a function, attempt to move any casts to the
1437 // arguments of the call/invoke.
1438 Value *Callee = CS.getCalledValue();
1439 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1442 if (Function *CalleeF = dyn_cast<Function>(Callee))
1443 // If the call and callee calling conventions don't match, this call must
1444 // be unreachable, as the call is undefined.
1445 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1446 // Only do this for calls to a function with a body. A prototype may
1447 // not actually end up matching the implementation's calling conv for a
1448 // variety of reasons (e.g. it may be written in assembly).
1449 !CalleeF->isDeclaration()) {
1450 Instruction *OldCall = CS.getInstruction();
1451 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1452 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1454 // If OldCall does not return void then replaceAllUsesWith undef.
1455 // This allows ValueHandlers and custom metadata to adjust itself.
1456 if (!OldCall->getType()->isVoidTy())
1457 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1458 if (isa<CallInst>(OldCall))
1459 return EraseInstFromFunction(*OldCall);
1461 // We cannot remove an invoke, because it would change the CFG, just
1462 // change the callee to a null pointer.
1463 cast<InvokeInst>(OldCall)->setCalledFunction(
1464 Constant::getNullValue(CalleeF->getType()));
1468 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1469 // If CS does not return void then replaceAllUsesWith undef.
1470 // This allows ValueHandlers and custom metadata to adjust itself.
1471 if (!CS.getInstruction()->getType()->isVoidTy())
1472 ReplaceInstUsesWith(*CS.getInstruction(),
1473 UndefValue::get(CS.getInstruction()->getType()));
1475 if (isa<InvokeInst>(CS.getInstruction())) {
1476 // Can't remove an invoke because we cannot change the CFG.
1480 // This instruction is not reachable, just remove it. We insert a store to
1481 // undef so that we know that this code is not reachable, despite the fact
1482 // that we can't modify the CFG here.
1483 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1484 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1485 CS.getInstruction());
1487 return EraseInstFromFunction(*CS.getInstruction());
1490 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1491 return transformCallThroughTrampoline(CS, II);
1493 PointerType *PTy = cast<PointerType>(Callee->getType());
1494 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1495 if (FTy->isVarArg()) {
1496 int ix = FTy->getNumParams();
1497 // See if we can optimize any arguments passed through the varargs area of
1499 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1500 E = CS.arg_end(); I != E; ++I, ++ix) {
1501 CastInst *CI = dyn_cast<CastInst>(*I);
1502 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1503 *I = CI->getOperand(0);
1509 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1510 // Inline asm calls cannot throw - mark them 'nounwind'.
1511 CS.setDoesNotThrow();
1515 // Try to optimize the call if possible, we require DataLayout for most of
1516 // this. None of these calls are seen as possibly dead so go ahead and
1517 // delete the instruction now.
1518 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1519 Instruction *I = tryOptimizeCall(CI);
1520 // If we changed something return the result, etc. Otherwise let
1521 // the fallthrough check.
1522 if (I) return EraseInstFromFunction(*I);
1525 return Changed ? CS.getInstruction() : nullptr;
1528 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1529 // attempt to move the cast to the arguments of the call/invoke.
1531 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1533 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1536 // The prototype of thunks are a lie, don't try to directly call such
1538 if (Callee->hasFnAttribute("thunk"))
1540 Instruction *Caller = CS.getInstruction();
1541 const AttributeSet &CallerPAL = CS.getAttributes();
1543 // Okay, this is a cast from a function to a different type. Unless doing so
1544 // would cause a type conversion of one of our arguments, change this call to
1545 // be a direct call with arguments casted to the appropriate types.
1547 FunctionType *FT = Callee->getFunctionType();
1548 Type *OldRetTy = Caller->getType();
1549 Type *NewRetTy = FT->getReturnType();
1551 // Check to see if we are changing the return type...
1552 if (OldRetTy != NewRetTy) {
1554 if (NewRetTy->isStructTy())
1555 return false; // TODO: Handle multiple return values.
1557 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1558 if (Callee->isDeclaration())
1559 return false; // Cannot transform this return value.
1561 if (!Caller->use_empty() &&
1562 // void -> non-void is handled specially
1563 !NewRetTy->isVoidTy())
1564 return false; // Cannot transform this return value.
1567 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1568 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1569 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1570 return false; // Attribute not compatible with transformed value.
1573 // If the callsite is an invoke instruction, and the return value is used by
1574 // a PHI node in a successor, we cannot change the return type of the call
1575 // because there is no place to put the cast instruction (without breaking
1576 // the critical edge). Bail out in this case.
1577 if (!Caller->use_empty())
1578 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1579 for (User *U : II->users())
1580 if (PHINode *PN = dyn_cast<PHINode>(U))
1581 if (PN->getParent() == II->getNormalDest() ||
1582 PN->getParent() == II->getUnwindDest())
1586 unsigned NumActualArgs = CS.arg_size();
1587 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1589 // Prevent us turning:
1590 // declare void @takes_i32_inalloca(i32* inalloca)
1591 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1594 // call void @takes_i32_inalloca(i32* null)
1596 // Similarly, avoid folding away bitcasts of byval calls.
1597 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1598 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1601 CallSite::arg_iterator AI = CS.arg_begin();
1602 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1603 Type *ParamTy = FT->getParamType(i);
1604 Type *ActTy = (*AI)->getType();
1606 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1607 return false; // Cannot transform this parameter value.
1609 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1610 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1611 return false; // Attribute not compatible with transformed value.
1613 if (CS.isInAllocaArgument(i))
1614 return false; // Cannot transform to and from inalloca.
1616 // If the parameter is passed as a byval argument, then we have to have a
1617 // sized type and the sized type has to have the same size as the old type.
1618 if (ParamTy != ActTy &&
1619 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1620 Attribute::ByVal)) {
1621 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1622 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1625 Type *CurElTy = ActTy->getPointerElementType();
1626 if (DL.getTypeAllocSize(CurElTy) !=
1627 DL.getTypeAllocSize(ParamPTy->getElementType()))
1632 if (Callee->isDeclaration()) {
1633 // Do not delete arguments unless we have a function body.
1634 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1637 // If the callee is just a declaration, don't change the varargsness of the
1638 // call. We don't want to introduce a varargs call where one doesn't
1640 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1641 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1644 // If both the callee and the cast type are varargs, we still have to make
1645 // sure the number of fixed parameters are the same or we have the same
1646 // ABI issues as if we introduce a varargs call.
1647 if (FT->isVarArg() &&
1648 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1649 FT->getNumParams() !=
1650 cast<FunctionType>(APTy->getElementType())->getNumParams())
1654 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1655 !CallerPAL.isEmpty())
1656 // In this case we have more arguments than the new function type, but we
1657 // won't be dropping them. Check that these extra arguments have attributes
1658 // that are compatible with being a vararg call argument.
1659 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1660 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1661 if (Index <= FT->getNumParams())
1664 // Check if it has an attribute that's incompatible with varargs.
1665 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1666 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1671 // Okay, we decided that this is a safe thing to do: go ahead and start
1672 // inserting cast instructions as necessary.
1673 std::vector<Value*> Args;
1674 Args.reserve(NumActualArgs);
1675 SmallVector<AttributeSet, 8> attrVec;
1676 attrVec.reserve(NumCommonArgs);
1678 // Get any return attributes.
1679 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1681 // If the return value is not being used, the type may not be compatible
1682 // with the existing attributes. Wipe out any problematic attributes.
1683 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1685 // Add the new return attributes.
1686 if (RAttrs.hasAttributes())
1687 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1688 AttributeSet::ReturnIndex, RAttrs));
1690 AI = CS.arg_begin();
1691 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1692 Type *ParamTy = FT->getParamType(i);
1694 if ((*AI)->getType() == ParamTy) {
1695 Args.push_back(*AI);
1697 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1700 // Add any parameter attributes.
1701 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1702 if (PAttrs.hasAttributes())
1703 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1707 // If the function takes more arguments than the call was taking, add them
1709 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1710 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1712 // If we are removing arguments to the function, emit an obnoxious warning.
1713 if (FT->getNumParams() < NumActualArgs) {
1714 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1715 if (FT->isVarArg()) {
1716 // Add all of the arguments in their promoted form to the arg list.
1717 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1718 Type *PTy = getPromotedType((*AI)->getType());
1719 if (PTy != (*AI)->getType()) {
1720 // Must promote to pass through va_arg area!
1721 Instruction::CastOps opcode =
1722 CastInst::getCastOpcode(*AI, false, PTy, false);
1723 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1725 Args.push_back(*AI);
1728 // Add any parameter attributes.
1729 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1730 if (PAttrs.hasAttributes())
1731 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1737 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1738 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1739 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1741 if (NewRetTy->isVoidTy())
1742 Caller->setName(""); // Void type should not have a name.
1744 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1748 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1749 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1750 II->getUnwindDest(), Args);
1752 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1753 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1755 CallInst *CI = cast<CallInst>(Caller);
1756 NC = Builder->CreateCall(Callee, Args);
1758 if (CI->isTailCall())
1759 cast<CallInst>(NC)->setTailCall();
1760 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1761 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1764 // Insert a cast of the return type as necessary.
1766 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1767 if (!NV->getType()->isVoidTy()) {
1768 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1769 NC->setDebugLoc(Caller->getDebugLoc());
1771 // If this is an invoke instruction, we should insert it after the first
1772 // non-phi, instruction in the normal successor block.
1773 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1774 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1775 InsertNewInstBefore(NC, *I);
1777 // Otherwise, it's a call, just insert cast right after the call.
1778 InsertNewInstBefore(NC, *Caller);
1780 Worklist.AddUsersToWorkList(*Caller);
1782 NV = UndefValue::get(Caller->getType());
1786 if (!Caller->use_empty())
1787 ReplaceInstUsesWith(*Caller, NV);
1788 else if (Caller->hasValueHandle()) {
1789 if (OldRetTy == NV->getType())
1790 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1792 // We cannot call ValueIsRAUWd with a different type, and the
1793 // actual tracked value will disappear.
1794 ValueHandleBase::ValueIsDeleted(Caller);
1797 EraseInstFromFunction(*Caller);
1801 // transformCallThroughTrampoline - Turn a call to a function created by
1802 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1803 // underlying function.
1806 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1807 IntrinsicInst *Tramp) {
1808 Value *Callee = CS.getCalledValue();
1809 PointerType *PTy = cast<PointerType>(Callee->getType());
1810 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1811 const AttributeSet &Attrs = CS.getAttributes();
1813 // If the call already has the 'nest' attribute somewhere then give up -
1814 // otherwise 'nest' would occur twice after splicing in the chain.
1815 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1819 "transformCallThroughTrampoline called with incorrect CallSite.");
1821 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1822 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1823 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1825 const AttributeSet &NestAttrs = NestF->getAttributes();
1826 if (!NestAttrs.isEmpty()) {
1827 unsigned NestIdx = 1;
1828 Type *NestTy = nullptr;
1829 AttributeSet NestAttr;
1831 // Look for a parameter marked with the 'nest' attribute.
1832 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1833 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1834 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1835 // Record the parameter type and any other attributes.
1837 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1842 Instruction *Caller = CS.getInstruction();
1843 std::vector<Value*> NewArgs;
1844 NewArgs.reserve(CS.arg_size() + 1);
1846 SmallVector<AttributeSet, 8> NewAttrs;
1847 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1849 // Insert the nest argument into the call argument list, which may
1850 // mean appending it. Likewise for attributes.
1852 // Add any result attributes.
1853 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1854 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1855 Attrs.getRetAttributes()));
1859 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1861 if (Idx == NestIdx) {
1862 // Add the chain argument and attributes.
1863 Value *NestVal = Tramp->getArgOperand(2);
1864 if (NestVal->getType() != NestTy)
1865 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1866 NewArgs.push_back(NestVal);
1867 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1874 // Add the original argument and attributes.
1875 NewArgs.push_back(*I);
1876 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1877 if (Attr.hasAttributes(Idx)) {
1878 AttrBuilder B(Attr, Idx);
1879 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1880 Idx + (Idx >= NestIdx), B));
1887 // Add any function attributes.
1888 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1889 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1890 Attrs.getFnAttributes()));
1892 // The trampoline may have been bitcast to a bogus type (FTy).
1893 // Handle this by synthesizing a new function type, equal to FTy
1894 // with the chain parameter inserted.
1896 std::vector<Type*> NewTypes;
1897 NewTypes.reserve(FTy->getNumParams()+1);
1899 // Insert the chain's type into the list of parameter types, which may
1900 // mean appending it.
1903 FunctionType::param_iterator I = FTy->param_begin(),
1904 E = FTy->param_end();
1908 // Add the chain's type.
1909 NewTypes.push_back(NestTy);
1914 // Add the original type.
1915 NewTypes.push_back(*I);
1921 // Replace the trampoline call with a direct call. Let the generic
1922 // code sort out any function type mismatches.
1923 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1925 Constant *NewCallee =
1926 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1927 NestF : ConstantExpr::getBitCast(NestF,
1928 PointerType::getUnqual(NewFTy));
1929 const AttributeSet &NewPAL =
1930 AttributeSet::get(FTy->getContext(), NewAttrs);
1932 Instruction *NewCaller;
1933 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1934 NewCaller = InvokeInst::Create(NewCallee,
1935 II->getNormalDest(), II->getUnwindDest(),
1937 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1938 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1940 NewCaller = CallInst::Create(NewCallee, NewArgs);
1941 if (cast<CallInst>(Caller)->isTailCall())
1942 cast<CallInst>(NewCaller)->setTailCall();
1943 cast<CallInst>(NewCaller)->
1944 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1945 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1952 // Replace the trampoline call with a direct call. Since there is no 'nest'
1953 // parameter, there is no need to adjust the argument list. Let the generic
1954 // code sort out any function type mismatches.
1955 Constant *NewCallee =
1956 NestF->getType() == PTy ? NestF :
1957 ConstantExpr::getBitCast(NestF, PTy);
1958 CS.setCalledFunction(NewCallee);
1959 return CS.getInstruction();