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 *SimplifyX86extend(const IntrinsicInst &II,
201 InstCombiner::BuilderTy &Builder,
203 VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType());
204 VectorType *DstTy = cast<VectorType>(II.getType());
205 unsigned NumDstElts = DstTy->getNumElements();
207 // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
208 SmallVector<int, 8> ShuffleMask;
209 for (int i = 0; i != (int)NumDstElts; ++i)
210 ShuffleMask.push_back(i);
212 Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0),
213 UndefValue::get(SrcTy), ShuffleMask);
214 return SignExtend ? Builder.CreateSExt(SV, DstTy)
215 : Builder.CreateZExt(SV, DstTy);
218 static Value *SimplifyX86insertps(const IntrinsicInst &II,
219 InstCombiner::BuilderTy &Builder) {
220 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
221 VectorType *VecTy = cast<VectorType>(II.getType());
222 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
224 // The immediate permute control byte looks like this:
225 // [3:0] - zero mask for each 32-bit lane
226 // [5:4] - select one 32-bit destination lane
227 // [7:6] - select one 32-bit source lane
229 uint8_t Imm = CInt->getZExtValue();
230 uint8_t ZMask = Imm & 0xf;
231 uint8_t DestLane = (Imm >> 4) & 0x3;
232 uint8_t SourceLane = (Imm >> 6) & 0x3;
234 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
236 // If all zero mask bits are set, this was just a weird way to
237 // generate a zero vector.
241 // Initialize by passing all of the first source bits through.
242 int ShuffleMask[4] = { 0, 1, 2, 3 };
244 // We may replace the second operand with the zero vector.
245 Value *V1 = II.getArgOperand(1);
248 // If the zero mask is being used with a single input or the zero mask
249 // overrides the destination lane, this is a shuffle with the zero vector.
250 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
251 (ZMask & (1 << DestLane))) {
253 // We may still move 32-bits of the first source vector from one lane
255 ShuffleMask[DestLane] = SourceLane;
256 // The zero mask may override the previous insert operation.
257 for (unsigned i = 0; i < 4; ++i)
258 if ((ZMask >> i) & 0x1)
259 ShuffleMask[i] = i + 4;
261 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
265 // Replace the selected destination lane with the selected source lane.
266 ShuffleMask[DestLane] = SourceLane + 4;
269 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
274 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
275 /// source vectors, unless a zero bit is set. If a zero bit is set,
276 /// then ignore that half of the mask and clear that half of the vector.
277 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
278 InstCombiner::BuilderTy &Builder) {
279 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
280 VectorType *VecTy = cast<VectorType>(II.getType());
281 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
283 // The immediate permute control byte looks like this:
284 // [1:0] - select 128 bits from sources for low half of destination
286 // [3] - zero low half of destination
287 // [5:4] - select 128 bits from sources for high half of destination
289 // [7] - zero high half of destination
291 uint8_t Imm = CInt->getZExtValue();
293 bool LowHalfZero = Imm & 0x08;
294 bool HighHalfZero = Imm & 0x80;
296 // If both zero mask bits are set, this was just a weird way to
297 // generate a zero vector.
298 if (LowHalfZero && HighHalfZero)
301 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
302 unsigned NumElts = VecTy->getNumElements();
303 unsigned HalfSize = NumElts / 2;
304 SmallVector<int, 8> ShuffleMask(NumElts);
306 // The high bit of the selection field chooses the 1st or 2nd operand.
307 bool LowInputSelect = Imm & 0x02;
308 bool HighInputSelect = Imm & 0x20;
310 // The low bit of the selection field chooses the low or high half
311 // of the selected operand.
312 bool LowHalfSelect = Imm & 0x01;
313 bool HighHalfSelect = Imm & 0x10;
315 // Determine which operand(s) are actually in use for this instruction.
316 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
317 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
319 // If needed, replace operands based on zero mask.
320 V0 = LowHalfZero ? ZeroVector : V0;
321 V1 = HighHalfZero ? ZeroVector : V1;
323 // Permute low half of result.
324 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
325 for (unsigned i = 0; i < HalfSize; ++i)
326 ShuffleMask[i] = StartIndex + i;
328 // Permute high half of result.
329 StartIndex = HighHalfSelect ? HalfSize : 0;
330 StartIndex += NumElts;
331 for (unsigned i = 0; i < HalfSize; ++i)
332 ShuffleMask[i + HalfSize] = StartIndex + i;
334 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
339 /// visitCallInst - CallInst simplification. This mostly only handles folding
340 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
341 /// the heavy lifting.
343 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
344 auto Args = CI.arg_operands();
345 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
347 return ReplaceInstUsesWith(CI, V);
349 if (isFreeCall(&CI, TLI))
350 return visitFree(CI);
352 // If the caller function is nounwind, mark the call as nounwind, even if the
354 if (CI.getParent()->getParent()->doesNotThrow() &&
355 !CI.doesNotThrow()) {
356 CI.setDoesNotThrow();
360 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
361 if (!II) return visitCallSite(&CI);
363 // Intrinsics cannot occur in an invoke, so handle them here instead of in
365 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
366 bool Changed = false;
368 // memmove/cpy/set of zero bytes is a noop.
369 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
370 if (NumBytes->isNullValue())
371 return EraseInstFromFunction(CI);
373 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
374 if (CI->getZExtValue() == 1) {
375 // Replace the instruction with just byte operations. We would
376 // transform other cases to loads/stores, but we don't know if
377 // alignment is sufficient.
381 // No other transformations apply to volatile transfers.
382 if (MI->isVolatile())
385 // If we have a memmove and the source operation is a constant global,
386 // then the source and dest pointers can't alias, so we can change this
387 // into a call to memcpy.
388 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
389 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
390 if (GVSrc->isConstant()) {
391 Module *M = CI.getParent()->getParent()->getParent();
392 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
393 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
394 CI.getArgOperand(1)->getType(),
395 CI.getArgOperand(2)->getType() };
396 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
401 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
402 // memmove(x,x,size) -> noop.
403 if (MTI->getSource() == MTI->getDest())
404 return EraseInstFromFunction(CI);
407 // If we can determine a pointer alignment that is bigger than currently
408 // set, update the alignment.
409 if (isa<MemTransferInst>(MI)) {
410 if (Instruction *I = SimplifyMemTransfer(MI))
412 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
413 if (Instruction *I = SimplifyMemSet(MSI))
417 if (Changed) return II;
420 switch (II->getIntrinsicID()) {
422 case Intrinsic::objectsize: {
424 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
425 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
428 case Intrinsic::bswap: {
429 Value *IIOperand = II->getArgOperand(0);
432 // bswap(bswap(x)) -> x
433 if (match(IIOperand, m_BSwap(m_Value(X))))
434 return ReplaceInstUsesWith(CI, X);
436 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
437 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
438 unsigned C = X->getType()->getPrimitiveSizeInBits() -
439 IIOperand->getType()->getPrimitiveSizeInBits();
440 Value *CV = ConstantInt::get(X->getType(), C);
441 Value *V = Builder->CreateLShr(X, CV);
442 return new TruncInst(V, IIOperand->getType());
447 case Intrinsic::powi:
448 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
451 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
454 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
455 // powi(x, -1) -> 1/x
456 if (Power->isAllOnesValue())
457 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
458 II->getArgOperand(0));
461 case Intrinsic::cttz: {
462 // If all bits below the first known one are known zero,
463 // this value is constant.
464 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
465 // FIXME: Try to simplify vectors of integers.
467 uint32_t BitWidth = IT->getBitWidth();
468 APInt KnownZero(BitWidth, 0);
469 APInt KnownOne(BitWidth, 0);
470 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
471 unsigned TrailingZeros = KnownOne.countTrailingZeros();
472 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
473 if ((Mask & KnownZero) == Mask)
474 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
475 APInt(BitWidth, TrailingZeros)));
479 case Intrinsic::ctlz: {
480 // If all bits above the first known one are known zero,
481 // this value is constant.
482 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
483 // FIXME: Try to simplify vectors of integers.
485 uint32_t BitWidth = IT->getBitWidth();
486 APInt KnownZero(BitWidth, 0);
487 APInt KnownOne(BitWidth, 0);
488 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
489 unsigned LeadingZeros = KnownOne.countLeadingZeros();
490 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
491 if ((Mask & KnownZero) == Mask)
492 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
493 APInt(BitWidth, LeadingZeros)));
498 case Intrinsic::uadd_with_overflow:
499 case Intrinsic::sadd_with_overflow:
500 case Intrinsic::umul_with_overflow:
501 case Intrinsic::smul_with_overflow:
502 if (isa<Constant>(II->getArgOperand(0)) &&
503 !isa<Constant>(II->getArgOperand(1))) {
504 // Canonicalize constants into the RHS.
505 Value *LHS = II->getArgOperand(0);
506 II->setArgOperand(0, II->getArgOperand(1));
507 II->setArgOperand(1, LHS);
512 case Intrinsic::usub_with_overflow:
513 case Intrinsic::ssub_with_overflow: {
514 OverflowCheckFlavor OCF =
515 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
516 assert(OCF != OCF_INVALID && "unexpected!");
518 Value *OperationResult = nullptr;
519 Constant *OverflowResult = nullptr;
520 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
521 *II, OperationResult, OverflowResult))
522 return CreateOverflowTuple(II, OperationResult, OverflowResult);
527 case Intrinsic::minnum:
528 case Intrinsic::maxnum: {
529 Value *Arg0 = II->getArgOperand(0);
530 Value *Arg1 = II->getArgOperand(1);
534 return ReplaceInstUsesWith(CI, Arg0);
536 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
537 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
539 // Canonicalize constants into the RHS.
541 II->setArgOperand(0, Arg1);
542 II->setArgOperand(1, Arg0);
547 if (C1 && C1->isNaN())
548 return ReplaceInstUsesWith(CI, Arg0);
550 // This is the value because if undef were NaN, we would return the other
551 // value and cannot return a NaN unless both operands are.
553 // fmin(undef, x) -> x
554 if (isa<UndefValue>(Arg0))
555 return ReplaceInstUsesWith(CI, Arg1);
557 // fmin(x, undef) -> x
558 if (isa<UndefValue>(Arg1))
559 return ReplaceInstUsesWith(CI, Arg0);
563 if (II->getIntrinsicID() == Intrinsic::minnum) {
564 // fmin(x, fmin(x, y)) -> fmin(x, y)
565 // fmin(y, fmin(x, y)) -> fmin(x, y)
566 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
567 if (Arg0 == X || Arg0 == Y)
568 return ReplaceInstUsesWith(CI, Arg1);
571 // fmin(fmin(x, y), x) -> fmin(x, y)
572 // fmin(fmin(x, y), y) -> fmin(x, y)
573 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
574 if (Arg1 == X || Arg1 == Y)
575 return ReplaceInstUsesWith(CI, Arg0);
578 // TODO: fmin(nnan x, inf) -> x
579 // TODO: fmin(nnan ninf x, flt_max) -> x
580 if (C1 && C1->isInfinity()) {
581 // fmin(x, -inf) -> -inf
582 if (C1->isNegative())
583 return ReplaceInstUsesWith(CI, Arg1);
586 assert(II->getIntrinsicID() == Intrinsic::maxnum);
587 // fmax(x, fmax(x, y)) -> fmax(x, y)
588 // fmax(y, fmax(x, y)) -> fmax(x, y)
589 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
590 if (Arg0 == X || Arg0 == Y)
591 return ReplaceInstUsesWith(CI, Arg1);
594 // fmax(fmax(x, y), x) -> fmax(x, y)
595 // fmax(fmax(x, y), y) -> fmax(x, y)
596 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
597 if (Arg1 == X || Arg1 == Y)
598 return ReplaceInstUsesWith(CI, Arg0);
601 // TODO: fmax(nnan x, -inf) -> x
602 // TODO: fmax(nnan ninf x, -flt_max) -> x
603 if (C1 && C1->isInfinity()) {
604 // fmax(x, inf) -> inf
605 if (!C1->isNegative())
606 return ReplaceInstUsesWith(CI, Arg1);
611 case Intrinsic::ppc_altivec_lvx:
612 case Intrinsic::ppc_altivec_lvxl:
613 // Turn PPC lvx -> load if the pointer is known aligned.
614 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
616 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
617 PointerType::getUnqual(II->getType()));
618 return new LoadInst(Ptr);
621 case Intrinsic::ppc_vsx_lxvw4x:
622 case Intrinsic::ppc_vsx_lxvd2x: {
623 // Turn PPC VSX loads into normal loads.
624 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
625 PointerType::getUnqual(II->getType()));
626 return new LoadInst(Ptr, Twine(""), false, 1);
628 case Intrinsic::ppc_altivec_stvx:
629 case Intrinsic::ppc_altivec_stvxl:
630 // Turn stvx -> store if the pointer is known aligned.
631 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
634 PointerType::getUnqual(II->getArgOperand(0)->getType());
635 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
636 return new StoreInst(II->getArgOperand(0), Ptr);
639 case Intrinsic::ppc_vsx_stxvw4x:
640 case Intrinsic::ppc_vsx_stxvd2x: {
641 // Turn PPC VSX stores into normal stores.
642 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
643 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
644 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
646 case Intrinsic::ppc_qpx_qvlfs:
647 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
648 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
650 Type *VTy = VectorType::get(Builder->getFloatTy(),
651 II->getType()->getVectorNumElements());
652 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
653 PointerType::getUnqual(VTy));
654 Value *Load = Builder->CreateLoad(Ptr);
655 return new FPExtInst(Load, II->getType());
658 case Intrinsic::ppc_qpx_qvlfd:
659 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
660 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
662 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
663 PointerType::getUnqual(II->getType()));
664 return new LoadInst(Ptr);
667 case Intrinsic::ppc_qpx_qvstfs:
668 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
669 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
671 Type *VTy = VectorType::get(Builder->getFloatTy(),
672 II->getArgOperand(0)->getType()->getVectorNumElements());
673 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
674 Type *OpPtrTy = PointerType::getUnqual(VTy);
675 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
676 return new StoreInst(TOp, Ptr);
679 case Intrinsic::ppc_qpx_qvstfd:
680 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
681 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
684 PointerType::getUnqual(II->getArgOperand(0)->getType());
685 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
686 return new StoreInst(II->getArgOperand(0), Ptr);
689 case Intrinsic::x86_sse_storeu_ps:
690 case Intrinsic::x86_sse2_storeu_pd:
691 case Intrinsic::x86_sse2_storeu_dq:
692 // Turn X86 storeu -> store if the pointer is known aligned.
693 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
696 PointerType::getUnqual(II->getArgOperand(1)->getType());
697 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
698 return new StoreInst(II->getArgOperand(1), Ptr);
702 case Intrinsic::x86_sse_cvtss2si:
703 case Intrinsic::x86_sse_cvtss2si64:
704 case Intrinsic::x86_sse_cvttss2si:
705 case Intrinsic::x86_sse_cvttss2si64:
706 case Intrinsic::x86_sse2_cvtsd2si:
707 case Intrinsic::x86_sse2_cvtsd2si64:
708 case Intrinsic::x86_sse2_cvttsd2si:
709 case Intrinsic::x86_sse2_cvttsd2si64: {
710 // These intrinsics only demand the 0th element of their input vectors. If
711 // we can simplify the input based on that, do so now.
713 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
714 APInt DemandedElts(VWidth, 1);
715 APInt UndefElts(VWidth, 0);
716 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
717 DemandedElts, UndefElts)) {
718 II->setArgOperand(0, V);
724 // Constant fold <A x Bi> << Ci.
725 // FIXME: We don't handle _dq because it's a shift of an i128, but is
726 // represented in the IR as <2 x i64>. A per element shift is wrong.
727 case Intrinsic::x86_sse2_psll_d:
728 case Intrinsic::x86_sse2_psll_q:
729 case Intrinsic::x86_sse2_psll_w:
730 case Intrinsic::x86_sse2_pslli_d:
731 case Intrinsic::x86_sse2_pslli_q:
732 case Intrinsic::x86_sse2_pslli_w:
733 case Intrinsic::x86_avx2_psll_d:
734 case Intrinsic::x86_avx2_psll_q:
735 case Intrinsic::x86_avx2_psll_w:
736 case Intrinsic::x86_avx2_pslli_d:
737 case Intrinsic::x86_avx2_pslli_q:
738 case Intrinsic::x86_avx2_pslli_w:
739 case Intrinsic::x86_sse2_psrl_d:
740 case Intrinsic::x86_sse2_psrl_q:
741 case Intrinsic::x86_sse2_psrl_w:
742 case Intrinsic::x86_sse2_psrli_d:
743 case Intrinsic::x86_sse2_psrli_q:
744 case Intrinsic::x86_sse2_psrli_w:
745 case Intrinsic::x86_avx2_psrl_d:
746 case Intrinsic::x86_avx2_psrl_q:
747 case Intrinsic::x86_avx2_psrl_w:
748 case Intrinsic::x86_avx2_psrli_d:
749 case Intrinsic::x86_avx2_psrli_q:
750 case Intrinsic::x86_avx2_psrli_w: {
751 // Simplify if count is constant. To 0 if >= BitWidth,
752 // otherwise to shl/lshr.
753 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
754 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
759 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
763 auto Vec = II->getArgOperand(0);
764 auto VT = cast<VectorType>(Vec->getType());
765 if (Count->getZExtValue() >
766 VT->getElementType()->getPrimitiveSizeInBits() - 1)
767 return ReplaceInstUsesWith(
768 CI, ConstantAggregateZero::get(Vec->getType()));
770 bool isPackedShiftLeft = true;
771 switch (II->getIntrinsicID()) {
773 case Intrinsic::x86_sse2_psrl_d:
774 case Intrinsic::x86_sse2_psrl_q:
775 case Intrinsic::x86_sse2_psrl_w:
776 case Intrinsic::x86_sse2_psrli_d:
777 case Intrinsic::x86_sse2_psrli_q:
778 case Intrinsic::x86_sse2_psrli_w:
779 case Intrinsic::x86_avx2_psrl_d:
780 case Intrinsic::x86_avx2_psrl_q:
781 case Intrinsic::x86_avx2_psrl_w:
782 case Intrinsic::x86_avx2_psrli_d:
783 case Intrinsic::x86_avx2_psrli_q:
784 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
787 unsigned VWidth = VT->getNumElements();
788 // Get a constant vector of the same type as the first operand.
789 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
790 if (isPackedShiftLeft)
791 return BinaryOperator::CreateShl(Vec,
792 Builder->CreateVectorSplat(VWidth, VTCI));
794 return BinaryOperator::CreateLShr(Vec,
795 Builder->CreateVectorSplat(VWidth, VTCI));
798 case Intrinsic::x86_sse41_pmovsxbd:
799 case Intrinsic::x86_sse41_pmovsxbq:
800 case Intrinsic::x86_sse41_pmovsxbw:
801 case Intrinsic::x86_sse41_pmovsxdq:
802 case Intrinsic::x86_sse41_pmovsxwd:
803 case Intrinsic::x86_sse41_pmovsxwq:
804 case Intrinsic::x86_avx2_pmovsxbd:
805 case Intrinsic::x86_avx2_pmovsxbq:
806 case Intrinsic::x86_avx2_pmovsxbw:
807 case Intrinsic::x86_avx2_pmovsxdq:
808 case Intrinsic::x86_avx2_pmovsxwd:
809 case Intrinsic::x86_avx2_pmovsxwq:
810 if (Value *V = SimplifyX86extend(*II, *Builder, true))
811 return ReplaceInstUsesWith(*II, V);
814 case Intrinsic::x86_sse41_pmovzxbd:
815 case Intrinsic::x86_sse41_pmovzxbq:
816 case Intrinsic::x86_sse41_pmovzxbw:
817 case Intrinsic::x86_sse41_pmovzxdq:
818 case Intrinsic::x86_sse41_pmovzxwd:
819 case Intrinsic::x86_sse41_pmovzxwq:
820 case Intrinsic::x86_avx2_pmovzxbd:
821 case Intrinsic::x86_avx2_pmovzxbq:
822 case Intrinsic::x86_avx2_pmovzxbw:
823 case Intrinsic::x86_avx2_pmovzxdq:
824 case Intrinsic::x86_avx2_pmovzxwd:
825 case Intrinsic::x86_avx2_pmovzxwq:
826 if (Value *V = SimplifyX86extend(*II, *Builder, false))
827 return ReplaceInstUsesWith(*II, V);
830 case Intrinsic::x86_sse41_insertps:
831 if (Value *V = SimplifyX86insertps(*II, *Builder))
832 return ReplaceInstUsesWith(*II, V);
835 case Intrinsic::x86_sse4a_insertqi: {
836 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
838 // TODO: eventually we should lower this intrinsic to IR
839 if (auto CILength = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
840 if (auto CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
841 unsigned Index = CIIndex->getZExtValue();
842 // From AMD documentation: "a value of zero in the field length is
843 // defined as length of 64".
844 unsigned Length = CILength->equalsInt(0) ? 64 : CILength->getZExtValue();
846 // From AMD documentation: "If the sum of the bit index + length field
847 // is greater than 64, the results are undefined".
848 unsigned End = Index + Length;
850 // Note that both field index and field length are 8-bit quantities.
851 // Since variables 'Index' and 'Length' are unsigned values
852 // obtained from zero-extending field index and field length
853 // respectively, their sum should never wrap around.
855 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
857 if (Length == 64 && Index == 0) {
858 Value *Vec = II->getArgOperand(1);
859 Value *Undef = UndefValue::get(Vec->getType());
860 const uint32_t Mask[] = { 0, 2 };
861 return ReplaceInstUsesWith(
863 Builder->CreateShuffleVector(
864 Vec, Undef, ConstantDataVector::get(
865 II->getContext(), makeArrayRef(Mask))));
866 } else if (auto Source =
867 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
868 if (Source->hasOneUse() &&
869 Source->getArgOperand(1) == II->getArgOperand(1)) {
870 // If the source of the insert has only one use and it's another
871 // insert (and they're both inserting from the same vector), try to
872 // bundle both together.
873 auto CISourceLength =
874 dyn_cast<ConstantInt>(Source->getArgOperand(2));
876 dyn_cast<ConstantInt>(Source->getArgOperand(3));
877 if (CISourceIndex && CISourceLength) {
878 unsigned SourceIndex = CISourceIndex->getZExtValue();
879 unsigned SourceLength = CISourceLength->getZExtValue();
880 unsigned SourceEnd = SourceIndex + SourceLength;
881 unsigned NewIndex, NewLength;
882 bool ShouldReplace = false;
883 if (Index <= SourceIndex && SourceIndex <= End) {
885 NewLength = std::max(End, SourceEnd) - NewIndex;
886 ShouldReplace = true;
887 } else if (SourceIndex <= Index && Index <= SourceEnd) {
888 NewIndex = SourceIndex;
889 NewLength = std::max(SourceEnd, End) - NewIndex;
890 ShouldReplace = true;
894 Constant *ConstantLength = ConstantInt::get(
895 II->getArgOperand(2)->getType(), NewLength, false);
896 Constant *ConstantIndex = ConstantInt::get(
897 II->getArgOperand(3)->getType(), NewIndex, false);
898 Value *Args[4] = { Source->getArgOperand(0),
899 II->getArgOperand(1), ConstantLength,
901 Module *M = CI.getParent()->getParent()->getParent();
903 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
904 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
914 case Intrinsic::x86_sse41_pblendvb:
915 case Intrinsic::x86_sse41_blendvps:
916 case Intrinsic::x86_sse41_blendvpd:
917 case Intrinsic::x86_avx_blendv_ps_256:
918 case Intrinsic::x86_avx_blendv_pd_256:
919 case Intrinsic::x86_avx2_pblendvb: {
920 // Convert blendv* to vector selects if the mask is constant.
921 // This optimization is convoluted because the intrinsic is defined as
922 // getting a vector of floats or doubles for the ps and pd versions.
923 // FIXME: That should be changed.
924 Value *Mask = II->getArgOperand(2);
925 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
926 auto Tyi1 = Builder->getInt1Ty();
927 auto SelectorType = cast<VectorType>(Mask->getType());
928 auto EltTy = SelectorType->getElementType();
929 unsigned Size = SelectorType->getNumElements();
933 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
934 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
935 "Wrong arguments for variable blend intrinsic");
936 SmallVector<Constant *, 32> Selectors;
937 for (unsigned I = 0; I < Size; ++I) {
938 // The intrinsics only read the top bit
941 Selector = C->getElementAsInteger(I);
943 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
944 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
946 auto NewSelector = ConstantVector::get(Selectors);
947 return SelectInst::Create(NewSelector, II->getArgOperand(1),
948 II->getArgOperand(0), "blendv");
954 case Intrinsic::x86_avx_vpermilvar_ps:
955 case Intrinsic::x86_avx_vpermilvar_ps_256:
956 case Intrinsic::x86_avx_vpermilvar_pd:
957 case Intrinsic::x86_avx_vpermilvar_pd_256: {
958 // Convert vpermil* to shufflevector if the mask is constant.
959 Value *V = II->getArgOperand(1);
960 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
961 assert(Size == 8 || Size == 4 || Size == 2);
963 if (auto C = dyn_cast<ConstantDataVector>(V)) {
964 // The intrinsics only read one or two bits, clear the rest.
965 for (unsigned I = 0; I < Size; ++I) {
966 uint32_t Index = C->getElementAsInteger(I) & 0x3;
967 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
968 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
972 } else if (isa<ConstantAggregateZero>(V)) {
973 for (unsigned I = 0; I < Size; ++I)
978 // The _256 variants are a bit trickier since the mask bits always index
979 // into the corresponding 128 half. In order to convert to a generic
980 // shuffle, we have to make that explicit.
981 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
982 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
983 for (unsigned I = Size / 2; I < Size; ++I)
984 Indexes[I] += Size / 2;
987 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
988 auto V1 = II->getArgOperand(0);
989 auto V2 = UndefValue::get(V1->getType());
990 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
991 return ReplaceInstUsesWith(CI, Shuffle);
994 case Intrinsic::x86_avx_vperm2f128_pd_256:
995 case Intrinsic::x86_avx_vperm2f128_ps_256:
996 case Intrinsic::x86_avx_vperm2f128_si_256:
997 case Intrinsic::x86_avx2_vperm2i128:
998 if (Value *V = SimplifyX86vperm2(*II, *Builder))
999 return ReplaceInstUsesWith(*II, V);
1002 case Intrinsic::ppc_altivec_vperm:
1003 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
1004 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
1005 // a vectorshuffle for little endian, we must undo the transformation
1006 // performed on vec_perm in altivec.h. That is, we must complement
1007 // the permutation mask with respect to 31 and reverse the order of
1009 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
1010 assert(Mask->getType()->getVectorNumElements() == 16 &&
1011 "Bad type for intrinsic!");
1013 // Check that all of the elements are integer constants or undefs.
1014 bool AllEltsOk = true;
1015 for (unsigned i = 0; i != 16; ++i) {
1016 Constant *Elt = Mask->getAggregateElement(i);
1017 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1024 // Cast the input vectors to byte vectors.
1025 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1027 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1029 Value *Result = UndefValue::get(Op0->getType());
1031 // Only extract each element once.
1032 Value *ExtractedElts[32];
1033 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1035 for (unsigned i = 0; i != 16; ++i) {
1036 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1039 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1040 Idx &= 31; // Match the hardware behavior.
1041 if (DL.isLittleEndian())
1044 if (!ExtractedElts[Idx]) {
1045 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1046 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1047 ExtractedElts[Idx] =
1048 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1049 Builder->getInt32(Idx&15));
1052 // Insert this value into the result vector.
1053 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1054 Builder->getInt32(i));
1056 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1061 case Intrinsic::arm_neon_vld1:
1062 case Intrinsic::arm_neon_vld2:
1063 case Intrinsic::arm_neon_vld3:
1064 case Intrinsic::arm_neon_vld4:
1065 case Intrinsic::arm_neon_vld2lane:
1066 case Intrinsic::arm_neon_vld3lane:
1067 case Intrinsic::arm_neon_vld4lane:
1068 case Intrinsic::arm_neon_vst1:
1069 case Intrinsic::arm_neon_vst2:
1070 case Intrinsic::arm_neon_vst3:
1071 case Intrinsic::arm_neon_vst4:
1072 case Intrinsic::arm_neon_vst2lane:
1073 case Intrinsic::arm_neon_vst3lane:
1074 case Intrinsic::arm_neon_vst4lane: {
1075 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1076 unsigned AlignArg = II->getNumArgOperands() - 1;
1077 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1078 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1079 II->setArgOperand(AlignArg,
1080 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1087 case Intrinsic::arm_neon_vmulls:
1088 case Intrinsic::arm_neon_vmullu:
1089 case Intrinsic::aarch64_neon_smull:
1090 case Intrinsic::aarch64_neon_umull: {
1091 Value *Arg0 = II->getArgOperand(0);
1092 Value *Arg1 = II->getArgOperand(1);
1094 // Handle mul by zero first:
1095 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1096 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1099 // Check for constant LHS & RHS - in this case we just simplify.
1100 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1101 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1102 VectorType *NewVT = cast<VectorType>(II->getType());
1103 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1104 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1105 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1106 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1108 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1111 // Couldn't simplify - canonicalize constant to the RHS.
1112 std::swap(Arg0, Arg1);
1115 // Handle mul by one:
1116 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1117 if (ConstantInt *Splat =
1118 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1120 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1121 /*isSigned=*/!Zext);
1126 case Intrinsic::AMDGPU_rcp: {
1127 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1128 const APFloat &ArgVal = C->getValueAPF();
1129 APFloat Val(ArgVal.getSemantics(), 1.0);
1130 APFloat::opStatus Status = Val.divide(ArgVal,
1131 APFloat::rmNearestTiesToEven);
1132 // Only do this if it was exact and therefore not dependent on the
1134 if (Status == APFloat::opOK)
1135 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1140 case Intrinsic::stackrestore: {
1141 // If the save is right next to the restore, remove the restore. This can
1142 // happen when variable allocas are DCE'd.
1143 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1144 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1145 BasicBlock::iterator BI = SS;
1147 return EraseInstFromFunction(CI);
1151 // Scan down this block to see if there is another stack restore in the
1152 // same block without an intervening call/alloca.
1153 BasicBlock::iterator BI = II;
1154 TerminatorInst *TI = II->getParent()->getTerminator();
1155 bool CannotRemove = false;
1156 for (++BI; &*BI != TI; ++BI) {
1157 if (isa<AllocaInst>(BI)) {
1158 CannotRemove = true;
1161 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1162 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1163 // If there is a stackrestore below this one, remove this one.
1164 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1165 return EraseInstFromFunction(CI);
1166 // Otherwise, ignore the intrinsic.
1168 // If we found a non-intrinsic call, we can't remove the stack
1170 CannotRemove = true;
1176 // If the stack restore is in a return, resume, or unwind block and if there
1177 // are no allocas or calls between the restore and the return, nuke the
1179 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1180 return EraseInstFromFunction(CI);
1183 case Intrinsic::assume: {
1184 // Canonicalize assume(a && b) -> assume(a); assume(b);
1185 // Note: New assumption intrinsics created here are registered by
1186 // the InstCombineIRInserter object.
1187 Value *IIOperand = II->getArgOperand(0), *A, *B,
1188 *AssumeIntrinsic = II->getCalledValue();
1189 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1190 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1191 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1192 return EraseInstFromFunction(*II);
1194 // assume(!(a || b)) -> assume(!a); assume(!b);
1195 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1196 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1198 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1200 return EraseInstFromFunction(*II);
1203 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1204 // (if assume is valid at the load)
1205 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1206 Value *LHS = ICmp->getOperand(0);
1207 Value *RHS = ICmp->getOperand(1);
1208 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1209 isa<LoadInst>(LHS) &&
1210 isa<Constant>(RHS) &&
1211 RHS->getType()->isPointerTy() &&
1212 cast<Constant>(RHS)->isNullValue()) {
1213 LoadInst* LI = cast<LoadInst>(LHS);
1214 if (isValidAssumeForContext(II, LI, DT)) {
1215 MDNode *MD = MDNode::get(II->getContext(), None);
1216 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1217 return EraseInstFromFunction(*II);
1220 // TODO: apply nonnull return attributes to calls and invokes
1221 // TODO: apply range metadata for range check patterns?
1223 // If there is a dominating assume with the same condition as this one,
1224 // then this one is redundant, and should be removed.
1225 APInt KnownZero(1, 0), KnownOne(1, 0);
1226 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1227 if (KnownOne.isAllOnesValue())
1228 return EraseInstFromFunction(*II);
1232 case Intrinsic::experimental_gc_relocate: {
1233 // Translate facts known about a pointer before relocating into
1234 // facts about the relocate value, while being careful to
1235 // preserve relocation semantics.
1236 GCRelocateOperands Operands(II);
1237 Value *DerivedPtr = Operands.getDerivedPtr();
1238 auto *GCRelocateType = cast<PointerType>(II->getType());
1240 // Remove the relocation if unused, note that this check is required
1241 // to prevent the cases below from looping forever.
1242 if (II->use_empty())
1243 return EraseInstFromFunction(*II);
1245 // Undef is undef, even after relocation.
1246 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1247 // most practical collectors, but there was discussion in the review thread
1248 // about whether it was legal for all possible collectors.
1249 if (isa<UndefValue>(DerivedPtr)) {
1250 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1251 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1254 // The relocation of null will be null for most any collector.
1255 // TODO: provide a hook for this in GCStrategy. There might be some weird
1256 // collector this property does not hold for.
1257 if (isa<ConstantPointerNull>(DerivedPtr)) {
1258 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1259 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1262 // isKnownNonNull -> nonnull attribute
1263 if (isKnownNonNull(DerivedPtr))
1264 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1266 // isDereferenceablePointer -> deref attribute
1267 if (isDereferenceablePointer(DerivedPtr, DL)) {
1268 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1269 uint64_t Bytes = A->getDereferenceableBytes();
1270 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1274 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1275 // Canonicalize on the type from the uses to the defs
1277 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1281 return visitCallSite(II);
1284 // InvokeInst simplification
1286 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1287 return visitCallSite(&II);
1290 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1291 /// passed through the varargs area, we can eliminate the use of the cast.
1292 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1293 const DataLayout &DL,
1294 const CastInst *const CI,
1296 if (!CI->isLosslessCast())
1299 // If this is a GC intrinsic, avoid munging types. We need types for
1300 // statepoint reconstruction in SelectionDAG.
1301 // TODO: This is probably something which should be expanded to all
1302 // intrinsics since the entire point of intrinsics is that
1303 // they are understandable by the optimizer.
1304 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1307 // The size of ByVal or InAlloca arguments is derived from the type, so we
1308 // can't change to a type with a different size. If the size were
1309 // passed explicitly we could avoid this check.
1310 if (!CS.isByValOrInAllocaArgument(ix))
1314 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1315 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1316 if (!SrcTy->isSized() || !DstTy->isSized())
1318 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1323 // Try to fold some different type of calls here.
1324 // Currently we're only working with the checking functions, memcpy_chk,
1325 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1326 // strcat_chk and strncat_chk.
1327 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1328 if (!CI->getCalledFunction()) return nullptr;
1330 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1331 ReplaceInstUsesWith(*From, With);
1333 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1334 if (Value *With = Simplifier.optimizeCall(CI)) {
1336 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1342 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1343 // Strip off at most one level of pointer casts, looking for an alloca. This
1344 // is good enough in practice and simpler than handling any number of casts.
1345 Value *Underlying = TrampMem->stripPointerCasts();
1346 if (Underlying != TrampMem &&
1347 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1349 if (!isa<AllocaInst>(Underlying))
1352 IntrinsicInst *InitTrampoline = nullptr;
1353 for (User *U : TrampMem->users()) {
1354 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1357 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1359 // More than one init_trampoline writes to this value. Give up.
1361 InitTrampoline = II;
1364 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1365 // Allow any number of calls to adjust.trampoline.
1370 // No call to init.trampoline found.
1371 if (!InitTrampoline)
1374 // Check that the alloca is being used in the expected way.
1375 if (InitTrampoline->getOperand(0) != TrampMem)
1378 return InitTrampoline;
1381 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1383 // Visit all the previous instructions in the basic block, and try to find a
1384 // init.trampoline which has a direct path to the adjust.trampoline.
1385 for (BasicBlock::iterator I = AdjustTramp,
1386 E = AdjustTramp->getParent()->begin(); I != E; ) {
1387 Instruction *Inst = --I;
1388 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1389 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1390 II->getOperand(0) == TrampMem)
1392 if (Inst->mayWriteToMemory())
1398 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1399 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1400 // to a direct call to a function. Otherwise return NULL.
1402 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1403 Callee = Callee->stripPointerCasts();
1404 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1406 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1409 Value *TrampMem = AdjustTramp->getOperand(0);
1411 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1413 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1418 // visitCallSite - Improvements for call and invoke instructions.
1420 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1422 if (isAllocLikeFn(CS.getInstruction(), TLI))
1423 return visitAllocSite(*CS.getInstruction());
1425 bool Changed = false;
1427 // Mark any parameters that are known to be non-null with the nonnull
1428 // attribute. This is helpful for inlining calls to functions with null
1429 // checks on their arguments.
1431 for (Value *V : CS.args()) {
1432 if (!CS.paramHasAttr(ArgNo+1, Attribute::NonNull) &&
1433 isKnownNonNull(V)) {
1434 AttributeSet AS = CS.getAttributes();
1435 AS = AS.addAttribute(CS.getInstruction()->getContext(), ArgNo+1,
1436 Attribute::NonNull);
1437 CS.setAttributes(AS);
1442 assert(ArgNo == CS.arg_size() && "sanity check");
1444 // If the callee is a pointer to a function, attempt to move any casts to the
1445 // arguments of the call/invoke.
1446 Value *Callee = CS.getCalledValue();
1447 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1450 if (Function *CalleeF = dyn_cast<Function>(Callee))
1451 // If the call and callee calling conventions don't match, this call must
1452 // be unreachable, as the call is undefined.
1453 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1454 // Only do this for calls to a function with a body. A prototype may
1455 // not actually end up matching the implementation's calling conv for a
1456 // variety of reasons (e.g. it may be written in assembly).
1457 !CalleeF->isDeclaration()) {
1458 Instruction *OldCall = CS.getInstruction();
1459 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1460 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1462 // If OldCall does not return void then replaceAllUsesWith undef.
1463 // This allows ValueHandlers and custom metadata to adjust itself.
1464 if (!OldCall->getType()->isVoidTy())
1465 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1466 if (isa<CallInst>(OldCall))
1467 return EraseInstFromFunction(*OldCall);
1469 // We cannot remove an invoke, because it would change the CFG, just
1470 // change the callee to a null pointer.
1471 cast<InvokeInst>(OldCall)->setCalledFunction(
1472 Constant::getNullValue(CalleeF->getType()));
1476 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1477 // If CS does not return void then replaceAllUsesWith undef.
1478 // This allows ValueHandlers and custom metadata to adjust itself.
1479 if (!CS.getInstruction()->getType()->isVoidTy())
1480 ReplaceInstUsesWith(*CS.getInstruction(),
1481 UndefValue::get(CS.getInstruction()->getType()));
1483 if (isa<InvokeInst>(CS.getInstruction())) {
1484 // Can't remove an invoke because we cannot change the CFG.
1488 // This instruction is not reachable, just remove it. We insert a store to
1489 // undef so that we know that this code is not reachable, despite the fact
1490 // that we can't modify the CFG here.
1491 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1492 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1493 CS.getInstruction());
1495 return EraseInstFromFunction(*CS.getInstruction());
1498 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1499 return transformCallThroughTrampoline(CS, II);
1501 PointerType *PTy = cast<PointerType>(Callee->getType());
1502 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1503 if (FTy->isVarArg()) {
1504 int ix = FTy->getNumParams();
1505 // See if we can optimize any arguments passed through the varargs area of
1507 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1508 E = CS.arg_end(); I != E; ++I, ++ix) {
1509 CastInst *CI = dyn_cast<CastInst>(*I);
1510 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1511 *I = CI->getOperand(0);
1517 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1518 // Inline asm calls cannot throw - mark them 'nounwind'.
1519 CS.setDoesNotThrow();
1523 // Try to optimize the call if possible, we require DataLayout for most of
1524 // this. None of these calls are seen as possibly dead so go ahead and
1525 // delete the instruction now.
1526 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1527 Instruction *I = tryOptimizeCall(CI);
1528 // If we changed something return the result, etc. Otherwise let
1529 // the fallthrough check.
1530 if (I) return EraseInstFromFunction(*I);
1533 return Changed ? CS.getInstruction() : nullptr;
1536 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1537 // attempt to move the cast to the arguments of the call/invoke.
1539 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1541 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1544 // The prototype of thunks are a lie, don't try to directly call such
1546 if (Callee->hasFnAttribute("thunk"))
1548 Instruction *Caller = CS.getInstruction();
1549 const AttributeSet &CallerPAL = CS.getAttributes();
1551 // Okay, this is a cast from a function to a different type. Unless doing so
1552 // would cause a type conversion of one of our arguments, change this call to
1553 // be a direct call with arguments casted to the appropriate types.
1555 FunctionType *FT = Callee->getFunctionType();
1556 Type *OldRetTy = Caller->getType();
1557 Type *NewRetTy = FT->getReturnType();
1559 // Check to see if we are changing the return type...
1560 if (OldRetTy != NewRetTy) {
1562 if (NewRetTy->isStructTy())
1563 return false; // TODO: Handle multiple return values.
1565 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1566 if (Callee->isDeclaration())
1567 return false; // Cannot transform this return value.
1569 if (!Caller->use_empty() &&
1570 // void -> non-void is handled specially
1571 !NewRetTy->isVoidTy())
1572 return false; // Cannot transform this return value.
1575 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1576 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1577 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1578 return false; // Attribute not compatible with transformed value.
1581 // If the callsite is an invoke instruction, and the return value is used by
1582 // a PHI node in a successor, we cannot change the return type of the call
1583 // because there is no place to put the cast instruction (without breaking
1584 // the critical edge). Bail out in this case.
1585 if (!Caller->use_empty())
1586 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1587 for (User *U : II->users())
1588 if (PHINode *PN = dyn_cast<PHINode>(U))
1589 if (PN->getParent() == II->getNormalDest() ||
1590 PN->getParent() == II->getUnwindDest())
1594 unsigned NumActualArgs = CS.arg_size();
1595 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1597 // Prevent us turning:
1598 // declare void @takes_i32_inalloca(i32* inalloca)
1599 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1602 // call void @takes_i32_inalloca(i32* null)
1604 // Similarly, avoid folding away bitcasts of byval calls.
1605 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1606 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1609 CallSite::arg_iterator AI = CS.arg_begin();
1610 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1611 Type *ParamTy = FT->getParamType(i);
1612 Type *ActTy = (*AI)->getType();
1614 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1615 return false; // Cannot transform this parameter value.
1617 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1618 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1619 return false; // Attribute not compatible with transformed value.
1621 if (CS.isInAllocaArgument(i))
1622 return false; // Cannot transform to and from inalloca.
1624 // If the parameter is passed as a byval argument, then we have to have a
1625 // sized type and the sized type has to have the same size as the old type.
1626 if (ParamTy != ActTy &&
1627 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1628 Attribute::ByVal)) {
1629 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1630 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1633 Type *CurElTy = ActTy->getPointerElementType();
1634 if (DL.getTypeAllocSize(CurElTy) !=
1635 DL.getTypeAllocSize(ParamPTy->getElementType()))
1640 if (Callee->isDeclaration()) {
1641 // Do not delete arguments unless we have a function body.
1642 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1645 // If the callee is just a declaration, don't change the varargsness of the
1646 // call. We don't want to introduce a varargs call where one doesn't
1648 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1649 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1652 // If both the callee and the cast type are varargs, we still have to make
1653 // sure the number of fixed parameters are the same or we have the same
1654 // ABI issues as if we introduce a varargs call.
1655 if (FT->isVarArg() &&
1656 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1657 FT->getNumParams() !=
1658 cast<FunctionType>(APTy->getElementType())->getNumParams())
1662 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1663 !CallerPAL.isEmpty())
1664 // In this case we have more arguments than the new function type, but we
1665 // won't be dropping them. Check that these extra arguments have attributes
1666 // that are compatible with being a vararg call argument.
1667 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1668 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1669 if (Index <= FT->getNumParams())
1672 // Check if it has an attribute that's incompatible with varargs.
1673 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1674 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1679 // Okay, we decided that this is a safe thing to do: go ahead and start
1680 // inserting cast instructions as necessary.
1681 std::vector<Value*> Args;
1682 Args.reserve(NumActualArgs);
1683 SmallVector<AttributeSet, 8> attrVec;
1684 attrVec.reserve(NumCommonArgs);
1686 // Get any return attributes.
1687 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1689 // If the return value is not being used, the type may not be compatible
1690 // with the existing attributes. Wipe out any problematic attributes.
1691 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1693 // Add the new return attributes.
1694 if (RAttrs.hasAttributes())
1695 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1696 AttributeSet::ReturnIndex, RAttrs));
1698 AI = CS.arg_begin();
1699 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1700 Type *ParamTy = FT->getParamType(i);
1702 if ((*AI)->getType() == ParamTy) {
1703 Args.push_back(*AI);
1705 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1708 // Add any parameter attributes.
1709 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1710 if (PAttrs.hasAttributes())
1711 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1715 // If the function takes more arguments than the call was taking, add them
1717 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1718 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1720 // If we are removing arguments to the function, emit an obnoxious warning.
1721 if (FT->getNumParams() < NumActualArgs) {
1722 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1723 if (FT->isVarArg()) {
1724 // Add all of the arguments in their promoted form to the arg list.
1725 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1726 Type *PTy = getPromotedType((*AI)->getType());
1727 if (PTy != (*AI)->getType()) {
1728 // Must promote to pass through va_arg area!
1729 Instruction::CastOps opcode =
1730 CastInst::getCastOpcode(*AI, false, PTy, false);
1731 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1733 Args.push_back(*AI);
1736 // Add any parameter attributes.
1737 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1738 if (PAttrs.hasAttributes())
1739 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1745 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1746 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1747 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1749 if (NewRetTy->isVoidTy())
1750 Caller->setName(""); // Void type should not have a name.
1752 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1756 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1757 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1758 II->getUnwindDest(), Args);
1760 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1761 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1763 CallInst *CI = cast<CallInst>(Caller);
1764 NC = Builder->CreateCall(Callee, Args);
1766 if (CI->isTailCall())
1767 cast<CallInst>(NC)->setTailCall();
1768 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1769 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1772 // Insert a cast of the return type as necessary.
1774 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1775 if (!NV->getType()->isVoidTy()) {
1776 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1777 NC->setDebugLoc(Caller->getDebugLoc());
1779 // If this is an invoke instruction, we should insert it after the first
1780 // non-phi, instruction in the normal successor block.
1781 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1782 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1783 InsertNewInstBefore(NC, *I);
1785 // Otherwise, it's a call, just insert cast right after the call.
1786 InsertNewInstBefore(NC, *Caller);
1788 Worklist.AddUsersToWorkList(*Caller);
1790 NV = UndefValue::get(Caller->getType());
1794 if (!Caller->use_empty())
1795 ReplaceInstUsesWith(*Caller, NV);
1796 else if (Caller->hasValueHandle()) {
1797 if (OldRetTy == NV->getType())
1798 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1800 // We cannot call ValueIsRAUWd with a different type, and the
1801 // actual tracked value will disappear.
1802 ValueHandleBase::ValueIsDeleted(Caller);
1805 EraseInstFromFunction(*Caller);
1809 // transformCallThroughTrampoline - Turn a call to a function created by
1810 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1811 // underlying function.
1814 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1815 IntrinsicInst *Tramp) {
1816 Value *Callee = CS.getCalledValue();
1817 PointerType *PTy = cast<PointerType>(Callee->getType());
1818 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1819 const AttributeSet &Attrs = CS.getAttributes();
1821 // If the call already has the 'nest' attribute somewhere then give up -
1822 // otherwise 'nest' would occur twice after splicing in the chain.
1823 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1827 "transformCallThroughTrampoline called with incorrect CallSite.");
1829 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1830 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1831 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1833 const AttributeSet &NestAttrs = NestF->getAttributes();
1834 if (!NestAttrs.isEmpty()) {
1835 unsigned NestIdx = 1;
1836 Type *NestTy = nullptr;
1837 AttributeSet NestAttr;
1839 // Look for a parameter marked with the 'nest' attribute.
1840 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1841 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1842 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1843 // Record the parameter type and any other attributes.
1845 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1850 Instruction *Caller = CS.getInstruction();
1851 std::vector<Value*> NewArgs;
1852 NewArgs.reserve(CS.arg_size() + 1);
1854 SmallVector<AttributeSet, 8> NewAttrs;
1855 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1857 // Insert the nest argument into the call argument list, which may
1858 // mean appending it. Likewise for attributes.
1860 // Add any result attributes.
1861 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1862 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1863 Attrs.getRetAttributes()));
1867 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1869 if (Idx == NestIdx) {
1870 // Add the chain argument and attributes.
1871 Value *NestVal = Tramp->getArgOperand(2);
1872 if (NestVal->getType() != NestTy)
1873 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1874 NewArgs.push_back(NestVal);
1875 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1882 // Add the original argument and attributes.
1883 NewArgs.push_back(*I);
1884 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1885 if (Attr.hasAttributes(Idx)) {
1886 AttrBuilder B(Attr, Idx);
1887 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1888 Idx + (Idx >= NestIdx), B));
1895 // Add any function attributes.
1896 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1897 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1898 Attrs.getFnAttributes()));
1900 // The trampoline may have been bitcast to a bogus type (FTy).
1901 // Handle this by synthesizing a new function type, equal to FTy
1902 // with the chain parameter inserted.
1904 std::vector<Type*> NewTypes;
1905 NewTypes.reserve(FTy->getNumParams()+1);
1907 // Insert the chain's type into the list of parameter types, which may
1908 // mean appending it.
1911 FunctionType::param_iterator I = FTy->param_begin(),
1912 E = FTy->param_end();
1916 // Add the chain's type.
1917 NewTypes.push_back(NestTy);
1922 // Add the original type.
1923 NewTypes.push_back(*I);
1929 // Replace the trampoline call with a direct call. Let the generic
1930 // code sort out any function type mismatches.
1931 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1933 Constant *NewCallee =
1934 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1935 NestF : ConstantExpr::getBitCast(NestF,
1936 PointerType::getUnqual(NewFTy));
1937 const AttributeSet &NewPAL =
1938 AttributeSet::get(FTy->getContext(), NewAttrs);
1940 Instruction *NewCaller;
1941 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1942 NewCaller = InvokeInst::Create(NewCallee,
1943 II->getNormalDest(), II->getUnwindDest(),
1945 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1946 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1948 NewCaller = CallInst::Create(NewCallee, NewArgs);
1949 if (cast<CallInst>(Caller)->isTailCall())
1950 cast<CallInst>(NewCaller)->setTailCall();
1951 cast<CallInst>(NewCaller)->
1952 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1953 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1960 // Replace the trampoline call with a direct call. Since there is no 'nest'
1961 // parameter, there is no need to adjust the argument list. Let the generic
1962 // code sort out any function type mismatches.
1963 Constant *NewCallee =
1964 NestF->getType() == PTy ? NestF :
1965 ConstantExpr::getBitCast(NestF, PTy);
1966 CS.setCalledFunction(NewCallee);
1967 return CS.getInstruction();