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(),
75 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
77 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
78 if (!MemOpLength) return nullptr;
80 // Source and destination pointer types are always "i8*" for intrinsic. See
81 // if the size is something we can handle with a single primitive load/store.
82 // A single load+store correctly handles overlapping memory in the memmove
84 uint64_t Size = MemOpLength->getLimitedValue();
85 assert(Size && "0-sized memory transferring should be removed already.");
87 if (Size > 8 || (Size&(Size-1)))
88 return nullptr; // If not 1/2/4/8 bytes, exit.
90 // Use an integer load+store unless we can find something better.
92 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
94 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
96 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
97 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
98 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
100 // Memcpy forces the use of i8* for the source and destination. That means
101 // that if you're using memcpy to move one double around, you'll get a cast
102 // from double* to i8*. We'd much rather use a double load+store rather than
103 // an i64 load+store, here because this improves the odds that the source or
104 // dest address will be promotable. See if we can find a better type than the
106 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
107 MDNode *CopyMD = nullptr;
108 if (StrippedDest != MI->getArgOperand(0)) {
109 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
111 if (SrcETy->isSized() && DL.getTypeStoreSize(SrcETy) == Size) {
112 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
113 // down through these levels if so.
114 SrcETy = reduceToSingleValueType(SrcETy);
116 if (SrcETy->isSingleValueType()) {
117 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
118 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
120 // If the memcpy has metadata describing the members, see if we can
121 // get the TBAA tag describing our copy.
122 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
123 if (M->getNumOperands() == 3 && M->getOperand(0) &&
124 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
125 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() &&
127 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
128 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
130 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
131 CopyMD = cast<MDNode>(M->getOperand(2));
137 // If the memcpy/memmove provides better alignment info than we can
139 SrcAlign = std::max(SrcAlign, CopyAlign);
140 DstAlign = std::max(DstAlign, CopyAlign);
142 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
143 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
144 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
145 L->setAlignment(SrcAlign);
147 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
148 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
149 S->setAlignment(DstAlign);
151 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
153 // Set the size of the copy to 0, it will be deleted on the next iteration.
154 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
158 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
159 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, AC, DT);
160 if (MI->getAlignment() < Alignment) {
161 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
166 // Extract the length and alignment and fill if they are constant.
167 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
168 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
169 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
171 uint64_t Len = LenC->getLimitedValue();
172 Alignment = MI->getAlignment();
173 assert(Len && "0-sized memory setting should be removed already.");
175 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
176 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
177 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
179 Value *Dest = MI->getDest();
180 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
181 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
182 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
184 // Alignment 0 is identity for alignment 1 for memset, but not store.
185 if (Alignment == 0) Alignment = 1;
187 // Extract the fill value and store.
188 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
189 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
191 S->setAlignment(Alignment);
193 // Set the size of the copy to 0, it will be deleted on the next iteration.
194 MI->setLength(Constant::getNullValue(LenC->getType()));
201 static Value *SimplifyX86insertps(const IntrinsicInst &II,
202 InstCombiner::BuilderTy &Builder) {
203 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
204 VectorType *VecTy = cast<VectorType>(II.getType());
205 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
207 // The immediate permute control byte looks like this:
208 // [3:0] - zero mask for each 32-bit lane
209 // [5:4] - select one 32-bit destination lane
210 // [7:6] - select one 32-bit source lane
212 uint8_t Imm = CInt->getZExtValue();
213 uint8_t ZMask = Imm & 0xf;
214 uint8_t DestLane = (Imm >> 4) & 0x3;
215 uint8_t SourceLane = (Imm >> 6) & 0x3;
217 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
219 // If all zero mask bits are set, this was just a weird way to
220 // generate a zero vector.
224 // Initialize by passing all of the first source bits through.
225 int ShuffleMask[4] = { 0, 1, 2, 3 };
227 // We may replace the second operand with the zero vector.
228 Value *V1 = II.getArgOperand(1);
231 // If the zero mask is being used with a single input or the zero mask
232 // overrides the destination lane, this is a shuffle with the zero vector.
233 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
234 (ZMask & (1 << DestLane))) {
236 // We may still move 32-bits of the first source vector from one lane
238 ShuffleMask[DestLane] = SourceLane;
239 // The zero mask may override the previous insert operation.
240 for (unsigned i = 0; i < 4; ++i)
241 if ((ZMask >> i) & 0x1)
242 ShuffleMask[i] = i + 4;
244 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
248 // Replace the selected destination lane with the selected source lane.
249 ShuffleMask[DestLane] = SourceLane + 4;
252 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
257 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
258 /// source vectors, unless a zero bit is set. If a zero bit is set,
259 /// then ignore that half of the mask and clear that half of the vector.
260 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
261 InstCombiner::BuilderTy &Builder) {
262 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
263 VectorType *VecTy = cast<VectorType>(II.getType());
264 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
266 // The immediate permute control byte looks like this:
267 // [1:0] - select 128 bits from sources for low half of destination
269 // [3] - zero low half of destination
270 // [5:4] - select 128 bits from sources for high half of destination
272 // [7] - zero high half of destination
274 uint8_t Imm = CInt->getZExtValue();
276 bool LowHalfZero = Imm & 0x08;
277 bool HighHalfZero = Imm & 0x80;
279 // If both zero mask bits are set, this was just a weird way to
280 // generate a zero vector.
281 if (LowHalfZero && HighHalfZero)
284 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
285 unsigned NumElts = VecTy->getNumElements();
286 unsigned HalfSize = NumElts / 2;
287 SmallVector<int, 8> ShuffleMask(NumElts);
289 // The high bit of the selection field chooses the 1st or 2nd operand.
290 bool LowInputSelect = Imm & 0x02;
291 bool HighInputSelect = Imm & 0x20;
293 // The low bit of the selection field chooses the low or high half
294 // of the selected operand.
295 bool LowHalfSelect = Imm & 0x01;
296 bool HighHalfSelect = Imm & 0x10;
298 // Determine which operand(s) are actually in use for this instruction.
299 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
300 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
302 // If needed, replace operands based on zero mask.
303 V0 = LowHalfZero ? ZeroVector : V0;
304 V1 = HighHalfZero ? ZeroVector : V1;
306 // Permute low half of result.
307 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
308 for (unsigned i = 0; i < HalfSize; ++i)
309 ShuffleMask[i] = StartIndex + i;
311 // Permute high half of result.
312 StartIndex = HighHalfSelect ? HalfSize : 0;
313 StartIndex += NumElts;
314 for (unsigned i = 0; i < HalfSize; ++i)
315 ShuffleMask[i + HalfSize] = StartIndex + i;
317 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
322 /// visitCallInst - CallInst simplification. This mostly only handles folding
323 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
324 /// the heavy lifting.
326 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
327 auto Args = CI.arg_operands();
328 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL,
330 return ReplaceInstUsesWith(CI, V);
332 if (isFreeCall(&CI, TLI))
333 return visitFree(CI);
335 // If the caller function is nounwind, mark the call as nounwind, even if the
337 if (CI.getParent()->getParent()->doesNotThrow() &&
338 !CI.doesNotThrow()) {
339 CI.setDoesNotThrow();
343 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
344 if (!II) return visitCallSite(&CI);
346 // Intrinsics cannot occur in an invoke, so handle them here instead of in
348 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
349 bool Changed = false;
351 // memmove/cpy/set of zero bytes is a noop.
352 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
353 if (NumBytes->isNullValue())
354 return EraseInstFromFunction(CI);
356 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
357 if (CI->getZExtValue() == 1) {
358 // Replace the instruction with just byte operations. We would
359 // transform other cases to loads/stores, but we don't know if
360 // alignment is sufficient.
364 // No other transformations apply to volatile transfers.
365 if (MI->isVolatile())
368 // If we have a memmove and the source operation is a constant global,
369 // then the source and dest pointers can't alias, so we can change this
370 // into a call to memcpy.
371 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
372 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
373 if (GVSrc->isConstant()) {
374 Module *M = CI.getParent()->getParent()->getParent();
375 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
376 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
377 CI.getArgOperand(1)->getType(),
378 CI.getArgOperand(2)->getType() };
379 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
384 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
385 // memmove(x,x,size) -> noop.
386 if (MTI->getSource() == MTI->getDest())
387 return EraseInstFromFunction(CI);
390 // If we can determine a pointer alignment that is bigger than currently
391 // set, update the alignment.
392 if (isa<MemTransferInst>(MI)) {
393 if (Instruction *I = SimplifyMemTransfer(MI))
395 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
396 if (Instruction *I = SimplifyMemSet(MSI))
400 if (Changed) return II;
403 switch (II->getIntrinsicID()) {
405 case Intrinsic::objectsize: {
407 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
408 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
411 case Intrinsic::bswap: {
412 Value *IIOperand = II->getArgOperand(0);
415 // bswap(bswap(x)) -> x
416 if (match(IIOperand, m_BSwap(m_Value(X))))
417 return ReplaceInstUsesWith(CI, X);
419 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
420 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
421 unsigned C = X->getType()->getPrimitiveSizeInBits() -
422 IIOperand->getType()->getPrimitiveSizeInBits();
423 Value *CV = ConstantInt::get(X->getType(), C);
424 Value *V = Builder->CreateLShr(X, CV);
425 return new TruncInst(V, IIOperand->getType());
430 case Intrinsic::powi:
431 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
434 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
437 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
438 // powi(x, -1) -> 1/x
439 if (Power->isAllOnesValue())
440 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
441 II->getArgOperand(0));
444 case Intrinsic::cttz: {
445 // If all bits below the first known one are known zero,
446 // this value is constant.
447 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
448 // FIXME: Try to simplify vectors of integers.
450 uint32_t BitWidth = IT->getBitWidth();
451 APInt KnownZero(BitWidth, 0);
452 APInt KnownOne(BitWidth, 0);
453 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
454 unsigned TrailingZeros = KnownOne.countTrailingZeros();
455 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
456 if ((Mask & KnownZero) == Mask)
457 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
458 APInt(BitWidth, TrailingZeros)));
462 case Intrinsic::ctlz: {
463 // If all bits above the first known one are known zero,
464 // this value is constant.
465 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
466 // FIXME: Try to simplify vectors of integers.
468 uint32_t BitWidth = IT->getBitWidth();
469 APInt KnownZero(BitWidth, 0);
470 APInt KnownOne(BitWidth, 0);
471 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
472 unsigned LeadingZeros = KnownOne.countLeadingZeros();
473 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
474 if ((Mask & KnownZero) == Mask)
475 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
476 APInt(BitWidth, LeadingZeros)));
481 case Intrinsic::uadd_with_overflow:
482 case Intrinsic::sadd_with_overflow:
483 case Intrinsic::umul_with_overflow:
484 case Intrinsic::smul_with_overflow:
485 if (isa<Constant>(II->getArgOperand(0)) &&
486 !isa<Constant>(II->getArgOperand(1))) {
487 // Canonicalize constants into the RHS.
488 Value *LHS = II->getArgOperand(0);
489 II->setArgOperand(0, II->getArgOperand(1));
490 II->setArgOperand(1, LHS);
495 case Intrinsic::usub_with_overflow:
496 case Intrinsic::ssub_with_overflow: {
497 OverflowCheckFlavor OCF =
498 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
499 assert(OCF != OCF_INVALID && "unexpected!");
501 Value *OperationResult = nullptr;
502 Constant *OverflowResult = nullptr;
503 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
504 *II, OperationResult, OverflowResult))
505 return CreateOverflowTuple(II, OperationResult, OverflowResult);
510 case Intrinsic::minnum:
511 case Intrinsic::maxnum: {
512 Value *Arg0 = II->getArgOperand(0);
513 Value *Arg1 = II->getArgOperand(1);
517 return ReplaceInstUsesWith(CI, Arg0);
519 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
520 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
522 // Canonicalize constants into the RHS.
524 II->setArgOperand(0, Arg1);
525 II->setArgOperand(1, Arg0);
530 if (C1 && C1->isNaN())
531 return ReplaceInstUsesWith(CI, Arg0);
533 // This is the value because if undef were NaN, we would return the other
534 // value and cannot return a NaN unless both operands are.
536 // fmin(undef, x) -> x
537 if (isa<UndefValue>(Arg0))
538 return ReplaceInstUsesWith(CI, Arg1);
540 // fmin(x, undef) -> x
541 if (isa<UndefValue>(Arg1))
542 return ReplaceInstUsesWith(CI, Arg0);
546 if (II->getIntrinsicID() == Intrinsic::minnum) {
547 // fmin(x, fmin(x, y)) -> fmin(x, y)
548 // fmin(y, fmin(x, y)) -> fmin(x, y)
549 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
550 if (Arg0 == X || Arg0 == Y)
551 return ReplaceInstUsesWith(CI, Arg1);
554 // fmin(fmin(x, y), x) -> fmin(x, y)
555 // fmin(fmin(x, y), y) -> fmin(x, y)
556 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
557 if (Arg1 == X || Arg1 == Y)
558 return ReplaceInstUsesWith(CI, Arg0);
561 // TODO: fmin(nnan x, inf) -> x
562 // TODO: fmin(nnan ninf x, flt_max) -> x
563 if (C1 && C1->isInfinity()) {
564 // fmin(x, -inf) -> -inf
565 if (C1->isNegative())
566 return ReplaceInstUsesWith(CI, Arg1);
569 assert(II->getIntrinsicID() == Intrinsic::maxnum);
570 // fmax(x, fmax(x, y)) -> fmax(x, y)
571 // fmax(y, fmax(x, y)) -> fmax(x, y)
572 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
573 if (Arg0 == X || Arg0 == Y)
574 return ReplaceInstUsesWith(CI, Arg1);
577 // fmax(fmax(x, y), x) -> fmax(x, y)
578 // fmax(fmax(x, y), y) -> fmax(x, y)
579 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
580 if (Arg1 == X || Arg1 == Y)
581 return ReplaceInstUsesWith(CI, Arg0);
584 // TODO: fmax(nnan x, -inf) -> x
585 // TODO: fmax(nnan ninf x, -flt_max) -> x
586 if (C1 && C1->isInfinity()) {
587 // fmax(x, inf) -> inf
588 if (!C1->isNegative())
589 return ReplaceInstUsesWith(CI, Arg1);
594 case Intrinsic::ppc_altivec_lvx:
595 case Intrinsic::ppc_altivec_lvxl:
596 // Turn PPC lvx -> load if the pointer is known aligned.
597 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
599 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
600 PointerType::getUnqual(II->getType()));
601 return new LoadInst(Ptr);
604 case Intrinsic::ppc_vsx_lxvw4x:
605 case Intrinsic::ppc_vsx_lxvd2x: {
606 // Turn PPC VSX loads into normal loads.
607 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
608 PointerType::getUnqual(II->getType()));
609 return new LoadInst(Ptr, Twine(""), false, 1);
611 case Intrinsic::ppc_altivec_stvx:
612 case Intrinsic::ppc_altivec_stvxl:
613 // Turn stvx -> store if the pointer is known aligned.
614 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
617 PointerType::getUnqual(II->getArgOperand(0)->getType());
618 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
619 return new StoreInst(II->getArgOperand(0), Ptr);
622 case Intrinsic::ppc_vsx_stxvw4x:
623 case Intrinsic::ppc_vsx_stxvd2x: {
624 // Turn PPC VSX stores into normal stores.
625 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
626 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
627 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
629 case Intrinsic::ppc_qpx_qvlfs:
630 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
631 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
633 Type *VTy = VectorType::get(Builder->getFloatTy(),
634 II->getType()->getVectorNumElements());
635 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
636 PointerType::getUnqual(VTy));
637 Value *Load = Builder->CreateLoad(Ptr);
638 return new FPExtInst(Load, II->getType());
641 case Intrinsic::ppc_qpx_qvlfd:
642 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
643 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
645 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
646 PointerType::getUnqual(II->getType()));
647 return new LoadInst(Ptr);
650 case Intrinsic::ppc_qpx_qvstfs:
651 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
652 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
654 Type *VTy = VectorType::get(Builder->getFloatTy(),
655 II->getArgOperand(0)->getType()->getVectorNumElements());
656 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
657 Type *OpPtrTy = PointerType::getUnqual(VTy);
658 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
659 return new StoreInst(TOp, Ptr);
662 case Intrinsic::ppc_qpx_qvstfd:
663 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
664 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
667 PointerType::getUnqual(II->getArgOperand(0)->getType());
668 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
669 return new StoreInst(II->getArgOperand(0), Ptr);
672 case Intrinsic::x86_sse_storeu_ps:
673 case Intrinsic::x86_sse2_storeu_pd:
674 case Intrinsic::x86_sse2_storeu_dq:
675 // Turn X86 storeu -> store if the pointer is known aligned.
676 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
679 PointerType::getUnqual(II->getArgOperand(1)->getType());
680 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
681 return new StoreInst(II->getArgOperand(1), Ptr);
685 case Intrinsic::x86_sse_cvtss2si:
686 case Intrinsic::x86_sse_cvtss2si64:
687 case Intrinsic::x86_sse_cvttss2si:
688 case Intrinsic::x86_sse_cvttss2si64:
689 case Intrinsic::x86_sse2_cvtsd2si:
690 case Intrinsic::x86_sse2_cvtsd2si64:
691 case Intrinsic::x86_sse2_cvttsd2si:
692 case Intrinsic::x86_sse2_cvttsd2si64: {
693 // These intrinsics only demand the 0th element of their input vectors. If
694 // we can simplify the input based on that, do so now.
696 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
697 APInt DemandedElts(VWidth, 1);
698 APInt UndefElts(VWidth, 0);
699 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
700 DemandedElts, UndefElts)) {
701 II->setArgOperand(0, V);
707 // Constant fold <A x Bi> << Ci.
708 // FIXME: We don't handle _dq because it's a shift of an i128, but is
709 // represented in the IR as <2 x i64>. A per element shift is wrong.
710 case Intrinsic::x86_sse2_psll_d:
711 case Intrinsic::x86_sse2_psll_q:
712 case Intrinsic::x86_sse2_psll_w:
713 case Intrinsic::x86_sse2_pslli_d:
714 case Intrinsic::x86_sse2_pslli_q:
715 case Intrinsic::x86_sse2_pslli_w:
716 case Intrinsic::x86_avx2_psll_d:
717 case Intrinsic::x86_avx2_psll_q:
718 case Intrinsic::x86_avx2_psll_w:
719 case Intrinsic::x86_avx2_pslli_d:
720 case Intrinsic::x86_avx2_pslli_q:
721 case Intrinsic::x86_avx2_pslli_w:
722 case Intrinsic::x86_sse2_psrl_d:
723 case Intrinsic::x86_sse2_psrl_q:
724 case Intrinsic::x86_sse2_psrl_w:
725 case Intrinsic::x86_sse2_psrli_d:
726 case Intrinsic::x86_sse2_psrli_q:
727 case Intrinsic::x86_sse2_psrli_w:
728 case Intrinsic::x86_avx2_psrl_d:
729 case Intrinsic::x86_avx2_psrl_q:
730 case Intrinsic::x86_avx2_psrl_w:
731 case Intrinsic::x86_avx2_psrli_d:
732 case Intrinsic::x86_avx2_psrli_q:
733 case Intrinsic::x86_avx2_psrli_w: {
734 // Simplify if count is constant. To 0 if >= BitWidth,
735 // otherwise to shl/lshr.
736 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
737 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
742 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
746 auto Vec = II->getArgOperand(0);
747 auto VT = cast<VectorType>(Vec->getType());
748 if (Count->getZExtValue() >
749 VT->getElementType()->getPrimitiveSizeInBits() - 1)
750 return ReplaceInstUsesWith(
751 CI, ConstantAggregateZero::get(Vec->getType()));
753 bool isPackedShiftLeft = true;
754 switch (II->getIntrinsicID()) {
756 case Intrinsic::x86_sse2_psrl_d:
757 case Intrinsic::x86_sse2_psrl_q:
758 case Intrinsic::x86_sse2_psrl_w:
759 case Intrinsic::x86_sse2_psrli_d:
760 case Intrinsic::x86_sse2_psrli_q:
761 case Intrinsic::x86_sse2_psrli_w:
762 case Intrinsic::x86_avx2_psrl_d:
763 case Intrinsic::x86_avx2_psrl_q:
764 case Intrinsic::x86_avx2_psrl_w:
765 case Intrinsic::x86_avx2_psrli_d:
766 case Intrinsic::x86_avx2_psrli_q:
767 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
770 unsigned VWidth = VT->getNumElements();
771 // Get a constant vector of the same type as the first operand.
772 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
773 if (isPackedShiftLeft)
774 return BinaryOperator::CreateShl(Vec,
775 Builder->CreateVectorSplat(VWidth, VTCI));
777 return BinaryOperator::CreateLShr(Vec,
778 Builder->CreateVectorSplat(VWidth, VTCI));
781 case Intrinsic::x86_sse41_pmovsxbw:
782 case Intrinsic::x86_sse41_pmovsxwd:
783 case Intrinsic::x86_sse41_pmovsxdq:
784 case Intrinsic::x86_sse41_pmovzxbw:
785 case Intrinsic::x86_sse41_pmovzxwd:
786 case Intrinsic::x86_sse41_pmovzxdq: {
787 // pmov{s|z}x ignores the upper half of their input vectors.
789 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
790 unsigned LowHalfElts = VWidth / 2;
791 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
792 APInt UndefElts(VWidth, 0);
793 if (Value *TmpV = SimplifyDemandedVectorElts(
794 II->getArgOperand(0), InputDemandedElts, UndefElts)) {
795 II->setArgOperand(0, TmpV);
800 case Intrinsic::x86_sse41_insertps:
801 if (Value *V = SimplifyX86insertps(*II, *Builder))
802 return ReplaceInstUsesWith(*II, V);
805 case Intrinsic::x86_sse4a_insertqi: {
806 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
808 // TODO: eventually we should lower this intrinsic to IR
809 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
810 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
811 unsigned Index = CIStart->getZExtValue();
812 // From AMD documentation: "a value of zero in the field length is
813 // defined as length of 64".
814 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
816 // From AMD documentation: "If the sum of the bit index + length field
817 // is greater than 64, the results are undefined".
819 // Note that both field index and field length are 8-bit quantities.
820 // Since variables 'Index' and 'Length' are unsigned values
821 // obtained from zero-extending field index and field length
822 // respectively, their sum should never wrap around.
823 if ((Index + Length) > 64)
824 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
826 if (Length == 64 && Index == 0) {
827 Value *Vec = II->getArgOperand(1);
828 Value *Undef = UndefValue::get(Vec->getType());
829 const uint32_t Mask[] = { 0, 2 };
830 return ReplaceInstUsesWith(
832 Builder->CreateShuffleVector(
833 Vec, Undef, ConstantDataVector::get(
834 II->getContext(), makeArrayRef(Mask))));
836 } else if (auto Source =
837 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
838 if (Source->hasOneUse() &&
839 Source->getArgOperand(1) == II->getArgOperand(1)) {
840 // If the source of the insert has only one use and it's another
841 // insert (and they're both inserting from the same vector), try to
842 // bundle both together.
844 dyn_cast<ConstantInt>(Source->getArgOperand(2));
846 dyn_cast<ConstantInt>(Source->getArgOperand(3));
847 if (CISourceStart && CISourceWidth) {
848 unsigned Start = CIStart->getZExtValue();
849 unsigned Width = CIWidth->getZExtValue();
850 unsigned End = Start + Width;
851 unsigned SourceStart = CISourceStart->getZExtValue();
852 unsigned SourceWidth = CISourceWidth->getZExtValue();
853 unsigned SourceEnd = SourceStart + SourceWidth;
854 unsigned NewStart, NewWidth;
855 bool ShouldReplace = false;
856 if (Start <= SourceStart && SourceStart <= End) {
858 NewWidth = std::max(End, SourceEnd) - NewStart;
859 ShouldReplace = true;
860 } else if (SourceStart <= Start && Start <= SourceEnd) {
861 NewStart = SourceStart;
862 NewWidth = std::max(SourceEnd, End) - NewStart;
863 ShouldReplace = true;
867 Constant *ConstantWidth = ConstantInt::get(
868 II->getArgOperand(2)->getType(), NewWidth, false);
869 Constant *ConstantStart = ConstantInt::get(
870 II->getArgOperand(3)->getType(), NewStart, false);
871 Value *Args[4] = { Source->getArgOperand(0),
872 II->getArgOperand(1), ConstantWidth,
874 Module *M = CI.getParent()->getParent()->getParent();
876 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
877 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
887 case Intrinsic::x86_sse41_pblendvb:
888 case Intrinsic::x86_sse41_blendvps:
889 case Intrinsic::x86_sse41_blendvpd:
890 case Intrinsic::x86_avx_blendv_ps_256:
891 case Intrinsic::x86_avx_blendv_pd_256:
892 case Intrinsic::x86_avx2_pblendvb: {
893 // Convert blendv* to vector selects if the mask is constant.
894 // This optimization is convoluted because the intrinsic is defined as
895 // getting a vector of floats or doubles for the ps and pd versions.
896 // FIXME: That should be changed.
897 Value *Mask = II->getArgOperand(2);
898 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
899 auto Tyi1 = Builder->getInt1Ty();
900 auto SelectorType = cast<VectorType>(Mask->getType());
901 auto EltTy = SelectorType->getElementType();
902 unsigned Size = SelectorType->getNumElements();
906 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
907 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
908 "Wrong arguments for variable blend intrinsic");
909 SmallVector<Constant *, 32> Selectors;
910 for (unsigned I = 0; I < Size; ++I) {
911 // The intrinsics only read the top bit
914 Selector = C->getElementAsInteger(I);
916 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
917 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
919 auto NewSelector = ConstantVector::get(Selectors);
920 return SelectInst::Create(NewSelector, II->getArgOperand(1),
921 II->getArgOperand(0), "blendv");
927 case Intrinsic::x86_avx_vpermilvar_ps:
928 case Intrinsic::x86_avx_vpermilvar_ps_256:
929 case Intrinsic::x86_avx_vpermilvar_pd:
930 case Intrinsic::x86_avx_vpermilvar_pd_256: {
931 // Convert vpermil* to shufflevector if the mask is constant.
932 Value *V = II->getArgOperand(1);
933 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
934 assert(Size == 8 || Size == 4 || Size == 2);
936 if (auto C = dyn_cast<ConstantDataVector>(V)) {
937 // The intrinsics only read one or two bits, clear the rest.
938 for (unsigned I = 0; I < Size; ++I) {
939 uint32_t Index = C->getElementAsInteger(I) & 0x3;
940 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
941 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
945 } else if (isa<ConstantAggregateZero>(V)) {
946 for (unsigned I = 0; I < Size; ++I)
951 // The _256 variants are a bit trickier since the mask bits always index
952 // into the corresponding 128 half. In order to convert to a generic
953 // shuffle, we have to make that explicit.
954 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
955 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
956 for (unsigned I = Size / 2; I < Size; ++I)
957 Indexes[I] += Size / 2;
960 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
961 auto V1 = II->getArgOperand(0);
962 auto V2 = UndefValue::get(V1->getType());
963 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
964 return ReplaceInstUsesWith(CI, Shuffle);
967 case Intrinsic::x86_avx_vperm2f128_pd_256:
968 case Intrinsic::x86_avx_vperm2f128_ps_256:
969 case Intrinsic::x86_avx_vperm2f128_si_256:
970 case Intrinsic::x86_avx2_vperm2i128:
971 if (Value *V = SimplifyX86vperm2(*II, *Builder))
972 return ReplaceInstUsesWith(*II, V);
975 case Intrinsic::ppc_altivec_vperm:
976 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
977 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
978 // a vectorshuffle for little endian, we must undo the transformation
979 // performed on vec_perm in altivec.h. That is, we must complement
980 // the permutation mask with respect to 31 and reverse the order of
982 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
983 assert(Mask->getType()->getVectorNumElements() == 16 &&
984 "Bad type for intrinsic!");
986 // Check that all of the elements are integer constants or undefs.
987 bool AllEltsOk = true;
988 for (unsigned i = 0; i != 16; ++i) {
989 Constant *Elt = Mask->getAggregateElement(i);
990 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
997 // Cast the input vectors to byte vectors.
998 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1000 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1002 Value *Result = UndefValue::get(Op0->getType());
1004 // Only extract each element once.
1005 Value *ExtractedElts[32];
1006 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1008 for (unsigned i = 0; i != 16; ++i) {
1009 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1012 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1013 Idx &= 31; // Match the hardware behavior.
1014 if (DL.isLittleEndian())
1017 if (!ExtractedElts[Idx]) {
1018 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1019 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1020 ExtractedElts[Idx] =
1021 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1022 Builder->getInt32(Idx&15));
1025 // Insert this value into the result vector.
1026 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1027 Builder->getInt32(i));
1029 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1034 case Intrinsic::arm_neon_vld1:
1035 case Intrinsic::arm_neon_vld2:
1036 case Intrinsic::arm_neon_vld3:
1037 case Intrinsic::arm_neon_vld4:
1038 case Intrinsic::arm_neon_vld2lane:
1039 case Intrinsic::arm_neon_vld3lane:
1040 case Intrinsic::arm_neon_vld4lane:
1041 case Intrinsic::arm_neon_vst1:
1042 case Intrinsic::arm_neon_vst2:
1043 case Intrinsic::arm_neon_vst3:
1044 case Intrinsic::arm_neon_vst4:
1045 case Intrinsic::arm_neon_vst2lane:
1046 case Intrinsic::arm_neon_vst3lane:
1047 case Intrinsic::arm_neon_vst4lane: {
1048 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1049 unsigned AlignArg = II->getNumArgOperands() - 1;
1050 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1051 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1052 II->setArgOperand(AlignArg,
1053 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1060 case Intrinsic::arm_neon_vmulls:
1061 case Intrinsic::arm_neon_vmullu:
1062 case Intrinsic::aarch64_neon_smull:
1063 case Intrinsic::aarch64_neon_umull: {
1064 Value *Arg0 = II->getArgOperand(0);
1065 Value *Arg1 = II->getArgOperand(1);
1067 // Handle mul by zero first:
1068 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1069 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1072 // Check for constant LHS & RHS - in this case we just simplify.
1073 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1074 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1075 VectorType *NewVT = cast<VectorType>(II->getType());
1076 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1077 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1078 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1079 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1081 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1084 // Couldn't simplify - canonicalize constant to the RHS.
1085 std::swap(Arg0, Arg1);
1088 // Handle mul by one:
1089 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1090 if (ConstantInt *Splat =
1091 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1093 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1094 /*isSigned=*/!Zext);
1099 case Intrinsic::AMDGPU_rcp: {
1100 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1101 const APFloat &ArgVal = C->getValueAPF();
1102 APFloat Val(ArgVal.getSemantics(), 1.0);
1103 APFloat::opStatus Status = Val.divide(ArgVal,
1104 APFloat::rmNearestTiesToEven);
1105 // Only do this if it was exact and therefore not dependent on the
1107 if (Status == APFloat::opOK)
1108 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1113 case Intrinsic::stackrestore: {
1114 // If the save is right next to the restore, remove the restore. This can
1115 // happen when variable allocas are DCE'd.
1116 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1117 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1118 BasicBlock::iterator BI = SS;
1120 return EraseInstFromFunction(CI);
1124 // Scan down this block to see if there is another stack restore in the
1125 // same block without an intervening call/alloca.
1126 BasicBlock::iterator BI = II;
1127 TerminatorInst *TI = II->getParent()->getTerminator();
1128 bool CannotRemove = false;
1129 for (++BI; &*BI != TI; ++BI) {
1130 if (isa<AllocaInst>(BI)) {
1131 CannotRemove = true;
1134 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1135 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1136 // If there is a stackrestore below this one, remove this one.
1137 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1138 return EraseInstFromFunction(CI);
1139 // Otherwise, ignore the intrinsic.
1141 // If we found a non-intrinsic call, we can't remove the stack
1143 CannotRemove = true;
1149 // If the stack restore is in a return, resume, or unwind block and if there
1150 // are no allocas or calls between the restore and the return, nuke the
1152 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1153 return EraseInstFromFunction(CI);
1156 case Intrinsic::assume: {
1157 // Canonicalize assume(a && b) -> assume(a); assume(b);
1158 // Note: New assumption intrinsics created here are registered by
1159 // the InstCombineIRInserter object.
1160 Value *IIOperand = II->getArgOperand(0), *A, *B,
1161 *AssumeIntrinsic = II->getCalledValue();
1162 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1163 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1164 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1165 return EraseInstFromFunction(*II);
1167 // assume(!(a || b)) -> assume(!a); assume(!b);
1168 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1169 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1171 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1173 return EraseInstFromFunction(*II);
1176 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1177 // (if assume is valid at the load)
1178 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1179 Value *LHS = ICmp->getOperand(0);
1180 Value *RHS = ICmp->getOperand(1);
1181 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1182 isa<LoadInst>(LHS) &&
1183 isa<Constant>(RHS) &&
1184 RHS->getType()->isPointerTy() &&
1185 cast<Constant>(RHS)->isNullValue()) {
1186 LoadInst* LI = cast<LoadInst>(LHS);
1187 if (isValidAssumeForContext(II, LI, DT)) {
1188 MDNode *MD = MDNode::get(II->getContext(), None);
1189 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1190 return EraseInstFromFunction(*II);
1193 // TODO: apply nonnull return attributes to calls and invokes
1194 // TODO: apply range metadata for range check patterns?
1196 // If there is a dominating assume with the same condition as this one,
1197 // then this one is redundant, and should be removed.
1198 APInt KnownZero(1, 0), KnownOne(1, 0);
1199 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1200 if (KnownOne.isAllOnesValue())
1201 return EraseInstFromFunction(*II);
1205 case Intrinsic::experimental_gc_relocate: {
1206 // Translate facts known about a pointer before relocating into
1207 // facts about the relocate value, while being careful to
1208 // preserve relocation semantics.
1209 GCRelocateOperands Operands(II);
1210 Value *DerivedPtr = Operands.getDerivedPtr();
1211 auto *GCRelocateType = cast<PointerType>(II->getType());
1213 // Remove the relocation if unused, note that this check is required
1214 // to prevent the cases below from looping forever.
1215 if (II->use_empty())
1216 return EraseInstFromFunction(*II);
1218 // Undef is undef, even after relocation.
1219 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1220 // most practical collectors, but there was discussion in the review thread
1221 // about whether it was legal for all possible collectors.
1222 if (isa<UndefValue>(DerivedPtr)) {
1223 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1224 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1227 // The relocation of null will be null for most any collector.
1228 // TODO: provide a hook for this in GCStrategy. There might be some weird
1229 // collector this property does not hold for.
1230 if (isa<ConstantPointerNull>(DerivedPtr)) {
1231 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1232 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1235 // isKnownNonNull -> nonnull attribute
1236 if (isKnownNonNull(DerivedPtr))
1237 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1239 // isDereferenceablePointer -> deref attribute
1240 if (isDereferenceablePointer(DerivedPtr, DL)) {
1241 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1242 uint64_t Bytes = A->getDereferenceableBytes();
1243 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1247 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1248 // Canonicalize on the type from the uses to the defs
1250 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1254 return visitCallSite(II);
1257 // InvokeInst simplification
1259 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1260 return visitCallSite(&II);
1263 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1264 /// passed through the varargs area, we can eliminate the use of the cast.
1265 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1266 const DataLayout &DL,
1267 const CastInst *const CI,
1269 if (!CI->isLosslessCast())
1272 // If this is a GC intrinsic, avoid munging types. We need types for
1273 // statepoint reconstruction in SelectionDAG.
1274 // TODO: This is probably something which should be expanded to all
1275 // intrinsics since the entire point of intrinsics is that
1276 // they are understandable by the optimizer.
1277 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1280 // The size of ByVal or InAlloca arguments is derived from the type, so we
1281 // can't change to a type with a different size. If the size were
1282 // passed explicitly we could avoid this check.
1283 if (!CS.isByValOrInAllocaArgument(ix))
1287 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1288 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1289 if (!SrcTy->isSized() || !DstTy->isSized())
1291 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1296 // Try to fold some different type of calls here.
1297 // Currently we're only working with the checking functions, memcpy_chk,
1298 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1299 // strcat_chk and strncat_chk.
1300 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1301 if (!CI->getCalledFunction()) return nullptr;
1303 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1304 ReplaceInstUsesWith(*From, With);
1306 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1307 if (Value *With = Simplifier.optimizeCall(CI)) {
1309 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1315 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1316 // Strip off at most one level of pointer casts, looking for an alloca. This
1317 // is good enough in practice and simpler than handling any number of casts.
1318 Value *Underlying = TrampMem->stripPointerCasts();
1319 if (Underlying != TrampMem &&
1320 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1322 if (!isa<AllocaInst>(Underlying))
1325 IntrinsicInst *InitTrampoline = nullptr;
1326 for (User *U : TrampMem->users()) {
1327 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1330 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1332 // More than one init_trampoline writes to this value. Give up.
1334 InitTrampoline = II;
1337 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1338 // Allow any number of calls to adjust.trampoline.
1343 // No call to init.trampoline found.
1344 if (!InitTrampoline)
1347 // Check that the alloca is being used in the expected way.
1348 if (InitTrampoline->getOperand(0) != TrampMem)
1351 return InitTrampoline;
1354 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1356 // Visit all the previous instructions in the basic block, and try to find a
1357 // init.trampoline which has a direct path to the adjust.trampoline.
1358 for (BasicBlock::iterator I = AdjustTramp,
1359 E = AdjustTramp->getParent()->begin(); I != E; ) {
1360 Instruction *Inst = --I;
1361 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1362 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1363 II->getOperand(0) == TrampMem)
1365 if (Inst->mayWriteToMemory())
1371 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1372 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1373 // to a direct call to a function. Otherwise return NULL.
1375 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1376 Callee = Callee->stripPointerCasts();
1377 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1379 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1382 Value *TrampMem = AdjustTramp->getOperand(0);
1384 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1386 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1391 // visitCallSite - Improvements for call and invoke instructions.
1393 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1394 if (isAllocLikeFn(CS.getInstruction(), TLI))
1395 return visitAllocSite(*CS.getInstruction());
1397 bool Changed = false;
1399 // If the callee is a pointer to a function, attempt to move any casts to the
1400 // arguments of the call/invoke.
1401 Value *Callee = CS.getCalledValue();
1402 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1405 if (Function *CalleeF = dyn_cast<Function>(Callee))
1406 // If the call and callee calling conventions don't match, this call must
1407 // be unreachable, as the call is undefined.
1408 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1409 // Only do this for calls to a function with a body. A prototype may
1410 // not actually end up matching the implementation's calling conv for a
1411 // variety of reasons (e.g. it may be written in assembly).
1412 !CalleeF->isDeclaration()) {
1413 Instruction *OldCall = CS.getInstruction();
1414 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1415 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1417 // If OldCall does not return void then replaceAllUsesWith undef.
1418 // This allows ValueHandlers and custom metadata to adjust itself.
1419 if (!OldCall->getType()->isVoidTy())
1420 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1421 if (isa<CallInst>(OldCall))
1422 return EraseInstFromFunction(*OldCall);
1424 // We cannot remove an invoke, because it would change the CFG, just
1425 // change the callee to a null pointer.
1426 cast<InvokeInst>(OldCall)->setCalledFunction(
1427 Constant::getNullValue(CalleeF->getType()));
1431 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1432 // If CS does not return void then replaceAllUsesWith undef.
1433 // This allows ValueHandlers and custom metadata to adjust itself.
1434 if (!CS.getInstruction()->getType()->isVoidTy())
1435 ReplaceInstUsesWith(*CS.getInstruction(),
1436 UndefValue::get(CS.getInstruction()->getType()));
1438 if (isa<InvokeInst>(CS.getInstruction())) {
1439 // Can't remove an invoke because we cannot change the CFG.
1443 // This instruction is not reachable, just remove it. We insert a store to
1444 // undef so that we know that this code is not reachable, despite the fact
1445 // that we can't modify the CFG here.
1446 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1447 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1448 CS.getInstruction());
1450 return EraseInstFromFunction(*CS.getInstruction());
1453 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1454 return transformCallThroughTrampoline(CS, II);
1456 PointerType *PTy = cast<PointerType>(Callee->getType());
1457 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1458 if (FTy->isVarArg()) {
1459 int ix = FTy->getNumParams();
1460 // See if we can optimize any arguments passed through the varargs area of
1462 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1463 E = CS.arg_end(); I != E; ++I, ++ix) {
1464 CastInst *CI = dyn_cast<CastInst>(*I);
1465 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1466 *I = CI->getOperand(0);
1472 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1473 // Inline asm calls cannot throw - mark them 'nounwind'.
1474 CS.setDoesNotThrow();
1478 // Try to optimize the call if possible, we require DataLayout for most of
1479 // this. None of these calls are seen as possibly dead so go ahead and
1480 // delete the instruction now.
1481 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1482 Instruction *I = tryOptimizeCall(CI);
1483 // If we changed something return the result, etc. Otherwise let
1484 // the fallthrough check.
1485 if (I) return EraseInstFromFunction(*I);
1488 return Changed ? CS.getInstruction() : nullptr;
1491 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1492 // attempt to move the cast to the arguments of the call/invoke.
1494 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1496 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1499 // The prototype of thunks are a lie, don't try to directly call such
1501 if (Callee->hasFnAttribute("thunk"))
1503 Instruction *Caller = CS.getInstruction();
1504 const AttributeSet &CallerPAL = CS.getAttributes();
1506 // Okay, this is a cast from a function to a different type. Unless doing so
1507 // would cause a type conversion of one of our arguments, change this call to
1508 // be a direct call with arguments casted to the appropriate types.
1510 FunctionType *FT = Callee->getFunctionType();
1511 Type *OldRetTy = Caller->getType();
1512 Type *NewRetTy = FT->getReturnType();
1514 // Check to see if we are changing the return type...
1515 if (OldRetTy != NewRetTy) {
1517 if (NewRetTy->isStructTy())
1518 return false; // TODO: Handle multiple return values.
1520 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1521 if (Callee->isDeclaration())
1522 return false; // Cannot transform this return value.
1524 if (!Caller->use_empty() &&
1525 // void -> non-void is handled specially
1526 !NewRetTy->isVoidTy())
1527 return false; // Cannot transform this return value.
1530 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1531 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1532 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1533 return false; // Attribute not compatible with transformed value.
1536 // If the callsite is an invoke instruction, and the return value is used by
1537 // a PHI node in a successor, we cannot change the return type of the call
1538 // because there is no place to put the cast instruction (without breaking
1539 // the critical edge). Bail out in this case.
1540 if (!Caller->use_empty())
1541 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1542 for (User *U : II->users())
1543 if (PHINode *PN = dyn_cast<PHINode>(U))
1544 if (PN->getParent() == II->getNormalDest() ||
1545 PN->getParent() == II->getUnwindDest())
1549 unsigned NumActualArgs = CS.arg_size();
1550 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1552 // Prevent us turning:
1553 // declare void @takes_i32_inalloca(i32* inalloca)
1554 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1557 // call void @takes_i32_inalloca(i32* null)
1559 // Similarly, avoid folding away bitcasts of byval calls.
1560 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1561 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1564 CallSite::arg_iterator AI = CS.arg_begin();
1565 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1566 Type *ParamTy = FT->getParamType(i);
1567 Type *ActTy = (*AI)->getType();
1569 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1570 return false; // Cannot transform this parameter value.
1572 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1573 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1574 return false; // Attribute not compatible with transformed value.
1576 if (CS.isInAllocaArgument(i))
1577 return false; // Cannot transform to and from inalloca.
1579 // If the parameter is passed as a byval argument, then we have to have a
1580 // sized type and the sized type has to have the same size as the old type.
1581 if (ParamTy != ActTy &&
1582 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1583 Attribute::ByVal)) {
1584 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1585 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1588 Type *CurElTy = ActTy->getPointerElementType();
1589 if (DL.getTypeAllocSize(CurElTy) !=
1590 DL.getTypeAllocSize(ParamPTy->getElementType()))
1595 if (Callee->isDeclaration()) {
1596 // Do not delete arguments unless we have a function body.
1597 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1600 // If the callee is just a declaration, don't change the varargsness of the
1601 // call. We don't want to introduce a varargs call where one doesn't
1603 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1604 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1607 // If both the callee and the cast type are varargs, we still have to make
1608 // sure the number of fixed parameters are the same or we have the same
1609 // ABI issues as if we introduce a varargs call.
1610 if (FT->isVarArg() &&
1611 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1612 FT->getNumParams() !=
1613 cast<FunctionType>(APTy->getElementType())->getNumParams())
1617 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1618 !CallerPAL.isEmpty())
1619 // In this case we have more arguments than the new function type, but we
1620 // won't be dropping them. Check that these extra arguments have attributes
1621 // that are compatible with being a vararg call argument.
1622 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1623 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1624 if (Index <= FT->getNumParams())
1627 // Check if it has an attribute that's incompatible with varargs.
1628 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1629 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1634 // Okay, we decided that this is a safe thing to do: go ahead and start
1635 // inserting cast instructions as necessary.
1636 std::vector<Value*> Args;
1637 Args.reserve(NumActualArgs);
1638 SmallVector<AttributeSet, 8> attrVec;
1639 attrVec.reserve(NumCommonArgs);
1641 // Get any return attributes.
1642 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1644 // If the return value is not being used, the type may not be compatible
1645 // with the existing attributes. Wipe out any problematic attributes.
1646 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1648 // Add the new return attributes.
1649 if (RAttrs.hasAttributes())
1650 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1651 AttributeSet::ReturnIndex, RAttrs));
1653 AI = CS.arg_begin();
1654 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1655 Type *ParamTy = FT->getParamType(i);
1657 if ((*AI)->getType() == ParamTy) {
1658 Args.push_back(*AI);
1660 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1663 // Add any parameter attributes.
1664 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1665 if (PAttrs.hasAttributes())
1666 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1670 // If the function takes more arguments than the call was taking, add them
1672 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1673 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1675 // If we are removing arguments to the function, emit an obnoxious warning.
1676 if (FT->getNumParams() < NumActualArgs) {
1677 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1678 if (FT->isVarArg()) {
1679 // Add all of the arguments in their promoted form to the arg list.
1680 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1681 Type *PTy = getPromotedType((*AI)->getType());
1682 if (PTy != (*AI)->getType()) {
1683 // Must promote to pass through va_arg area!
1684 Instruction::CastOps opcode =
1685 CastInst::getCastOpcode(*AI, false, PTy, false);
1686 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1688 Args.push_back(*AI);
1691 // Add any parameter attributes.
1692 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1693 if (PAttrs.hasAttributes())
1694 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1700 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1701 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1702 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1704 if (NewRetTy->isVoidTy())
1705 Caller->setName(""); // Void type should not have a name.
1707 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1711 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1712 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1713 II->getUnwindDest(), Args);
1715 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1716 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1718 CallInst *CI = cast<CallInst>(Caller);
1719 NC = Builder->CreateCall(Callee, Args);
1721 if (CI->isTailCall())
1722 cast<CallInst>(NC)->setTailCall();
1723 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1724 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1727 // Insert a cast of the return type as necessary.
1729 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1730 if (!NV->getType()->isVoidTy()) {
1731 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1732 NC->setDebugLoc(Caller->getDebugLoc());
1734 // If this is an invoke instruction, we should insert it after the first
1735 // non-phi, instruction in the normal successor block.
1736 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1737 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1738 InsertNewInstBefore(NC, *I);
1740 // Otherwise, it's a call, just insert cast right after the call.
1741 InsertNewInstBefore(NC, *Caller);
1743 Worklist.AddUsersToWorkList(*Caller);
1745 NV = UndefValue::get(Caller->getType());
1749 if (!Caller->use_empty())
1750 ReplaceInstUsesWith(*Caller, NV);
1751 else if (Caller->hasValueHandle()) {
1752 if (OldRetTy == NV->getType())
1753 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1755 // We cannot call ValueIsRAUWd with a different type, and the
1756 // actual tracked value will disappear.
1757 ValueHandleBase::ValueIsDeleted(Caller);
1760 EraseInstFromFunction(*Caller);
1764 // transformCallThroughTrampoline - Turn a call to a function created by
1765 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1766 // underlying function.
1769 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1770 IntrinsicInst *Tramp) {
1771 Value *Callee = CS.getCalledValue();
1772 PointerType *PTy = cast<PointerType>(Callee->getType());
1773 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1774 const AttributeSet &Attrs = CS.getAttributes();
1776 // If the call already has the 'nest' attribute somewhere then give up -
1777 // otherwise 'nest' would occur twice after splicing in the chain.
1778 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1782 "transformCallThroughTrampoline called with incorrect CallSite.");
1784 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1785 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1786 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1788 const AttributeSet &NestAttrs = NestF->getAttributes();
1789 if (!NestAttrs.isEmpty()) {
1790 unsigned NestIdx = 1;
1791 Type *NestTy = nullptr;
1792 AttributeSet NestAttr;
1794 // Look for a parameter marked with the 'nest' attribute.
1795 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1796 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1797 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1798 // Record the parameter type and any other attributes.
1800 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1805 Instruction *Caller = CS.getInstruction();
1806 std::vector<Value*> NewArgs;
1807 NewArgs.reserve(CS.arg_size() + 1);
1809 SmallVector<AttributeSet, 8> NewAttrs;
1810 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1812 // Insert the nest argument into the call argument list, which may
1813 // mean appending it. Likewise for attributes.
1815 // Add any result attributes.
1816 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1817 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1818 Attrs.getRetAttributes()));
1822 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1824 if (Idx == NestIdx) {
1825 // Add the chain argument and attributes.
1826 Value *NestVal = Tramp->getArgOperand(2);
1827 if (NestVal->getType() != NestTy)
1828 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1829 NewArgs.push_back(NestVal);
1830 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1837 // Add the original argument and attributes.
1838 NewArgs.push_back(*I);
1839 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1840 if (Attr.hasAttributes(Idx)) {
1841 AttrBuilder B(Attr, Idx);
1842 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1843 Idx + (Idx >= NestIdx), B));
1850 // Add any function attributes.
1851 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1852 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1853 Attrs.getFnAttributes()));
1855 // The trampoline may have been bitcast to a bogus type (FTy).
1856 // Handle this by synthesizing a new function type, equal to FTy
1857 // with the chain parameter inserted.
1859 std::vector<Type*> NewTypes;
1860 NewTypes.reserve(FTy->getNumParams()+1);
1862 // Insert the chain's type into the list of parameter types, which may
1863 // mean appending it.
1866 FunctionType::param_iterator I = FTy->param_begin(),
1867 E = FTy->param_end();
1871 // Add the chain's type.
1872 NewTypes.push_back(NestTy);
1877 // Add the original type.
1878 NewTypes.push_back(*I);
1884 // Replace the trampoline call with a direct call. Let the generic
1885 // code sort out any function type mismatches.
1886 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1888 Constant *NewCallee =
1889 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1890 NestF : ConstantExpr::getBitCast(NestF,
1891 PointerType::getUnqual(NewFTy));
1892 const AttributeSet &NewPAL =
1893 AttributeSet::get(FTy->getContext(), NewAttrs);
1895 Instruction *NewCaller;
1896 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1897 NewCaller = InvokeInst::Create(NewCallee,
1898 II->getNormalDest(), II->getUnwindDest(),
1900 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1901 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1903 NewCaller = CallInst::Create(NewCallee, NewArgs);
1904 if (cast<CallInst>(Caller)->isTailCall())
1905 cast<CallInst>(NewCaller)->setTailCall();
1906 cast<CallInst>(NewCaller)->
1907 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1908 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1915 // Replace the trampoline call with a direct call. Since there is no 'nest'
1916 // parameter, there is no need to adjust the argument list. Let the generic
1917 // code sort out any function type mismatches.
1918 Constant *NewCallee =
1919 NestF->getType() == PTy ? NestF :
1920 ConstantExpr::getBitCast(NestF, PTy);
1921 CS.setCalledFunction(NewCallee);
1922 return CS.getInstruction();