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/MemoryBuiltins.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/Dominators.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/IR/Statepoint.h"
21 #include "llvm/Transforms/Utils/BuildLibCalls.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
25 using namespace PatternMatch;
27 #define DEBUG_TYPE "instcombine"
29 STATISTIC(NumSimplified, "Number of library calls simplified");
31 /// getPromotedType - Return the specified type promoted as it would be to pass
32 /// though a va_arg area.
33 static Type *getPromotedType(Type *Ty) {
34 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
35 if (ITy->getBitWidth() < 32)
36 return Type::getInt32Ty(Ty->getContext());
41 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
42 /// single scalar element, like {{{type}}} or [1 x type], return type.
43 static Type *reduceToSingleValueType(Type *T) {
44 while (!T->isSingleValueType()) {
45 if (StructType *STy = dyn_cast<StructType>(T)) {
46 if (STy->getNumElements() == 1)
47 T = STy->getElementType(0);
50 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
51 if (ATy->getNumElements() == 1)
52 T = ATy->getElementType();
62 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
63 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, MI, AC, DT);
64 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, AC, DT);
65 unsigned MinAlign = std::min(DstAlign, SrcAlign);
66 unsigned CopyAlign = MI->getAlignment();
68 if (CopyAlign < MinAlign) {
69 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
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 *SimplifyX86insertps(const IntrinsicInst &II,
201 InstCombiner::BuilderTy &Builder) {
202 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
203 VectorType *VecTy = cast<VectorType>(II.getType());
204 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
206 // The immediate permute control byte looks like this:
207 // [3:0] - zero mask for each 32-bit lane
208 // [5:4] - select one 32-bit destination lane
209 // [7:6] - select one 32-bit source lane
211 uint8_t Imm = CInt->getZExtValue();
212 uint8_t ZMask = Imm & 0xf;
213 uint8_t DestLane = (Imm >> 4) & 0x3;
214 uint8_t SourceLane = (Imm >> 6) & 0x3;
216 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
218 // If all zero mask bits are set, this was just a weird way to
219 // generate a zero vector.
223 // Initialize by passing all of the first source bits through.
224 int ShuffleMask[4] = { 0, 1, 2, 3 };
226 // We may replace the second operand with the zero vector.
227 Value *V1 = II.getArgOperand(1);
230 // If the zero mask is being used with a single input or the zero mask
231 // overrides the destination lane, this is a shuffle with the zero vector.
232 if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
233 (ZMask & (1 << DestLane))) {
235 // We may still move 32-bits of the first source vector from one lane
237 ShuffleMask[DestLane] = SourceLane;
238 // The zero mask may override the previous insert operation.
239 for (unsigned i = 0; i < 4; ++i)
240 if ((ZMask >> i) & 0x1)
241 ShuffleMask[i] = i + 4;
243 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
247 // Replace the selected destination lane with the selected source lane.
248 ShuffleMask[DestLane] = SourceLane + 4;
251 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
256 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
257 /// source vectors, unless a zero bit is set. If a zero bit is set,
258 /// then ignore that half of the mask and clear that half of the vector.
259 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
260 InstCombiner::BuilderTy &Builder) {
261 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
262 VectorType *VecTy = cast<VectorType>(II.getType());
263 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
265 // The immediate permute control byte looks like this:
266 // [1:0] - select 128 bits from sources for low half of destination
268 // [3] - zero low half of destination
269 // [5:4] - select 128 bits from sources for high half of destination
271 // [7] - zero high half of destination
273 uint8_t Imm = CInt->getZExtValue();
275 bool LowHalfZero = Imm & 0x08;
276 bool HighHalfZero = Imm & 0x80;
278 // If both zero mask bits are set, this was just a weird way to
279 // generate a zero vector.
280 if (LowHalfZero && HighHalfZero)
283 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
284 unsigned NumElts = VecTy->getNumElements();
285 unsigned HalfSize = NumElts / 2;
286 SmallVector<int, 8> ShuffleMask(NumElts);
288 // The high bit of the selection field chooses the 1st or 2nd operand.
289 bool LowInputSelect = Imm & 0x02;
290 bool HighInputSelect = Imm & 0x20;
292 // The low bit of the selection field chooses the low or high half
293 // of the selected operand.
294 bool LowHalfSelect = Imm & 0x01;
295 bool HighHalfSelect = Imm & 0x10;
297 // Determine which operand(s) are actually in use for this instruction.
298 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
299 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
301 // If needed, replace operands based on zero mask.
302 V0 = LowHalfZero ? ZeroVector : V0;
303 V1 = HighHalfZero ? ZeroVector : V1;
305 // Permute low half of result.
306 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
307 for (unsigned i = 0; i < HalfSize; ++i)
308 ShuffleMask[i] = StartIndex + i;
310 // Permute high half of result.
311 StartIndex = HighHalfSelect ? HalfSize : 0;
312 StartIndex += NumElts;
313 for (unsigned i = 0; i < HalfSize; ++i)
314 ShuffleMask[i + HalfSize] = StartIndex + i;
316 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
321 /// visitCallInst - CallInst simplification. This mostly only handles folding
322 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
323 /// the heavy lifting.
325 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
326 if (isFreeCall(&CI, TLI))
327 return visitFree(CI);
329 // If the caller function is nounwind, mark the call as nounwind, even if the
331 if (CI.getParent()->getParent()->doesNotThrow() &&
332 !CI.doesNotThrow()) {
333 CI.setDoesNotThrow();
337 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
338 if (!II) return visitCallSite(&CI);
340 // Intrinsics cannot occur in an invoke, so handle them here instead of in
342 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
343 bool Changed = false;
345 // memmove/cpy/set of zero bytes is a noop.
346 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
347 if (NumBytes->isNullValue())
348 return EraseInstFromFunction(CI);
350 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
351 if (CI->getZExtValue() == 1) {
352 // Replace the instruction with just byte operations. We would
353 // transform other cases to loads/stores, but we don't know if
354 // alignment is sufficient.
358 // No other transformations apply to volatile transfers.
359 if (MI->isVolatile())
362 // If we have a memmove and the source operation is a constant global,
363 // then the source and dest pointers can't alias, so we can change this
364 // into a call to memcpy.
365 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
366 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
367 if (GVSrc->isConstant()) {
368 Module *M = CI.getParent()->getParent()->getParent();
369 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
370 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
371 CI.getArgOperand(1)->getType(),
372 CI.getArgOperand(2)->getType() };
373 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
378 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
379 // memmove(x,x,size) -> noop.
380 if (MTI->getSource() == MTI->getDest())
381 return EraseInstFromFunction(CI);
384 // If we can determine a pointer alignment that is bigger than currently
385 // set, update the alignment.
386 if (isa<MemTransferInst>(MI)) {
387 if (Instruction *I = SimplifyMemTransfer(MI))
389 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
390 if (Instruction *I = SimplifyMemSet(MSI))
394 if (Changed) return II;
397 switch (II->getIntrinsicID()) {
399 case Intrinsic::objectsize: {
401 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
402 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
405 case Intrinsic::bswap: {
406 Value *IIOperand = II->getArgOperand(0);
409 // bswap(bswap(x)) -> x
410 if (match(IIOperand, m_BSwap(m_Value(X))))
411 return ReplaceInstUsesWith(CI, X);
413 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
414 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
415 unsigned C = X->getType()->getPrimitiveSizeInBits() -
416 IIOperand->getType()->getPrimitiveSizeInBits();
417 Value *CV = ConstantInt::get(X->getType(), C);
418 Value *V = Builder->CreateLShr(X, CV);
419 return new TruncInst(V, IIOperand->getType());
424 case Intrinsic::powi:
425 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
428 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
431 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
432 // powi(x, -1) -> 1/x
433 if (Power->isAllOnesValue())
434 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
435 II->getArgOperand(0));
438 case Intrinsic::cttz: {
439 // If all bits below the first known one are known zero,
440 // this value is constant.
441 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
442 // FIXME: Try to simplify vectors of integers.
444 uint32_t BitWidth = IT->getBitWidth();
445 APInt KnownZero(BitWidth, 0);
446 APInt KnownOne(BitWidth, 0);
447 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
448 unsigned TrailingZeros = KnownOne.countTrailingZeros();
449 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
450 if ((Mask & KnownZero) == Mask)
451 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
452 APInt(BitWidth, TrailingZeros)));
456 case Intrinsic::ctlz: {
457 // If all bits above the first known one are known zero,
458 // this value is constant.
459 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
460 // FIXME: Try to simplify vectors of integers.
462 uint32_t BitWidth = IT->getBitWidth();
463 APInt KnownZero(BitWidth, 0);
464 APInt KnownOne(BitWidth, 0);
465 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
466 unsigned LeadingZeros = KnownOne.countLeadingZeros();
467 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
468 if ((Mask & KnownZero) == Mask)
469 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
470 APInt(BitWidth, LeadingZeros)));
475 case Intrinsic::uadd_with_overflow:
476 case Intrinsic::sadd_with_overflow:
477 case Intrinsic::umul_with_overflow:
478 case Intrinsic::smul_with_overflow:
479 if (isa<Constant>(II->getArgOperand(0)) &&
480 !isa<Constant>(II->getArgOperand(1))) {
481 // Canonicalize constants into the RHS.
482 Value *LHS = II->getArgOperand(0);
483 II->setArgOperand(0, II->getArgOperand(1));
484 II->setArgOperand(1, LHS);
489 case Intrinsic::usub_with_overflow:
490 case Intrinsic::ssub_with_overflow: {
491 OverflowCheckFlavor OCF =
492 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
493 assert(OCF != OCF_INVALID && "unexpected!");
495 Value *OperationResult = nullptr;
496 Constant *OverflowResult = nullptr;
497 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
498 *II, OperationResult, OverflowResult))
499 return CreateOverflowTuple(II, OperationResult, OverflowResult);
504 case Intrinsic::minnum:
505 case Intrinsic::maxnum: {
506 Value *Arg0 = II->getArgOperand(0);
507 Value *Arg1 = II->getArgOperand(1);
511 return ReplaceInstUsesWith(CI, Arg0);
513 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
514 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
516 // Canonicalize constants into the RHS.
518 II->setArgOperand(0, Arg1);
519 II->setArgOperand(1, Arg0);
524 if (C1 && C1->isNaN())
525 return ReplaceInstUsesWith(CI, Arg0);
527 // This is the value because if undef were NaN, we would return the other
528 // value and cannot return a NaN unless both operands are.
530 // fmin(undef, x) -> x
531 if (isa<UndefValue>(Arg0))
532 return ReplaceInstUsesWith(CI, Arg1);
534 // fmin(x, undef) -> x
535 if (isa<UndefValue>(Arg1))
536 return ReplaceInstUsesWith(CI, Arg0);
540 if (II->getIntrinsicID() == Intrinsic::minnum) {
541 // fmin(x, fmin(x, y)) -> fmin(x, y)
542 // fmin(y, fmin(x, y)) -> fmin(x, y)
543 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
544 if (Arg0 == X || Arg0 == Y)
545 return ReplaceInstUsesWith(CI, Arg1);
548 // fmin(fmin(x, y), x) -> fmin(x, y)
549 // fmin(fmin(x, y), y) -> fmin(x, y)
550 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
551 if (Arg1 == X || Arg1 == Y)
552 return ReplaceInstUsesWith(CI, Arg0);
555 // TODO: fmin(nnan x, inf) -> x
556 // TODO: fmin(nnan ninf x, flt_max) -> x
557 if (C1 && C1->isInfinity()) {
558 // fmin(x, -inf) -> -inf
559 if (C1->isNegative())
560 return ReplaceInstUsesWith(CI, Arg1);
563 assert(II->getIntrinsicID() == Intrinsic::maxnum);
564 // fmax(x, fmax(x, y)) -> fmax(x, y)
565 // fmax(y, fmax(x, y)) -> fmax(x, y)
566 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
567 if (Arg0 == X || Arg0 == Y)
568 return ReplaceInstUsesWith(CI, Arg1);
571 // fmax(fmax(x, y), x) -> fmax(x, y)
572 // fmax(fmax(x, y), y) -> fmax(x, y)
573 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
574 if (Arg1 == X || Arg1 == Y)
575 return ReplaceInstUsesWith(CI, Arg0);
578 // TODO: fmax(nnan x, -inf) -> x
579 // TODO: fmax(nnan ninf x, -flt_max) -> x
580 if (C1 && C1->isInfinity()) {
581 // fmax(x, inf) -> inf
582 if (!C1->isNegative())
583 return ReplaceInstUsesWith(CI, Arg1);
588 case Intrinsic::ppc_altivec_lvx:
589 case Intrinsic::ppc_altivec_lvxl:
590 // Turn PPC lvx -> load if the pointer is known aligned.
591 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
593 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
594 PointerType::getUnqual(II->getType()));
595 return new LoadInst(Ptr);
598 case Intrinsic::ppc_vsx_lxvw4x:
599 case Intrinsic::ppc_vsx_lxvd2x: {
600 // Turn PPC VSX loads into normal loads.
601 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
602 PointerType::getUnqual(II->getType()));
603 return new LoadInst(Ptr, Twine(""), false, 1);
605 case Intrinsic::ppc_altivec_stvx:
606 case Intrinsic::ppc_altivec_stvxl:
607 // Turn stvx -> store if the pointer is known aligned.
608 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
611 PointerType::getUnqual(II->getArgOperand(0)->getType());
612 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
613 return new StoreInst(II->getArgOperand(0), Ptr);
616 case Intrinsic::ppc_vsx_stxvw4x:
617 case Intrinsic::ppc_vsx_stxvd2x: {
618 // Turn PPC VSX stores into normal stores.
619 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
620 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
621 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
623 case Intrinsic::ppc_qpx_qvlfs:
624 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
625 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
627 Type *VTy = VectorType::get(Builder->getFloatTy(),
628 II->getType()->getVectorNumElements());
629 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
630 PointerType::getUnqual(VTy));
631 Value *Load = Builder->CreateLoad(Ptr);
632 return new FPExtInst(Load, II->getType());
635 case Intrinsic::ppc_qpx_qvlfd:
636 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
637 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
639 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
640 PointerType::getUnqual(II->getType()));
641 return new LoadInst(Ptr);
644 case Intrinsic::ppc_qpx_qvstfs:
645 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
646 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
648 Type *VTy = VectorType::get(Builder->getFloatTy(),
649 II->getArgOperand(0)->getType()->getVectorNumElements());
650 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy);
651 Type *OpPtrTy = PointerType::getUnqual(VTy);
652 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
653 return new StoreInst(TOp, Ptr);
656 case Intrinsic::ppc_qpx_qvstfd:
657 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
658 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
661 PointerType::getUnqual(II->getArgOperand(0)->getType());
662 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
663 return new StoreInst(II->getArgOperand(0), Ptr);
666 case Intrinsic::x86_sse_storeu_ps:
667 case Intrinsic::x86_sse2_storeu_pd:
668 case Intrinsic::x86_sse2_storeu_dq:
669 // Turn X86 storeu -> store if the pointer is known aligned.
670 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
673 PointerType::getUnqual(II->getArgOperand(1)->getType());
674 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
675 return new StoreInst(II->getArgOperand(1), Ptr);
679 case Intrinsic::x86_sse_cvtss2si:
680 case Intrinsic::x86_sse_cvtss2si64:
681 case Intrinsic::x86_sse_cvttss2si:
682 case Intrinsic::x86_sse_cvttss2si64:
683 case Intrinsic::x86_sse2_cvtsd2si:
684 case Intrinsic::x86_sse2_cvtsd2si64:
685 case Intrinsic::x86_sse2_cvttsd2si:
686 case Intrinsic::x86_sse2_cvttsd2si64: {
687 // These intrinsics only demand the 0th element of their input vectors. If
688 // we can simplify the input based on that, do so now.
690 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
691 APInt DemandedElts(VWidth, 1);
692 APInt UndefElts(VWidth, 0);
693 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
694 DemandedElts, UndefElts)) {
695 II->setArgOperand(0, V);
701 // Constant fold <A x Bi> << Ci.
702 // FIXME: We don't handle _dq because it's a shift of an i128, but is
703 // represented in the IR as <2 x i64>. A per element shift is wrong.
704 case Intrinsic::x86_sse2_psll_d:
705 case Intrinsic::x86_sse2_psll_q:
706 case Intrinsic::x86_sse2_psll_w:
707 case Intrinsic::x86_sse2_pslli_d:
708 case Intrinsic::x86_sse2_pslli_q:
709 case Intrinsic::x86_sse2_pslli_w:
710 case Intrinsic::x86_avx2_psll_d:
711 case Intrinsic::x86_avx2_psll_q:
712 case Intrinsic::x86_avx2_psll_w:
713 case Intrinsic::x86_avx2_pslli_d:
714 case Intrinsic::x86_avx2_pslli_q:
715 case Intrinsic::x86_avx2_pslli_w:
716 case Intrinsic::x86_sse2_psrl_d:
717 case Intrinsic::x86_sse2_psrl_q:
718 case Intrinsic::x86_sse2_psrl_w:
719 case Intrinsic::x86_sse2_psrli_d:
720 case Intrinsic::x86_sse2_psrli_q:
721 case Intrinsic::x86_sse2_psrli_w:
722 case Intrinsic::x86_avx2_psrl_d:
723 case Intrinsic::x86_avx2_psrl_q:
724 case Intrinsic::x86_avx2_psrl_w:
725 case Intrinsic::x86_avx2_psrli_d:
726 case Intrinsic::x86_avx2_psrli_q:
727 case Intrinsic::x86_avx2_psrli_w: {
728 // Simplify if count is constant. To 0 if >= BitWidth,
729 // otherwise to shl/lshr.
730 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
731 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
736 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
740 auto Vec = II->getArgOperand(0);
741 auto VT = cast<VectorType>(Vec->getType());
742 if (Count->getZExtValue() >
743 VT->getElementType()->getPrimitiveSizeInBits() - 1)
744 return ReplaceInstUsesWith(
745 CI, ConstantAggregateZero::get(Vec->getType()));
747 bool isPackedShiftLeft = true;
748 switch (II->getIntrinsicID()) {
750 case Intrinsic::x86_sse2_psrl_d:
751 case Intrinsic::x86_sse2_psrl_q:
752 case Intrinsic::x86_sse2_psrl_w:
753 case Intrinsic::x86_sse2_psrli_d:
754 case Intrinsic::x86_sse2_psrli_q:
755 case Intrinsic::x86_sse2_psrli_w:
756 case Intrinsic::x86_avx2_psrl_d:
757 case Intrinsic::x86_avx2_psrl_q:
758 case Intrinsic::x86_avx2_psrl_w:
759 case Intrinsic::x86_avx2_psrli_d:
760 case Intrinsic::x86_avx2_psrli_q:
761 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
764 unsigned VWidth = VT->getNumElements();
765 // Get a constant vector of the same type as the first operand.
766 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
767 if (isPackedShiftLeft)
768 return BinaryOperator::CreateShl(Vec,
769 Builder->CreateVectorSplat(VWidth, VTCI));
771 return BinaryOperator::CreateLShr(Vec,
772 Builder->CreateVectorSplat(VWidth, VTCI));
775 case Intrinsic::x86_sse41_pmovsxbw:
776 case Intrinsic::x86_sse41_pmovsxwd:
777 case Intrinsic::x86_sse41_pmovsxdq:
778 case Intrinsic::x86_sse41_pmovzxbw:
779 case Intrinsic::x86_sse41_pmovzxwd:
780 case Intrinsic::x86_sse41_pmovzxdq: {
781 // pmov{s|z}x ignores the upper half of their input vectors.
783 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
784 unsigned LowHalfElts = VWidth / 2;
785 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
786 APInt UndefElts(VWidth, 0);
787 if (Value *TmpV = SimplifyDemandedVectorElts(
788 II->getArgOperand(0), InputDemandedElts, UndefElts)) {
789 II->setArgOperand(0, TmpV);
794 case Intrinsic::x86_sse41_insertps:
795 if (Value *V = SimplifyX86insertps(*II, *Builder))
796 return ReplaceInstUsesWith(*II, V);
799 case Intrinsic::x86_sse4a_insertqi: {
800 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
802 // TODO: eventually we should lower this intrinsic to IR
803 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
804 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
805 unsigned Index = CIStart->getZExtValue();
806 // From AMD documentation: "a value of zero in the field length is
807 // defined as length of 64".
808 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
810 // From AMD documentation: "If the sum of the bit index + length field
811 // is greater than 64, the results are undefined".
813 // Note that both field index and field length are 8-bit quantities.
814 // Since variables 'Index' and 'Length' are unsigned values
815 // obtained from zero-extending field index and field length
816 // respectively, their sum should never wrap around.
817 if ((Index + Length) > 64)
818 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
820 if (Length == 64 && Index == 0) {
821 Value *Vec = II->getArgOperand(1);
822 Value *Undef = UndefValue::get(Vec->getType());
823 const uint32_t Mask[] = { 0, 2 };
824 return ReplaceInstUsesWith(
826 Builder->CreateShuffleVector(
827 Vec, Undef, ConstantDataVector::get(
828 II->getContext(), makeArrayRef(Mask))));
830 } else if (auto Source =
831 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
832 if (Source->hasOneUse() &&
833 Source->getArgOperand(1) == II->getArgOperand(1)) {
834 // If the source of the insert has only one use and it's another
835 // insert (and they're both inserting from the same vector), try to
836 // bundle both together.
838 dyn_cast<ConstantInt>(Source->getArgOperand(2));
840 dyn_cast<ConstantInt>(Source->getArgOperand(3));
841 if (CISourceStart && CISourceWidth) {
842 unsigned Start = CIStart->getZExtValue();
843 unsigned Width = CIWidth->getZExtValue();
844 unsigned End = Start + Width;
845 unsigned SourceStart = CISourceStart->getZExtValue();
846 unsigned SourceWidth = CISourceWidth->getZExtValue();
847 unsigned SourceEnd = SourceStart + SourceWidth;
848 unsigned NewStart, NewWidth;
849 bool ShouldReplace = false;
850 if (Start <= SourceStart && SourceStart <= End) {
852 NewWidth = std::max(End, SourceEnd) - NewStart;
853 ShouldReplace = true;
854 } else if (SourceStart <= Start && Start <= SourceEnd) {
855 NewStart = SourceStart;
856 NewWidth = std::max(SourceEnd, End) - NewStart;
857 ShouldReplace = true;
861 Constant *ConstantWidth = ConstantInt::get(
862 II->getArgOperand(2)->getType(), NewWidth, false);
863 Constant *ConstantStart = ConstantInt::get(
864 II->getArgOperand(3)->getType(), NewStart, false);
865 Value *Args[4] = { Source->getArgOperand(0),
866 II->getArgOperand(1), ConstantWidth,
868 Module *M = CI.getParent()->getParent()->getParent();
870 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
871 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
881 case Intrinsic::x86_sse41_pblendvb:
882 case Intrinsic::x86_sse41_blendvps:
883 case Intrinsic::x86_sse41_blendvpd:
884 case Intrinsic::x86_avx_blendv_ps_256:
885 case Intrinsic::x86_avx_blendv_pd_256:
886 case Intrinsic::x86_avx2_pblendvb: {
887 // Convert blendv* to vector selects if the mask is constant.
888 // This optimization is convoluted because the intrinsic is defined as
889 // getting a vector of floats or doubles for the ps and pd versions.
890 // FIXME: That should be changed.
891 Value *Mask = II->getArgOperand(2);
892 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
893 auto Tyi1 = Builder->getInt1Ty();
894 auto SelectorType = cast<VectorType>(Mask->getType());
895 auto EltTy = SelectorType->getElementType();
896 unsigned Size = SelectorType->getNumElements();
900 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
901 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
902 "Wrong arguments for variable blend intrinsic");
903 SmallVector<Constant *, 32> Selectors;
904 for (unsigned I = 0; I < Size; ++I) {
905 // The intrinsics only read the top bit
908 Selector = C->getElementAsInteger(I);
910 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
911 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
913 auto NewSelector = ConstantVector::get(Selectors);
914 return SelectInst::Create(NewSelector, II->getArgOperand(1),
915 II->getArgOperand(0), "blendv");
921 case Intrinsic::x86_avx_vpermilvar_ps:
922 case Intrinsic::x86_avx_vpermilvar_ps_256:
923 case Intrinsic::x86_avx_vpermilvar_pd:
924 case Intrinsic::x86_avx_vpermilvar_pd_256: {
925 // Convert vpermil* to shufflevector if the mask is constant.
926 Value *V = II->getArgOperand(1);
927 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
928 assert(Size == 8 || Size == 4 || Size == 2);
930 if (auto C = dyn_cast<ConstantDataVector>(V)) {
931 // The intrinsics only read one or two bits, clear the rest.
932 for (unsigned I = 0; I < Size; ++I) {
933 uint32_t Index = C->getElementAsInteger(I) & 0x3;
934 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
935 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
939 } else if (isa<ConstantAggregateZero>(V)) {
940 for (unsigned I = 0; I < Size; ++I)
945 // The _256 variants are a bit trickier since the mask bits always index
946 // into the corresponding 128 half. In order to convert to a generic
947 // shuffle, we have to make that explicit.
948 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
949 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
950 for (unsigned I = Size / 2; I < Size; ++I)
951 Indexes[I] += Size / 2;
954 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
955 auto V1 = II->getArgOperand(0);
956 auto V2 = UndefValue::get(V1->getType());
957 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
958 return ReplaceInstUsesWith(CI, Shuffle);
961 case Intrinsic::x86_avx_vperm2f128_pd_256:
962 case Intrinsic::x86_avx_vperm2f128_ps_256:
963 case Intrinsic::x86_avx_vperm2f128_si_256:
964 case Intrinsic::x86_avx2_vperm2i128:
965 if (Value *V = SimplifyX86vperm2(*II, *Builder))
966 return ReplaceInstUsesWith(*II, V);
969 case Intrinsic::ppc_altivec_vperm:
970 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
971 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
972 // a vectorshuffle for little endian, we must undo the transformation
973 // performed on vec_perm in altivec.h. That is, we must complement
974 // the permutation mask with respect to 31 and reverse the order of
976 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
977 assert(Mask->getType()->getVectorNumElements() == 16 &&
978 "Bad type for intrinsic!");
980 // Check that all of the elements are integer constants or undefs.
981 bool AllEltsOk = true;
982 for (unsigned i = 0; i != 16; ++i) {
983 Constant *Elt = Mask->getAggregateElement(i);
984 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
991 // Cast the input vectors to byte vectors.
992 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
994 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
996 Value *Result = UndefValue::get(Op0->getType());
998 // Only extract each element once.
999 Value *ExtractedElts[32];
1000 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1002 for (unsigned i = 0; i != 16; ++i) {
1003 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1006 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1007 Idx &= 31; // Match the hardware behavior.
1008 if (DL.isLittleEndian())
1011 if (!ExtractedElts[Idx]) {
1012 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1013 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1014 ExtractedElts[Idx] =
1015 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1016 Builder->getInt32(Idx&15));
1019 // Insert this value into the result vector.
1020 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1021 Builder->getInt32(i));
1023 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1028 case Intrinsic::arm_neon_vld1:
1029 case Intrinsic::arm_neon_vld2:
1030 case Intrinsic::arm_neon_vld3:
1031 case Intrinsic::arm_neon_vld4:
1032 case Intrinsic::arm_neon_vld2lane:
1033 case Intrinsic::arm_neon_vld3lane:
1034 case Intrinsic::arm_neon_vld4lane:
1035 case Intrinsic::arm_neon_vst1:
1036 case Intrinsic::arm_neon_vst2:
1037 case Intrinsic::arm_neon_vst3:
1038 case Intrinsic::arm_neon_vst4:
1039 case Intrinsic::arm_neon_vst2lane:
1040 case Intrinsic::arm_neon_vst3lane:
1041 case Intrinsic::arm_neon_vst4lane: {
1042 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1043 unsigned AlignArg = II->getNumArgOperands() - 1;
1044 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1045 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1046 II->setArgOperand(AlignArg,
1047 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1054 case Intrinsic::arm_neon_vmulls:
1055 case Intrinsic::arm_neon_vmullu:
1056 case Intrinsic::aarch64_neon_smull:
1057 case Intrinsic::aarch64_neon_umull: {
1058 Value *Arg0 = II->getArgOperand(0);
1059 Value *Arg1 = II->getArgOperand(1);
1061 // Handle mul by zero first:
1062 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1063 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1066 // Check for constant LHS & RHS - in this case we just simplify.
1067 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1068 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1069 VectorType *NewVT = cast<VectorType>(II->getType());
1070 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1071 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1072 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1073 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1075 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1078 // Couldn't simplify - canonicalize constant to the RHS.
1079 std::swap(Arg0, Arg1);
1082 // Handle mul by one:
1083 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1084 if (ConstantInt *Splat =
1085 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1087 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1088 /*isSigned=*/!Zext);
1093 case Intrinsic::AMDGPU_rcp: {
1094 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1095 const APFloat &ArgVal = C->getValueAPF();
1096 APFloat Val(ArgVal.getSemantics(), 1.0);
1097 APFloat::opStatus Status = Val.divide(ArgVal,
1098 APFloat::rmNearestTiesToEven);
1099 // Only do this if it was exact and therefore not dependent on the
1101 if (Status == APFloat::opOK)
1102 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1107 case Intrinsic::stackrestore: {
1108 // If the save is right next to the restore, remove the restore. This can
1109 // happen when variable allocas are DCE'd.
1110 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1111 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1112 BasicBlock::iterator BI = SS;
1114 return EraseInstFromFunction(CI);
1118 // Scan down this block to see if there is another stack restore in the
1119 // same block without an intervening call/alloca.
1120 BasicBlock::iterator BI = II;
1121 TerminatorInst *TI = II->getParent()->getTerminator();
1122 bool CannotRemove = false;
1123 for (++BI; &*BI != TI; ++BI) {
1124 if (isa<AllocaInst>(BI)) {
1125 CannotRemove = true;
1128 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1129 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1130 // If there is a stackrestore below this one, remove this one.
1131 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1132 return EraseInstFromFunction(CI);
1133 // Otherwise, ignore the intrinsic.
1135 // If we found a non-intrinsic call, we can't remove the stack
1137 CannotRemove = true;
1143 // If the stack restore is in a return, resume, or unwind block and if there
1144 // are no allocas or calls between the restore and the return, nuke the
1146 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1147 return EraseInstFromFunction(CI);
1150 case Intrinsic::assume: {
1151 // Canonicalize assume(a && b) -> assume(a); assume(b);
1152 // Note: New assumption intrinsics created here are registered by
1153 // the InstCombineIRInserter object.
1154 Value *IIOperand = II->getArgOperand(0), *A, *B,
1155 *AssumeIntrinsic = II->getCalledValue();
1156 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1157 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1158 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1159 return EraseInstFromFunction(*II);
1161 // assume(!(a || b)) -> assume(!a); assume(!b);
1162 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1163 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1165 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1167 return EraseInstFromFunction(*II);
1170 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1171 // (if assume is valid at the load)
1172 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1173 Value *LHS = ICmp->getOperand(0);
1174 Value *RHS = ICmp->getOperand(1);
1175 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1176 isa<LoadInst>(LHS) &&
1177 isa<Constant>(RHS) &&
1178 RHS->getType()->isPointerTy() &&
1179 cast<Constant>(RHS)->isNullValue()) {
1180 LoadInst* LI = cast<LoadInst>(LHS);
1181 if (isValidAssumeForContext(II, LI, DT)) {
1182 MDNode *MD = MDNode::get(II->getContext(), None);
1183 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1184 return EraseInstFromFunction(*II);
1187 // TODO: apply nonnull return attributes to calls and invokes
1188 // TODO: apply range metadata for range check patterns?
1190 // If there is a dominating assume with the same condition as this one,
1191 // then this one is redundant, and should be removed.
1192 APInt KnownZero(1, 0), KnownOne(1, 0);
1193 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1194 if (KnownOne.isAllOnesValue())
1195 return EraseInstFromFunction(*II);
1199 case Intrinsic::experimental_gc_relocate: {
1200 // Translate facts known about a pointer before relocating into
1201 // facts about the relocate value, while being careful to
1202 // preserve relocation semantics.
1203 GCRelocateOperands Operands(II);
1204 Value *DerivedPtr = Operands.getDerivedPtr();
1205 auto *GCRelocateType = cast<PointerType>(II->getType());
1207 // Remove the relocation if unused, note that this check is required
1208 // to prevent the cases below from looping forever.
1209 if (II->use_empty())
1210 return EraseInstFromFunction(*II);
1212 // Undef is undef, even after relocation.
1213 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1214 // most practical collectors, but there was discussion in the review thread
1215 // about whether it was legal for all possible collectors.
1216 if (isa<UndefValue>(DerivedPtr)) {
1217 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it.
1218 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType));
1221 // The relocation of null will be null for most any collector.
1222 // TODO: provide a hook for this in GCStrategy. There might be some weird
1223 // collector this property does not hold for.
1224 if (isa<ConstantPointerNull>(DerivedPtr)) {
1225 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it.
1226 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType));
1229 // isKnownNonNull -> nonnull attribute
1230 if (isKnownNonNull(DerivedPtr))
1231 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1233 // isDereferenceablePointer -> deref attribute
1234 if (isDereferenceablePointer(DerivedPtr, DL)) {
1235 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1236 uint64_t Bytes = A->getDereferenceableBytes();
1237 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1241 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1242 // Canonicalize on the type from the uses to the defs
1244 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1248 return visitCallSite(II);
1251 // InvokeInst simplification
1253 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1254 return visitCallSite(&II);
1257 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1258 /// passed through the varargs area, we can eliminate the use of the cast.
1259 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1260 const DataLayout &DL,
1261 const CastInst *const CI,
1263 if (!CI->isLosslessCast())
1266 // If this is a GC intrinsic, avoid munging types. We need types for
1267 // statepoint reconstruction in SelectionDAG.
1268 // TODO: This is probably something which should be expanded to all
1269 // intrinsics since the entire point of intrinsics is that
1270 // they are understandable by the optimizer.
1271 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1274 // The size of ByVal or InAlloca arguments is derived from the type, so we
1275 // can't change to a type with a different size. If the size were
1276 // passed explicitly we could avoid this check.
1277 if (!CS.isByValOrInAllocaArgument(ix))
1281 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1282 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1283 if (!SrcTy->isSized() || !DstTy->isSized())
1285 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1290 // Try to fold some different type of calls here.
1291 // Currently we're only working with the checking functions, memcpy_chk,
1292 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1293 // strcat_chk and strncat_chk.
1294 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1295 if (!CI->getCalledFunction()) return nullptr;
1297 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1298 ReplaceInstUsesWith(*From, With);
1300 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1301 if (Value *With = Simplifier.optimizeCall(CI)) {
1303 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1309 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1310 // Strip off at most one level of pointer casts, looking for an alloca. This
1311 // is good enough in practice and simpler than handling any number of casts.
1312 Value *Underlying = TrampMem->stripPointerCasts();
1313 if (Underlying != TrampMem &&
1314 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1316 if (!isa<AllocaInst>(Underlying))
1319 IntrinsicInst *InitTrampoline = nullptr;
1320 for (User *U : TrampMem->users()) {
1321 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1324 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1326 // More than one init_trampoline writes to this value. Give up.
1328 InitTrampoline = II;
1331 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1332 // Allow any number of calls to adjust.trampoline.
1337 // No call to init.trampoline found.
1338 if (!InitTrampoline)
1341 // Check that the alloca is being used in the expected way.
1342 if (InitTrampoline->getOperand(0) != TrampMem)
1345 return InitTrampoline;
1348 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1350 // Visit all the previous instructions in the basic block, and try to find a
1351 // init.trampoline which has a direct path to the adjust.trampoline.
1352 for (BasicBlock::iterator I = AdjustTramp,
1353 E = AdjustTramp->getParent()->begin(); I != E; ) {
1354 Instruction *Inst = --I;
1355 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1356 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1357 II->getOperand(0) == TrampMem)
1359 if (Inst->mayWriteToMemory())
1365 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1366 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1367 // to a direct call to a function. Otherwise return NULL.
1369 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1370 Callee = Callee->stripPointerCasts();
1371 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1373 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1376 Value *TrampMem = AdjustTramp->getOperand(0);
1378 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1380 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1385 // visitCallSite - Improvements for call and invoke instructions.
1387 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1388 if (isAllocLikeFn(CS.getInstruction(), TLI))
1389 return visitAllocSite(*CS.getInstruction());
1391 bool Changed = false;
1393 // If the callee is a pointer to a function, attempt to move any casts to the
1394 // arguments of the call/invoke.
1395 Value *Callee = CS.getCalledValue();
1396 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1399 if (Function *CalleeF = dyn_cast<Function>(Callee))
1400 // If the call and callee calling conventions don't match, this call must
1401 // be unreachable, as the call is undefined.
1402 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1403 // Only do this for calls to a function with a body. A prototype may
1404 // not actually end up matching the implementation's calling conv for a
1405 // variety of reasons (e.g. it may be written in assembly).
1406 !CalleeF->isDeclaration()) {
1407 Instruction *OldCall = CS.getInstruction();
1408 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1409 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1411 // If OldCall does not return void then replaceAllUsesWith undef.
1412 // This allows ValueHandlers and custom metadata to adjust itself.
1413 if (!OldCall->getType()->isVoidTy())
1414 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1415 if (isa<CallInst>(OldCall))
1416 return EraseInstFromFunction(*OldCall);
1418 // We cannot remove an invoke, because it would change the CFG, just
1419 // change the callee to a null pointer.
1420 cast<InvokeInst>(OldCall)->setCalledFunction(
1421 Constant::getNullValue(CalleeF->getType()));
1425 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1426 // If CS does not return void then replaceAllUsesWith undef.
1427 // This allows ValueHandlers and custom metadata to adjust itself.
1428 if (!CS.getInstruction()->getType()->isVoidTy())
1429 ReplaceInstUsesWith(*CS.getInstruction(),
1430 UndefValue::get(CS.getInstruction()->getType()));
1432 if (isa<InvokeInst>(CS.getInstruction())) {
1433 // Can't remove an invoke because we cannot change the CFG.
1437 // This instruction is not reachable, just remove it. We insert a store to
1438 // undef so that we know that this code is not reachable, despite the fact
1439 // that we can't modify the CFG here.
1440 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1441 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1442 CS.getInstruction());
1444 return EraseInstFromFunction(*CS.getInstruction());
1447 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1448 return transformCallThroughTrampoline(CS, II);
1450 PointerType *PTy = cast<PointerType>(Callee->getType());
1451 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1452 if (FTy->isVarArg()) {
1453 int ix = FTy->getNumParams();
1454 // See if we can optimize any arguments passed through the varargs area of
1456 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1457 E = CS.arg_end(); I != E; ++I, ++ix) {
1458 CastInst *CI = dyn_cast<CastInst>(*I);
1459 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1460 *I = CI->getOperand(0);
1466 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1467 // Inline asm calls cannot throw - mark them 'nounwind'.
1468 CS.setDoesNotThrow();
1472 // Try to optimize the call if possible, we require DataLayout for most of
1473 // this. None of these calls are seen as possibly dead so go ahead and
1474 // delete the instruction now.
1475 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1476 Instruction *I = tryOptimizeCall(CI);
1477 // If we changed something return the result, etc. Otherwise let
1478 // the fallthrough check.
1479 if (I) return EraseInstFromFunction(*I);
1482 return Changed ? CS.getInstruction() : nullptr;
1485 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1486 // attempt to move the cast to the arguments of the call/invoke.
1488 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1490 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1493 // The prototype of thunks are a lie, don't try to directly call such
1495 if (Callee->hasFnAttribute("thunk"))
1497 Instruction *Caller = CS.getInstruction();
1498 const AttributeSet &CallerPAL = CS.getAttributes();
1500 // Okay, this is a cast from a function to a different type. Unless doing so
1501 // would cause a type conversion of one of our arguments, change this call to
1502 // be a direct call with arguments casted to the appropriate types.
1504 FunctionType *FT = Callee->getFunctionType();
1505 Type *OldRetTy = Caller->getType();
1506 Type *NewRetTy = FT->getReturnType();
1508 // Check to see if we are changing the return type...
1509 if (OldRetTy != NewRetTy) {
1511 if (NewRetTy->isStructTy())
1512 return false; // TODO: Handle multiple return values.
1514 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1515 if (Callee->isDeclaration())
1516 return false; // Cannot transform this return value.
1518 if (!Caller->use_empty() &&
1519 // void -> non-void is handled specially
1520 !NewRetTy->isVoidTy())
1521 return false; // Cannot transform this return value.
1524 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1525 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1526 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
1527 return false; // Attribute not compatible with transformed value.
1530 // If the callsite is an invoke instruction, and the return value is used by
1531 // a PHI node in a successor, we cannot change the return type of the call
1532 // because there is no place to put the cast instruction (without breaking
1533 // the critical edge). Bail out in this case.
1534 if (!Caller->use_empty())
1535 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1536 for (User *U : II->users())
1537 if (PHINode *PN = dyn_cast<PHINode>(U))
1538 if (PN->getParent() == II->getNormalDest() ||
1539 PN->getParent() == II->getUnwindDest())
1543 unsigned NumActualArgs = CS.arg_size();
1544 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1546 // Prevent us turning:
1547 // declare void @takes_i32_inalloca(i32* inalloca)
1548 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1551 // call void @takes_i32_inalloca(i32* null)
1553 // Similarly, avoid folding away bitcasts of byval calls.
1554 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1555 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1558 CallSite::arg_iterator AI = CS.arg_begin();
1559 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1560 Type *ParamTy = FT->getParamType(i);
1561 Type *ActTy = (*AI)->getType();
1563 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1564 return false; // Cannot transform this parameter value.
1566 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1567 overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
1568 return false; // Attribute not compatible with transformed value.
1570 if (CS.isInAllocaArgument(i))
1571 return false; // Cannot transform to and from inalloca.
1573 // If the parameter is passed as a byval argument, then we have to have a
1574 // sized type and the sized type has to have the same size as the old type.
1575 if (ParamTy != ActTy &&
1576 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1577 Attribute::ByVal)) {
1578 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1579 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1582 Type *CurElTy = ActTy->getPointerElementType();
1583 if (DL.getTypeAllocSize(CurElTy) !=
1584 DL.getTypeAllocSize(ParamPTy->getElementType()))
1589 if (Callee->isDeclaration()) {
1590 // Do not delete arguments unless we have a function body.
1591 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1594 // If the callee is just a declaration, don't change the varargsness of the
1595 // call. We don't want to introduce a varargs call where one doesn't
1597 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1598 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1601 // If both the callee and the cast type are varargs, we still have to make
1602 // sure the number of fixed parameters are the same or we have the same
1603 // ABI issues as if we introduce a varargs call.
1604 if (FT->isVarArg() &&
1605 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1606 FT->getNumParams() !=
1607 cast<FunctionType>(APTy->getElementType())->getNumParams())
1611 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1612 !CallerPAL.isEmpty())
1613 // In this case we have more arguments than the new function type, but we
1614 // won't be dropping them. Check that these extra arguments have attributes
1615 // that are compatible with being a vararg call argument.
1616 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1617 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1618 if (Index <= FT->getNumParams())
1621 // Check if it has an attribute that's incompatible with varargs.
1622 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1623 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1628 // Okay, we decided that this is a safe thing to do: go ahead and start
1629 // inserting cast instructions as necessary.
1630 std::vector<Value*> Args;
1631 Args.reserve(NumActualArgs);
1632 SmallVector<AttributeSet, 8> attrVec;
1633 attrVec.reserve(NumCommonArgs);
1635 // Get any return attributes.
1636 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1638 // If the return value is not being used, the type may not be compatible
1639 // with the existing attributes. Wipe out any problematic attributes.
1640 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
1642 // Add the new return attributes.
1643 if (RAttrs.hasAttributes())
1644 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1645 AttributeSet::ReturnIndex, RAttrs));
1647 AI = CS.arg_begin();
1648 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1649 Type *ParamTy = FT->getParamType(i);
1651 if ((*AI)->getType() == ParamTy) {
1652 Args.push_back(*AI);
1654 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1657 // Add any parameter attributes.
1658 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1659 if (PAttrs.hasAttributes())
1660 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1664 // If the function takes more arguments than the call was taking, add them
1666 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1667 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1669 // If we are removing arguments to the function, emit an obnoxious warning.
1670 if (FT->getNumParams() < NumActualArgs) {
1671 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1672 if (FT->isVarArg()) {
1673 // Add all of the arguments in their promoted form to the arg list.
1674 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1675 Type *PTy = getPromotedType((*AI)->getType());
1676 if (PTy != (*AI)->getType()) {
1677 // Must promote to pass through va_arg area!
1678 Instruction::CastOps opcode =
1679 CastInst::getCastOpcode(*AI, false, PTy, false);
1680 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1682 Args.push_back(*AI);
1685 // Add any parameter attributes.
1686 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1687 if (PAttrs.hasAttributes())
1688 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1694 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1695 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1696 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1698 if (NewRetTy->isVoidTy())
1699 Caller->setName(""); // Void type should not have a name.
1701 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1705 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1706 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1707 II->getUnwindDest(), Args);
1709 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1710 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1712 CallInst *CI = cast<CallInst>(Caller);
1713 NC = Builder->CreateCall(Callee, Args);
1715 if (CI->isTailCall())
1716 cast<CallInst>(NC)->setTailCall();
1717 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1718 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1721 // Insert a cast of the return type as necessary.
1723 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1724 if (!NV->getType()->isVoidTy()) {
1725 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1726 NC->setDebugLoc(Caller->getDebugLoc());
1728 // If this is an invoke instruction, we should insert it after the first
1729 // non-phi, instruction in the normal successor block.
1730 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1731 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1732 InsertNewInstBefore(NC, *I);
1734 // Otherwise, it's a call, just insert cast right after the call.
1735 InsertNewInstBefore(NC, *Caller);
1737 Worklist.AddUsersToWorkList(*Caller);
1739 NV = UndefValue::get(Caller->getType());
1743 if (!Caller->use_empty())
1744 ReplaceInstUsesWith(*Caller, NV);
1745 else if (Caller->hasValueHandle()) {
1746 if (OldRetTy == NV->getType())
1747 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1749 // We cannot call ValueIsRAUWd with a different type, and the
1750 // actual tracked value will disappear.
1751 ValueHandleBase::ValueIsDeleted(Caller);
1754 EraseInstFromFunction(*Caller);
1758 // transformCallThroughTrampoline - Turn a call to a function created by
1759 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1760 // underlying function.
1763 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1764 IntrinsicInst *Tramp) {
1765 Value *Callee = CS.getCalledValue();
1766 PointerType *PTy = cast<PointerType>(Callee->getType());
1767 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1768 const AttributeSet &Attrs = CS.getAttributes();
1770 // If the call already has the 'nest' attribute somewhere then give up -
1771 // otherwise 'nest' would occur twice after splicing in the chain.
1772 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1776 "transformCallThroughTrampoline called with incorrect CallSite.");
1778 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1779 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1780 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1782 const AttributeSet &NestAttrs = NestF->getAttributes();
1783 if (!NestAttrs.isEmpty()) {
1784 unsigned NestIdx = 1;
1785 Type *NestTy = nullptr;
1786 AttributeSet NestAttr;
1788 // Look for a parameter marked with the 'nest' attribute.
1789 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1790 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1791 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1792 // Record the parameter type and any other attributes.
1794 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1799 Instruction *Caller = CS.getInstruction();
1800 std::vector<Value*> NewArgs;
1801 NewArgs.reserve(CS.arg_size() + 1);
1803 SmallVector<AttributeSet, 8> NewAttrs;
1804 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1806 // Insert the nest argument into the call argument list, which may
1807 // mean appending it. Likewise for attributes.
1809 // Add any result attributes.
1810 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1811 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1812 Attrs.getRetAttributes()));
1816 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1818 if (Idx == NestIdx) {
1819 // Add the chain argument and attributes.
1820 Value *NestVal = Tramp->getArgOperand(2);
1821 if (NestVal->getType() != NestTy)
1822 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1823 NewArgs.push_back(NestVal);
1824 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1831 // Add the original argument and attributes.
1832 NewArgs.push_back(*I);
1833 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1834 if (Attr.hasAttributes(Idx)) {
1835 AttrBuilder B(Attr, Idx);
1836 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1837 Idx + (Idx >= NestIdx), B));
1844 // Add any function attributes.
1845 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1846 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1847 Attrs.getFnAttributes()));
1849 // The trampoline may have been bitcast to a bogus type (FTy).
1850 // Handle this by synthesizing a new function type, equal to FTy
1851 // with the chain parameter inserted.
1853 std::vector<Type*> NewTypes;
1854 NewTypes.reserve(FTy->getNumParams()+1);
1856 // Insert the chain's type into the list of parameter types, which may
1857 // mean appending it.
1860 FunctionType::param_iterator I = FTy->param_begin(),
1861 E = FTy->param_end();
1865 // Add the chain's type.
1866 NewTypes.push_back(NestTy);
1871 // Add the original type.
1872 NewTypes.push_back(*I);
1878 // Replace the trampoline call with a direct call. Let the generic
1879 // code sort out any function type mismatches.
1880 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1882 Constant *NewCallee =
1883 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1884 NestF : ConstantExpr::getBitCast(NestF,
1885 PointerType::getUnqual(NewFTy));
1886 const AttributeSet &NewPAL =
1887 AttributeSet::get(FTy->getContext(), NewAttrs);
1889 Instruction *NewCaller;
1890 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1891 NewCaller = InvokeInst::Create(NewCallee,
1892 II->getNormalDest(), II->getUnwindDest(),
1894 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1895 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1897 NewCaller = CallInst::Create(NewCallee, NewArgs);
1898 if (cast<CallInst>(Caller)->isTailCall())
1899 cast<CallInst>(NewCaller)->setTailCall();
1900 cast<CallInst>(NewCaller)->
1901 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1902 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1909 // Replace the trampoline call with a direct call. Since there is no 'nest'
1910 // parameter, there is no need to adjust the argument list. Let the generic
1911 // code sort out any function type mismatches.
1912 Constant *NewCallee =
1913 NestF->getType() == PTy ? NestF :
1914 ConstantExpr::getBitCast(NestF, PTy);
1915 CS.setCalledFunction(NewCallee);
1916 return CS.getInstruction();