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 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
201 /// source vectors, unless a zero bit is set. If a zero bit is set,
202 /// then ignore that half of the mask and clear that half of the vector.
203 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
204 InstCombiner::BuilderTy &Builder) {
205 if (auto CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
206 VectorType *VecTy = cast<VectorType>(II.getType());
207 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
209 // The immediate permute control byte looks like this:
210 // [1:0] - select 128 bits from sources for low half of destination
212 // [3] - zero low half of destination
213 // [5:4] - select 128 bits from sources for high half of destination
215 // [7] - zero high half of destination
217 uint8_t Imm = CInt->getZExtValue();
219 bool LowHalfZero = Imm & 0x08;
220 bool HighHalfZero = Imm & 0x80;
222 // If both zero mask bits are set, this was just a weird way to
223 // generate a zero vector.
224 if (LowHalfZero && HighHalfZero)
227 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
228 unsigned NumElts = VecTy->getNumElements();
229 unsigned HalfSize = NumElts / 2;
230 SmallVector<int, 8> ShuffleMask(NumElts);
232 // The high bit of the selection field chooses the 1st or 2nd operand.
233 bool LowInputSelect = Imm & 0x02;
234 bool HighInputSelect = Imm & 0x20;
236 // The low bit of the selection field chooses the low or high half
237 // of the selected operand.
238 bool LowHalfSelect = Imm & 0x01;
239 bool HighHalfSelect = Imm & 0x10;
241 // Determine which operand(s) are actually in use for this instruction.
242 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
243 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
245 // If needed, replace operands based on zero mask.
246 V0 = LowHalfZero ? ZeroVector : V0;
247 V1 = HighHalfZero ? ZeroVector : V1;
249 // Permute low half of result.
250 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
251 for (unsigned i = 0; i < HalfSize; ++i)
252 ShuffleMask[i] = StartIndex + i;
254 // Permute high half of result.
255 StartIndex = HighHalfSelect ? HalfSize : 0;
256 StartIndex += NumElts;
257 for (unsigned i = 0; i < HalfSize; ++i)
258 ShuffleMask[i + HalfSize] = StartIndex + i;
260 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
265 /// visitCallInst - CallInst simplification. This mostly only handles folding
266 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
267 /// the heavy lifting.
269 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
270 if (isFreeCall(&CI, TLI))
271 return visitFree(CI);
273 // If the caller function is nounwind, mark the call as nounwind, even if the
275 if (CI.getParent()->getParent()->doesNotThrow() &&
276 !CI.doesNotThrow()) {
277 CI.setDoesNotThrow();
281 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
282 if (!II) return visitCallSite(&CI);
284 // Intrinsics cannot occur in an invoke, so handle them here instead of in
286 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
287 bool Changed = false;
289 // memmove/cpy/set of zero bytes is a noop.
290 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
291 if (NumBytes->isNullValue())
292 return EraseInstFromFunction(CI);
294 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
295 if (CI->getZExtValue() == 1) {
296 // Replace the instruction with just byte operations. We would
297 // transform other cases to loads/stores, but we don't know if
298 // alignment is sufficient.
302 // No other transformations apply to volatile transfers.
303 if (MI->isVolatile())
306 // If we have a memmove and the source operation is a constant global,
307 // then the source and dest pointers can't alias, so we can change this
308 // into a call to memcpy.
309 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
310 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
311 if (GVSrc->isConstant()) {
312 Module *M = CI.getParent()->getParent()->getParent();
313 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
314 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
315 CI.getArgOperand(1)->getType(),
316 CI.getArgOperand(2)->getType() };
317 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
322 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
323 // memmove(x,x,size) -> noop.
324 if (MTI->getSource() == MTI->getDest())
325 return EraseInstFromFunction(CI);
328 // If we can determine a pointer alignment that is bigger than currently
329 // set, update the alignment.
330 if (isa<MemTransferInst>(MI)) {
331 if (Instruction *I = SimplifyMemTransfer(MI))
333 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
334 if (Instruction *I = SimplifyMemSet(MSI))
338 if (Changed) return II;
341 switch (II->getIntrinsicID()) {
343 case Intrinsic::objectsize: {
345 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
346 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
349 case Intrinsic::bswap: {
350 Value *IIOperand = II->getArgOperand(0);
353 // bswap(bswap(x)) -> x
354 if (match(IIOperand, m_BSwap(m_Value(X))))
355 return ReplaceInstUsesWith(CI, X);
357 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
358 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
359 unsigned C = X->getType()->getPrimitiveSizeInBits() -
360 IIOperand->getType()->getPrimitiveSizeInBits();
361 Value *CV = ConstantInt::get(X->getType(), C);
362 Value *V = Builder->CreateLShr(X, CV);
363 return new TruncInst(V, IIOperand->getType());
368 case Intrinsic::powi:
369 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
372 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
375 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
376 // powi(x, -1) -> 1/x
377 if (Power->isAllOnesValue())
378 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
379 II->getArgOperand(0));
382 case Intrinsic::cttz: {
383 // If all bits below the first known one are known zero,
384 // this value is constant.
385 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
386 // FIXME: Try to simplify vectors of integers.
388 uint32_t BitWidth = IT->getBitWidth();
389 APInt KnownZero(BitWidth, 0);
390 APInt KnownOne(BitWidth, 0);
391 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
392 unsigned TrailingZeros = KnownOne.countTrailingZeros();
393 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
394 if ((Mask & KnownZero) == Mask)
395 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
396 APInt(BitWidth, TrailingZeros)));
400 case Intrinsic::ctlz: {
401 // If all bits above the first known one are known zero,
402 // this value is constant.
403 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
404 // FIXME: Try to simplify vectors of integers.
406 uint32_t BitWidth = IT->getBitWidth();
407 APInt KnownZero(BitWidth, 0);
408 APInt KnownOne(BitWidth, 0);
409 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
410 unsigned LeadingZeros = KnownOne.countLeadingZeros();
411 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
412 if ((Mask & KnownZero) == Mask)
413 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
414 APInt(BitWidth, LeadingZeros)));
419 case Intrinsic::uadd_with_overflow: // FALLTHROUGH
420 case Intrinsic::sadd_with_overflow: // FALLTHROUGH
421 case Intrinsic::usub_with_overflow: // FALLTHROUGH
422 case Intrinsic::ssub_with_overflow: // FALLTHROUGH
423 case Intrinsic::umul_with_overflow: // FALLTHROUGH
424 case Intrinsic::smul_with_overflow: {
425 if (isa<Constant>(II->getArgOperand(0)) &&
426 !isa<Constant>(II->getArgOperand(1))) {
427 // Canonicalize constants into the RHS.
428 Value *LHS = II->getArgOperand(0);
429 II->setArgOperand(0, II->getArgOperand(1));
430 II->setArgOperand(1, LHS);
434 OverflowCheckFlavor OCF =
435 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID());
436 assert(OCF != OCF_INVALID && "unexpected!");
438 Value *OperationResult = nullptr;
439 Constant *OverflowResult = nullptr;
440 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
441 *II, OperationResult, OverflowResult))
442 return CreateOverflowTuple(II, OperationResult, OverflowResult);
447 case Intrinsic::minnum:
448 case Intrinsic::maxnum: {
449 Value *Arg0 = II->getArgOperand(0);
450 Value *Arg1 = II->getArgOperand(1);
454 return ReplaceInstUsesWith(CI, Arg0);
456 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
457 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
459 // Canonicalize constants into the RHS.
461 II->setArgOperand(0, Arg1);
462 II->setArgOperand(1, Arg0);
467 if (C1 && C1->isNaN())
468 return ReplaceInstUsesWith(CI, Arg0);
470 // This is the value because if undef were NaN, we would return the other
471 // value and cannot return a NaN unless both operands are.
473 // fmin(undef, x) -> x
474 if (isa<UndefValue>(Arg0))
475 return ReplaceInstUsesWith(CI, Arg1);
477 // fmin(x, undef) -> x
478 if (isa<UndefValue>(Arg1))
479 return ReplaceInstUsesWith(CI, Arg0);
483 if (II->getIntrinsicID() == Intrinsic::minnum) {
484 // fmin(x, fmin(x, y)) -> fmin(x, y)
485 // fmin(y, fmin(x, y)) -> fmin(x, y)
486 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
487 if (Arg0 == X || Arg0 == Y)
488 return ReplaceInstUsesWith(CI, Arg1);
491 // fmin(fmin(x, y), x) -> fmin(x, y)
492 // fmin(fmin(x, y), y) -> fmin(x, y)
493 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
494 if (Arg1 == X || Arg1 == Y)
495 return ReplaceInstUsesWith(CI, Arg0);
498 // TODO: fmin(nnan x, inf) -> x
499 // TODO: fmin(nnan ninf x, flt_max) -> x
500 if (C1 && C1->isInfinity()) {
501 // fmin(x, -inf) -> -inf
502 if (C1->isNegative())
503 return ReplaceInstUsesWith(CI, Arg1);
506 assert(II->getIntrinsicID() == Intrinsic::maxnum);
507 // fmax(x, fmax(x, y)) -> fmax(x, y)
508 // fmax(y, fmax(x, y)) -> fmax(x, y)
509 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
510 if (Arg0 == X || Arg0 == Y)
511 return ReplaceInstUsesWith(CI, Arg1);
514 // fmax(fmax(x, y), x) -> fmax(x, y)
515 // fmax(fmax(x, y), y) -> fmax(x, y)
516 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
517 if (Arg1 == X || Arg1 == Y)
518 return ReplaceInstUsesWith(CI, Arg0);
521 // TODO: fmax(nnan x, -inf) -> x
522 // TODO: fmax(nnan ninf x, -flt_max) -> x
523 if (C1 && C1->isInfinity()) {
524 // fmax(x, inf) -> inf
525 if (!C1->isNegative())
526 return ReplaceInstUsesWith(CI, Arg1);
531 case Intrinsic::ppc_altivec_lvx:
532 case Intrinsic::ppc_altivec_lvxl:
533 // Turn PPC lvx -> load if the pointer is known aligned.
534 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
536 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
537 PointerType::getUnqual(II->getType()));
538 return new LoadInst(Ptr);
541 case Intrinsic::ppc_vsx_lxvw4x:
542 case Intrinsic::ppc_vsx_lxvd2x: {
543 // Turn PPC VSX loads into normal loads.
544 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
545 PointerType::getUnqual(II->getType()));
546 return new LoadInst(Ptr, Twine(""), false, 1);
548 case Intrinsic::ppc_altivec_stvx:
549 case Intrinsic::ppc_altivec_stvxl:
550 // Turn stvx -> store if the pointer is known aligned.
551 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
554 PointerType::getUnqual(II->getArgOperand(0)->getType());
555 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
556 return new StoreInst(II->getArgOperand(0), Ptr);
559 case Intrinsic::ppc_vsx_stxvw4x:
560 case Intrinsic::ppc_vsx_stxvd2x: {
561 // Turn PPC VSX stores into normal stores.
562 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
563 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
564 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
566 case Intrinsic::ppc_qpx_qvlfs:
567 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
568 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
570 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
571 PointerType::getUnqual(II->getType()));
572 return new LoadInst(Ptr);
575 case Intrinsic::ppc_qpx_qvlfd:
576 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
577 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
579 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
580 PointerType::getUnqual(II->getType()));
581 return new LoadInst(Ptr);
584 case Intrinsic::ppc_qpx_qvstfs:
585 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
586 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
589 PointerType::getUnqual(II->getArgOperand(0)->getType());
590 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
591 return new StoreInst(II->getArgOperand(0), Ptr);
594 case Intrinsic::ppc_qpx_qvstfd:
595 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
596 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
599 PointerType::getUnqual(II->getArgOperand(0)->getType());
600 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
601 return new StoreInst(II->getArgOperand(0), Ptr);
604 case Intrinsic::x86_sse_storeu_ps:
605 case Intrinsic::x86_sse2_storeu_pd:
606 case Intrinsic::x86_sse2_storeu_dq:
607 // Turn X86 storeu -> store if the pointer is known aligned.
608 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
611 PointerType::getUnqual(II->getArgOperand(1)->getType());
612 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
613 return new StoreInst(II->getArgOperand(1), Ptr);
617 case Intrinsic::x86_sse_cvtss2si:
618 case Intrinsic::x86_sse_cvtss2si64:
619 case Intrinsic::x86_sse_cvttss2si:
620 case Intrinsic::x86_sse_cvttss2si64:
621 case Intrinsic::x86_sse2_cvtsd2si:
622 case Intrinsic::x86_sse2_cvtsd2si64:
623 case Intrinsic::x86_sse2_cvttsd2si:
624 case Intrinsic::x86_sse2_cvttsd2si64: {
625 // These intrinsics only demand the 0th element of their input vectors. If
626 // we can simplify the input based on that, do so now.
628 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
629 APInt DemandedElts(VWidth, 1);
630 APInt UndefElts(VWidth, 0);
631 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
632 DemandedElts, UndefElts)) {
633 II->setArgOperand(0, V);
639 // Constant fold <A x Bi> << Ci.
640 // FIXME: We don't handle _dq because it's a shift of an i128, but is
641 // represented in the IR as <2 x i64>. A per element shift is wrong.
642 case Intrinsic::x86_sse2_psll_d:
643 case Intrinsic::x86_sse2_psll_q:
644 case Intrinsic::x86_sse2_psll_w:
645 case Intrinsic::x86_sse2_pslli_d:
646 case Intrinsic::x86_sse2_pslli_q:
647 case Intrinsic::x86_sse2_pslli_w:
648 case Intrinsic::x86_avx2_psll_d:
649 case Intrinsic::x86_avx2_psll_q:
650 case Intrinsic::x86_avx2_psll_w:
651 case Intrinsic::x86_avx2_pslli_d:
652 case Intrinsic::x86_avx2_pslli_q:
653 case Intrinsic::x86_avx2_pslli_w:
654 case Intrinsic::x86_sse2_psrl_d:
655 case Intrinsic::x86_sse2_psrl_q:
656 case Intrinsic::x86_sse2_psrl_w:
657 case Intrinsic::x86_sse2_psrli_d:
658 case Intrinsic::x86_sse2_psrli_q:
659 case Intrinsic::x86_sse2_psrli_w:
660 case Intrinsic::x86_avx2_psrl_d:
661 case Intrinsic::x86_avx2_psrl_q:
662 case Intrinsic::x86_avx2_psrl_w:
663 case Intrinsic::x86_avx2_psrli_d:
664 case Intrinsic::x86_avx2_psrli_q:
665 case Intrinsic::x86_avx2_psrli_w: {
666 // Simplify if count is constant. To 0 if >= BitWidth,
667 // otherwise to shl/lshr.
668 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
669 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
674 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
678 auto Vec = II->getArgOperand(0);
679 auto VT = cast<VectorType>(Vec->getType());
680 if (Count->getZExtValue() >
681 VT->getElementType()->getPrimitiveSizeInBits() - 1)
682 return ReplaceInstUsesWith(
683 CI, ConstantAggregateZero::get(Vec->getType()));
685 bool isPackedShiftLeft = true;
686 switch (II->getIntrinsicID()) {
688 case Intrinsic::x86_sse2_psrl_d:
689 case Intrinsic::x86_sse2_psrl_q:
690 case Intrinsic::x86_sse2_psrl_w:
691 case Intrinsic::x86_sse2_psrli_d:
692 case Intrinsic::x86_sse2_psrli_q:
693 case Intrinsic::x86_sse2_psrli_w:
694 case Intrinsic::x86_avx2_psrl_d:
695 case Intrinsic::x86_avx2_psrl_q:
696 case Intrinsic::x86_avx2_psrl_w:
697 case Intrinsic::x86_avx2_psrli_d:
698 case Intrinsic::x86_avx2_psrli_q:
699 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
702 unsigned VWidth = VT->getNumElements();
703 // Get a constant vector of the same type as the first operand.
704 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
705 if (isPackedShiftLeft)
706 return BinaryOperator::CreateShl(Vec,
707 Builder->CreateVectorSplat(VWidth, VTCI));
709 return BinaryOperator::CreateLShr(Vec,
710 Builder->CreateVectorSplat(VWidth, VTCI));
713 case Intrinsic::x86_sse41_pmovsxbw:
714 case Intrinsic::x86_sse41_pmovsxwd:
715 case Intrinsic::x86_sse41_pmovsxdq:
716 case Intrinsic::x86_sse41_pmovzxbw:
717 case Intrinsic::x86_sse41_pmovzxwd:
718 case Intrinsic::x86_sse41_pmovzxdq: {
719 // pmov{s|z}x ignores the upper half of their input vectors.
721 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
722 unsigned LowHalfElts = VWidth / 2;
723 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
724 APInt UndefElts(VWidth, 0);
725 if (Value *TmpV = SimplifyDemandedVectorElts(
726 II->getArgOperand(0), InputDemandedElts, UndefElts)) {
727 II->setArgOperand(0, TmpV);
733 case Intrinsic::x86_sse4a_insertqi: {
734 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
736 // TODO: eventually we should lower this intrinsic to IR
737 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
738 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
739 unsigned Index = CIStart->getZExtValue();
740 // From AMD documentation: "a value of zero in the field length is
741 // defined as length of 64".
742 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
744 // From AMD documentation: "If the sum of the bit index + length field
745 // is greater than 64, the results are undefined".
747 // Note that both field index and field length are 8-bit quantities.
748 // Since variables 'Index' and 'Length' are unsigned values
749 // obtained from zero-extending field index and field length
750 // respectively, their sum should never wrap around.
751 if ((Index + Length) > 64)
752 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
754 if (Length == 64 && Index == 0) {
755 Value *Vec = II->getArgOperand(1);
756 Value *Undef = UndefValue::get(Vec->getType());
757 const uint32_t Mask[] = { 0, 2 };
758 return ReplaceInstUsesWith(
760 Builder->CreateShuffleVector(
761 Vec, Undef, ConstantDataVector::get(
762 II->getContext(), makeArrayRef(Mask))));
764 } else if (auto Source =
765 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
766 if (Source->hasOneUse() &&
767 Source->getArgOperand(1) == II->getArgOperand(1)) {
768 // If the source of the insert has only one use and it's another
769 // insert (and they're both inserting from the same vector), try to
770 // bundle both together.
772 dyn_cast<ConstantInt>(Source->getArgOperand(2));
774 dyn_cast<ConstantInt>(Source->getArgOperand(3));
775 if (CISourceStart && CISourceWidth) {
776 unsigned Start = CIStart->getZExtValue();
777 unsigned Width = CIWidth->getZExtValue();
778 unsigned End = Start + Width;
779 unsigned SourceStart = CISourceStart->getZExtValue();
780 unsigned SourceWidth = CISourceWidth->getZExtValue();
781 unsigned SourceEnd = SourceStart + SourceWidth;
782 unsigned NewStart, NewWidth;
783 bool ShouldReplace = false;
784 if (Start <= SourceStart && SourceStart <= End) {
786 NewWidth = std::max(End, SourceEnd) - NewStart;
787 ShouldReplace = true;
788 } else if (SourceStart <= Start && Start <= SourceEnd) {
789 NewStart = SourceStart;
790 NewWidth = std::max(SourceEnd, End) - NewStart;
791 ShouldReplace = true;
795 Constant *ConstantWidth = ConstantInt::get(
796 II->getArgOperand(2)->getType(), NewWidth, false);
797 Constant *ConstantStart = ConstantInt::get(
798 II->getArgOperand(3)->getType(), NewStart, false);
799 Value *Args[4] = { Source->getArgOperand(0),
800 II->getArgOperand(1), ConstantWidth,
802 Module *M = CI.getParent()->getParent()->getParent();
804 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
805 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
815 case Intrinsic::x86_sse41_pblendvb:
816 case Intrinsic::x86_sse41_blendvps:
817 case Intrinsic::x86_sse41_blendvpd:
818 case Intrinsic::x86_avx_blendv_ps_256:
819 case Intrinsic::x86_avx_blendv_pd_256:
820 case Intrinsic::x86_avx2_pblendvb: {
821 // Convert blendv* to vector selects if the mask is constant.
822 // This optimization is convoluted because the intrinsic is defined as
823 // getting a vector of floats or doubles for the ps and pd versions.
824 // FIXME: That should be changed.
825 Value *Mask = II->getArgOperand(2);
826 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
827 auto Tyi1 = Builder->getInt1Ty();
828 auto SelectorType = cast<VectorType>(Mask->getType());
829 auto EltTy = SelectorType->getElementType();
830 unsigned Size = SelectorType->getNumElements();
834 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
835 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
836 "Wrong arguments for variable blend intrinsic");
837 SmallVector<Constant *, 32> Selectors;
838 for (unsigned I = 0; I < Size; ++I) {
839 // The intrinsics only read the top bit
842 Selector = C->getElementAsInteger(I);
844 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
845 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
847 auto NewSelector = ConstantVector::get(Selectors);
848 return SelectInst::Create(NewSelector, II->getArgOperand(1),
849 II->getArgOperand(0), "blendv");
855 case Intrinsic::x86_avx_vpermilvar_ps:
856 case Intrinsic::x86_avx_vpermilvar_ps_256:
857 case Intrinsic::x86_avx_vpermilvar_pd:
858 case Intrinsic::x86_avx_vpermilvar_pd_256: {
859 // Convert vpermil* to shufflevector if the mask is constant.
860 Value *V = II->getArgOperand(1);
861 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
862 assert(Size == 8 || Size == 4 || Size == 2);
864 if (auto C = dyn_cast<ConstantDataVector>(V)) {
865 // The intrinsics only read one or two bits, clear the rest.
866 for (unsigned I = 0; I < Size; ++I) {
867 uint32_t Index = C->getElementAsInteger(I) & 0x3;
868 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
869 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
873 } else if (isa<ConstantAggregateZero>(V)) {
874 for (unsigned I = 0; I < Size; ++I)
879 // The _256 variants are a bit trickier since the mask bits always index
880 // into the corresponding 128 half. In order to convert to a generic
881 // shuffle, we have to make that explicit.
882 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
883 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
884 for (unsigned I = Size / 2; I < Size; ++I)
885 Indexes[I] += Size / 2;
888 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
889 auto V1 = II->getArgOperand(0);
890 auto V2 = UndefValue::get(V1->getType());
891 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
892 return ReplaceInstUsesWith(CI, Shuffle);
895 case Intrinsic::x86_avx_vperm2f128_pd_256:
896 case Intrinsic::x86_avx_vperm2f128_ps_256:
897 case Intrinsic::x86_avx_vperm2f128_si_256:
898 case Intrinsic::x86_avx2_vperm2i128:
899 if (Value *V = SimplifyX86vperm2(*II, *Builder))
900 return ReplaceInstUsesWith(*II, V);
903 case Intrinsic::ppc_altivec_vperm:
904 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
905 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
906 // a vectorshuffle for little endian, we must undo the transformation
907 // performed on vec_perm in altivec.h. That is, we must complement
908 // the permutation mask with respect to 31 and reverse the order of
910 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
911 assert(Mask->getType()->getVectorNumElements() == 16 &&
912 "Bad type for intrinsic!");
914 // Check that all of the elements are integer constants or undefs.
915 bool AllEltsOk = true;
916 for (unsigned i = 0; i != 16; ++i) {
917 Constant *Elt = Mask->getAggregateElement(i);
918 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
925 // Cast the input vectors to byte vectors.
926 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
928 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
930 Value *Result = UndefValue::get(Op0->getType());
932 // Only extract each element once.
933 Value *ExtractedElts[32];
934 memset(ExtractedElts, 0, sizeof(ExtractedElts));
936 for (unsigned i = 0; i != 16; ++i) {
937 if (isa<UndefValue>(Mask->getAggregateElement(i)))
940 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
941 Idx &= 31; // Match the hardware behavior.
942 if (DL.isLittleEndian())
945 if (!ExtractedElts[Idx]) {
946 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
947 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
949 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
950 Builder->getInt32(Idx&15));
953 // Insert this value into the result vector.
954 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
955 Builder->getInt32(i));
957 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
962 case Intrinsic::arm_neon_vld1:
963 case Intrinsic::arm_neon_vld2:
964 case Intrinsic::arm_neon_vld3:
965 case Intrinsic::arm_neon_vld4:
966 case Intrinsic::arm_neon_vld2lane:
967 case Intrinsic::arm_neon_vld3lane:
968 case Intrinsic::arm_neon_vld4lane:
969 case Intrinsic::arm_neon_vst1:
970 case Intrinsic::arm_neon_vst2:
971 case Intrinsic::arm_neon_vst3:
972 case Intrinsic::arm_neon_vst4:
973 case Intrinsic::arm_neon_vst2lane:
974 case Intrinsic::arm_neon_vst3lane:
975 case Intrinsic::arm_neon_vst4lane: {
976 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
977 unsigned AlignArg = II->getNumArgOperands() - 1;
978 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
979 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
980 II->setArgOperand(AlignArg,
981 ConstantInt::get(Type::getInt32Ty(II->getContext()),
988 case Intrinsic::arm_neon_vmulls:
989 case Intrinsic::arm_neon_vmullu:
990 case Intrinsic::aarch64_neon_smull:
991 case Intrinsic::aarch64_neon_umull: {
992 Value *Arg0 = II->getArgOperand(0);
993 Value *Arg1 = II->getArgOperand(1);
995 // Handle mul by zero first:
996 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
997 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1000 // Check for constant LHS & RHS - in this case we just simplify.
1001 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1002 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1003 VectorType *NewVT = cast<VectorType>(II->getType());
1004 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1005 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1006 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1007 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1009 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1012 // Couldn't simplify - canonicalize constant to the RHS.
1013 std::swap(Arg0, Arg1);
1016 // Handle mul by one:
1017 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1018 if (ConstantInt *Splat =
1019 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1021 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1022 /*isSigned=*/!Zext);
1027 case Intrinsic::AMDGPU_rcp: {
1028 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1029 const APFloat &ArgVal = C->getValueAPF();
1030 APFloat Val(ArgVal.getSemantics(), 1.0);
1031 APFloat::opStatus Status = Val.divide(ArgVal,
1032 APFloat::rmNearestTiesToEven);
1033 // Only do this if it was exact and therefore not dependent on the
1035 if (Status == APFloat::opOK)
1036 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1041 case Intrinsic::stackrestore: {
1042 // If the save is right next to the restore, remove the restore. This can
1043 // happen when variable allocas are DCE'd.
1044 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1045 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1046 BasicBlock::iterator BI = SS;
1048 return EraseInstFromFunction(CI);
1052 // Scan down this block to see if there is another stack restore in the
1053 // same block without an intervening call/alloca.
1054 BasicBlock::iterator BI = II;
1055 TerminatorInst *TI = II->getParent()->getTerminator();
1056 bool CannotRemove = false;
1057 for (++BI; &*BI != TI; ++BI) {
1058 if (isa<AllocaInst>(BI)) {
1059 CannotRemove = true;
1062 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1063 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1064 // If there is a stackrestore below this one, remove this one.
1065 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1066 return EraseInstFromFunction(CI);
1067 // Otherwise, ignore the intrinsic.
1069 // If we found a non-intrinsic call, we can't remove the stack
1071 CannotRemove = true;
1077 // If the stack restore is in a return, resume, or unwind block and if there
1078 // are no allocas or calls between the restore and the return, nuke the
1080 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1081 return EraseInstFromFunction(CI);
1084 case Intrinsic::assume: {
1085 // Canonicalize assume(a && b) -> assume(a); assume(b);
1086 // Note: New assumption intrinsics created here are registered by
1087 // the InstCombineIRInserter object.
1088 Value *IIOperand = II->getArgOperand(0), *A, *B,
1089 *AssumeIntrinsic = II->getCalledValue();
1090 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1091 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1092 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1093 return EraseInstFromFunction(*II);
1095 // assume(!(a || b)) -> assume(!a); assume(!b);
1096 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1097 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1099 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1101 return EraseInstFromFunction(*II);
1104 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1105 // (if assume is valid at the load)
1106 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1107 Value *LHS = ICmp->getOperand(0);
1108 Value *RHS = ICmp->getOperand(1);
1109 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1110 isa<LoadInst>(LHS) &&
1111 isa<Constant>(RHS) &&
1112 RHS->getType()->isPointerTy() &&
1113 cast<Constant>(RHS)->isNullValue()) {
1114 LoadInst* LI = cast<LoadInst>(LHS);
1115 if (isValidAssumeForContext(II, LI, DT)) {
1116 MDNode *MD = MDNode::get(II->getContext(), None);
1117 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1118 return EraseInstFromFunction(*II);
1121 // TODO: apply nonnull return attributes to calls and invokes
1122 // TODO: apply range metadata for range check patterns?
1124 // If there is a dominating assume with the same condition as this one,
1125 // then this one is redundant, and should be removed.
1126 APInt KnownZero(1, 0), KnownOne(1, 0);
1127 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1128 if (KnownOne.isAllOnesValue())
1129 return EraseInstFromFunction(*II);
1133 case Intrinsic::experimental_gc_relocate: {
1134 // Translate facts known about a pointer before relocating into
1135 // facts about the relocate value, while being careful to
1136 // preserve relocation semantics.
1137 GCRelocateOperands Operands(II);
1138 Value *DerivedPtr = Operands.derivedPtr();
1140 // Remove the relocation if unused, note that this check is required
1141 // to prevent the cases below from looping forever.
1142 if (II->use_empty())
1143 return EraseInstFromFunction(*II);
1145 // Undef is undef, even after relocation.
1146 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1147 // most practical collectors, but there was discussion in the review thread
1148 // about whether it was legal for all possible collectors.
1149 if (isa<UndefValue>(DerivedPtr))
1150 return ReplaceInstUsesWith(*II, DerivedPtr);
1152 // The relocation of null will be null for most any collector.
1153 // TODO: provide a hook for this in GCStrategy. There might be some weird
1154 // collector this property does not hold for.
1155 if (isa<ConstantPointerNull>(DerivedPtr))
1156 return ReplaceInstUsesWith(*II, DerivedPtr);
1158 // isKnownNonNull -> nonnull attribute
1159 if (isKnownNonNull(DerivedPtr))
1160 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1162 // isDereferenceablePointer -> deref attribute
1163 if (DerivedPtr->isDereferenceablePointer(DL)) {
1164 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1165 uint64_t Bytes = A->getDereferenceableBytes();
1166 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1170 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1171 // Canonicalize on the type from the uses to the defs
1173 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1177 return visitCallSite(II);
1180 // InvokeInst simplification
1182 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1183 return visitCallSite(&II);
1186 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1187 /// passed through the varargs area, we can eliminate the use of the cast.
1188 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1189 const DataLayout &DL,
1190 const CastInst *const CI,
1192 if (!CI->isLosslessCast())
1195 // If this is a GC intrinsic, avoid munging types. We need types for
1196 // statepoint reconstruction in SelectionDAG.
1197 // TODO: This is probably something which should be expanded to all
1198 // intrinsics since the entire point of intrinsics is that
1199 // they are understandable by the optimizer.
1200 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1203 // The size of ByVal or InAlloca arguments is derived from the type, so we
1204 // can't change to a type with a different size. If the size were
1205 // passed explicitly we could avoid this check.
1206 if (!CS.isByValOrInAllocaArgument(ix))
1210 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1211 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1212 if (!SrcTy->isSized() || !DstTy->isSized())
1214 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1219 // Try to fold some different type of calls here.
1220 // Currently we're only working with the checking functions, memcpy_chk,
1221 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1222 // strcat_chk and strncat_chk.
1223 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1224 if (!CI->getCalledFunction()) return nullptr;
1226 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1227 ReplaceInstUsesWith(*From, With);
1229 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1230 if (Value *With = Simplifier.optimizeCall(CI)) {
1232 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1238 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1239 // Strip off at most one level of pointer casts, looking for an alloca. This
1240 // is good enough in practice and simpler than handling any number of casts.
1241 Value *Underlying = TrampMem->stripPointerCasts();
1242 if (Underlying != TrampMem &&
1243 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1245 if (!isa<AllocaInst>(Underlying))
1248 IntrinsicInst *InitTrampoline = nullptr;
1249 for (User *U : TrampMem->users()) {
1250 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1253 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1255 // More than one init_trampoline writes to this value. Give up.
1257 InitTrampoline = II;
1260 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1261 // Allow any number of calls to adjust.trampoline.
1266 // No call to init.trampoline found.
1267 if (!InitTrampoline)
1270 // Check that the alloca is being used in the expected way.
1271 if (InitTrampoline->getOperand(0) != TrampMem)
1274 return InitTrampoline;
1277 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1279 // Visit all the previous instructions in the basic block, and try to find a
1280 // init.trampoline which has a direct path to the adjust.trampoline.
1281 for (BasicBlock::iterator I = AdjustTramp,
1282 E = AdjustTramp->getParent()->begin(); I != E; ) {
1283 Instruction *Inst = --I;
1284 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1285 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1286 II->getOperand(0) == TrampMem)
1288 if (Inst->mayWriteToMemory())
1294 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1295 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1296 // to a direct call to a function. Otherwise return NULL.
1298 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1299 Callee = Callee->stripPointerCasts();
1300 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1302 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1305 Value *TrampMem = AdjustTramp->getOperand(0);
1307 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1309 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1314 // visitCallSite - Improvements for call and invoke instructions.
1316 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1317 if (isAllocLikeFn(CS.getInstruction(), TLI))
1318 return visitAllocSite(*CS.getInstruction());
1320 bool Changed = false;
1322 // If the callee is a pointer to a function, attempt to move any casts to the
1323 // arguments of the call/invoke.
1324 Value *Callee = CS.getCalledValue();
1325 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1328 if (Function *CalleeF = dyn_cast<Function>(Callee))
1329 // If the call and callee calling conventions don't match, this call must
1330 // be unreachable, as the call is undefined.
1331 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1332 // Only do this for calls to a function with a body. A prototype may
1333 // not actually end up matching the implementation's calling conv for a
1334 // variety of reasons (e.g. it may be written in assembly).
1335 !CalleeF->isDeclaration()) {
1336 Instruction *OldCall = CS.getInstruction();
1337 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1338 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1340 // If OldCall does not return void then replaceAllUsesWith undef.
1341 // This allows ValueHandlers and custom metadata to adjust itself.
1342 if (!OldCall->getType()->isVoidTy())
1343 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1344 if (isa<CallInst>(OldCall))
1345 return EraseInstFromFunction(*OldCall);
1347 // We cannot remove an invoke, because it would change the CFG, just
1348 // change the callee to a null pointer.
1349 cast<InvokeInst>(OldCall)->setCalledFunction(
1350 Constant::getNullValue(CalleeF->getType()));
1354 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1355 // If CS does not return void then replaceAllUsesWith undef.
1356 // This allows ValueHandlers and custom metadata to adjust itself.
1357 if (!CS.getInstruction()->getType()->isVoidTy())
1358 ReplaceInstUsesWith(*CS.getInstruction(),
1359 UndefValue::get(CS.getInstruction()->getType()));
1361 if (isa<InvokeInst>(CS.getInstruction())) {
1362 // Can't remove an invoke because we cannot change the CFG.
1366 // This instruction is not reachable, just remove it. We insert a store to
1367 // undef so that we know that this code is not reachable, despite the fact
1368 // that we can't modify the CFG here.
1369 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1370 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1371 CS.getInstruction());
1373 return EraseInstFromFunction(*CS.getInstruction());
1376 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1377 return transformCallThroughTrampoline(CS, II);
1379 PointerType *PTy = cast<PointerType>(Callee->getType());
1380 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1381 if (FTy->isVarArg()) {
1382 int ix = FTy->getNumParams();
1383 // See if we can optimize any arguments passed through the varargs area of
1385 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1386 E = CS.arg_end(); I != E; ++I, ++ix) {
1387 CastInst *CI = dyn_cast<CastInst>(*I);
1388 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1389 *I = CI->getOperand(0);
1395 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1396 // Inline asm calls cannot throw - mark them 'nounwind'.
1397 CS.setDoesNotThrow();
1401 // Try to optimize the call if possible, we require DataLayout for most of
1402 // this. None of these calls are seen as possibly dead so go ahead and
1403 // delete the instruction now.
1404 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1405 Instruction *I = tryOptimizeCall(CI);
1406 // If we changed something return the result, etc. Otherwise let
1407 // the fallthrough check.
1408 if (I) return EraseInstFromFunction(*I);
1411 return Changed ? CS.getInstruction() : nullptr;
1414 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1415 // attempt to move the cast to the arguments of the call/invoke.
1417 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1419 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1422 // The prototype of thunks are a lie, don't try to directly call such
1424 if (Callee->hasFnAttribute("thunk"))
1426 Instruction *Caller = CS.getInstruction();
1427 const AttributeSet &CallerPAL = CS.getAttributes();
1429 // Okay, this is a cast from a function to a different type. Unless doing so
1430 // would cause a type conversion of one of our arguments, change this call to
1431 // be a direct call with arguments casted to the appropriate types.
1433 FunctionType *FT = Callee->getFunctionType();
1434 Type *OldRetTy = Caller->getType();
1435 Type *NewRetTy = FT->getReturnType();
1437 // Check to see if we are changing the return type...
1438 if (OldRetTy != NewRetTy) {
1440 if (NewRetTy->isStructTy())
1441 return false; // TODO: Handle multiple return values.
1443 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1444 if (Callee->isDeclaration())
1445 return false; // Cannot transform this return value.
1447 if (!Caller->use_empty() &&
1448 // void -> non-void is handled specially
1449 !NewRetTy->isVoidTy())
1450 return false; // Cannot transform this return value.
1453 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1454 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1456 hasAttributes(AttributeFuncs::
1457 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1458 AttributeSet::ReturnIndex))
1459 return false; // Attribute not compatible with transformed value.
1462 // If the callsite is an invoke instruction, and the return value is used by
1463 // a PHI node in a successor, we cannot change the return type of the call
1464 // because there is no place to put the cast instruction (without breaking
1465 // the critical edge). Bail out in this case.
1466 if (!Caller->use_empty())
1467 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1468 for (User *U : II->users())
1469 if (PHINode *PN = dyn_cast<PHINode>(U))
1470 if (PN->getParent() == II->getNormalDest() ||
1471 PN->getParent() == II->getUnwindDest())
1475 unsigned NumActualArgs = CS.arg_size();
1476 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1478 // Prevent us turning:
1479 // declare void @takes_i32_inalloca(i32* inalloca)
1480 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1483 // call void @takes_i32_inalloca(i32* null)
1485 // Similarly, avoid folding away bitcasts of byval calls.
1486 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1487 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1490 CallSite::arg_iterator AI = CS.arg_begin();
1491 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1492 Type *ParamTy = FT->getParamType(i);
1493 Type *ActTy = (*AI)->getType();
1495 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1496 return false; // Cannot transform this parameter value.
1498 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1499 hasAttributes(AttributeFuncs::
1500 typeIncompatible(ParamTy, i + 1), i + 1))
1501 return false; // Attribute not compatible with transformed value.
1503 if (CS.isInAllocaArgument(i))
1504 return false; // Cannot transform to and from inalloca.
1506 // If the parameter is passed as a byval argument, then we have to have a
1507 // sized type and the sized type has to have the same size as the old type.
1508 if (ParamTy != ActTy &&
1509 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1510 Attribute::ByVal)) {
1511 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1512 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1515 Type *CurElTy = ActTy->getPointerElementType();
1516 if (DL.getTypeAllocSize(CurElTy) !=
1517 DL.getTypeAllocSize(ParamPTy->getElementType()))
1522 if (Callee->isDeclaration()) {
1523 // Do not delete arguments unless we have a function body.
1524 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1527 // If the callee is just a declaration, don't change the varargsness of the
1528 // call. We don't want to introduce a varargs call where one doesn't
1530 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1531 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1534 // If both the callee and the cast type are varargs, we still have to make
1535 // sure the number of fixed parameters are the same or we have the same
1536 // ABI issues as if we introduce a varargs call.
1537 if (FT->isVarArg() &&
1538 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1539 FT->getNumParams() !=
1540 cast<FunctionType>(APTy->getElementType())->getNumParams())
1544 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1545 !CallerPAL.isEmpty())
1546 // In this case we have more arguments than the new function type, but we
1547 // won't be dropping them. Check that these extra arguments have attributes
1548 // that are compatible with being a vararg call argument.
1549 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1550 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1551 if (Index <= FT->getNumParams())
1554 // Check if it has an attribute that's incompatible with varargs.
1555 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1556 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1561 // Okay, we decided that this is a safe thing to do: go ahead and start
1562 // inserting cast instructions as necessary.
1563 std::vector<Value*> Args;
1564 Args.reserve(NumActualArgs);
1565 SmallVector<AttributeSet, 8> attrVec;
1566 attrVec.reserve(NumCommonArgs);
1568 // Get any return attributes.
1569 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1571 // If the return value is not being used, the type may not be compatible
1572 // with the existing attributes. Wipe out any problematic attributes.
1574 removeAttributes(AttributeFuncs::
1575 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1576 AttributeSet::ReturnIndex);
1578 // Add the new return attributes.
1579 if (RAttrs.hasAttributes())
1580 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1581 AttributeSet::ReturnIndex, RAttrs));
1583 AI = CS.arg_begin();
1584 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1585 Type *ParamTy = FT->getParamType(i);
1587 if ((*AI)->getType() == ParamTy) {
1588 Args.push_back(*AI);
1590 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1593 // Add any parameter attributes.
1594 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1595 if (PAttrs.hasAttributes())
1596 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1600 // If the function takes more arguments than the call was taking, add them
1602 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1603 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1605 // If we are removing arguments to the function, emit an obnoxious warning.
1606 if (FT->getNumParams() < NumActualArgs) {
1607 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1608 if (FT->isVarArg()) {
1609 // Add all of the arguments in their promoted form to the arg list.
1610 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1611 Type *PTy = getPromotedType((*AI)->getType());
1612 if (PTy != (*AI)->getType()) {
1613 // Must promote to pass through va_arg area!
1614 Instruction::CastOps opcode =
1615 CastInst::getCastOpcode(*AI, false, PTy, false);
1616 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1618 Args.push_back(*AI);
1621 // Add any parameter attributes.
1622 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1623 if (PAttrs.hasAttributes())
1624 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1630 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1631 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1632 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1634 if (NewRetTy->isVoidTy())
1635 Caller->setName(""); // Void type should not have a name.
1637 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1641 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1642 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1643 II->getUnwindDest(), Args);
1645 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1646 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1648 CallInst *CI = cast<CallInst>(Caller);
1649 NC = Builder->CreateCall(Callee, Args);
1651 if (CI->isTailCall())
1652 cast<CallInst>(NC)->setTailCall();
1653 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1654 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1657 // Insert a cast of the return type as necessary.
1659 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1660 if (!NV->getType()->isVoidTy()) {
1661 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1662 NC->setDebugLoc(Caller->getDebugLoc());
1664 // If this is an invoke instruction, we should insert it after the first
1665 // non-phi, instruction in the normal successor block.
1666 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1667 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1668 InsertNewInstBefore(NC, *I);
1670 // Otherwise, it's a call, just insert cast right after the call.
1671 InsertNewInstBefore(NC, *Caller);
1673 Worklist.AddUsersToWorkList(*Caller);
1675 NV = UndefValue::get(Caller->getType());
1679 if (!Caller->use_empty())
1680 ReplaceInstUsesWith(*Caller, NV);
1681 else if (Caller->hasValueHandle()) {
1682 if (OldRetTy == NV->getType())
1683 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1685 // We cannot call ValueIsRAUWd with a different type, and the
1686 // actual tracked value will disappear.
1687 ValueHandleBase::ValueIsDeleted(Caller);
1690 EraseInstFromFunction(*Caller);
1694 // transformCallThroughTrampoline - Turn a call to a function created by
1695 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1696 // underlying function.
1699 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1700 IntrinsicInst *Tramp) {
1701 Value *Callee = CS.getCalledValue();
1702 PointerType *PTy = cast<PointerType>(Callee->getType());
1703 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1704 const AttributeSet &Attrs = CS.getAttributes();
1706 // If the call already has the 'nest' attribute somewhere then give up -
1707 // otherwise 'nest' would occur twice after splicing in the chain.
1708 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1712 "transformCallThroughTrampoline called with incorrect CallSite.");
1714 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1715 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1716 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1718 const AttributeSet &NestAttrs = NestF->getAttributes();
1719 if (!NestAttrs.isEmpty()) {
1720 unsigned NestIdx = 1;
1721 Type *NestTy = nullptr;
1722 AttributeSet NestAttr;
1724 // Look for a parameter marked with the 'nest' attribute.
1725 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1726 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1727 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1728 // Record the parameter type and any other attributes.
1730 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1735 Instruction *Caller = CS.getInstruction();
1736 std::vector<Value*> NewArgs;
1737 NewArgs.reserve(CS.arg_size() + 1);
1739 SmallVector<AttributeSet, 8> NewAttrs;
1740 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1742 // Insert the nest argument into the call argument list, which may
1743 // mean appending it. Likewise for attributes.
1745 // Add any result attributes.
1746 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1747 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1748 Attrs.getRetAttributes()));
1752 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1754 if (Idx == NestIdx) {
1755 // Add the chain argument and attributes.
1756 Value *NestVal = Tramp->getArgOperand(2);
1757 if (NestVal->getType() != NestTy)
1758 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1759 NewArgs.push_back(NestVal);
1760 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1767 // Add the original argument and attributes.
1768 NewArgs.push_back(*I);
1769 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1770 if (Attr.hasAttributes(Idx)) {
1771 AttrBuilder B(Attr, Idx);
1772 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1773 Idx + (Idx >= NestIdx), B));
1780 // Add any function attributes.
1781 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1782 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1783 Attrs.getFnAttributes()));
1785 // The trampoline may have been bitcast to a bogus type (FTy).
1786 // Handle this by synthesizing a new function type, equal to FTy
1787 // with the chain parameter inserted.
1789 std::vector<Type*> NewTypes;
1790 NewTypes.reserve(FTy->getNumParams()+1);
1792 // Insert the chain's type into the list of parameter types, which may
1793 // mean appending it.
1796 FunctionType::param_iterator I = FTy->param_begin(),
1797 E = FTy->param_end();
1801 // Add the chain's type.
1802 NewTypes.push_back(NestTy);
1807 // Add the original type.
1808 NewTypes.push_back(*I);
1814 // Replace the trampoline call with a direct call. Let the generic
1815 // code sort out any function type mismatches.
1816 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1818 Constant *NewCallee =
1819 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1820 NestF : ConstantExpr::getBitCast(NestF,
1821 PointerType::getUnqual(NewFTy));
1822 const AttributeSet &NewPAL =
1823 AttributeSet::get(FTy->getContext(), NewAttrs);
1825 Instruction *NewCaller;
1826 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1827 NewCaller = InvokeInst::Create(NewCallee,
1828 II->getNormalDest(), II->getUnwindDest(),
1830 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1831 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1833 NewCaller = CallInst::Create(NewCallee, NewArgs);
1834 if (cast<CallInst>(Caller)->isTailCall())
1835 cast<CallInst>(NewCaller)->setTailCall();
1836 cast<CallInst>(NewCaller)->
1837 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1838 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1845 // Replace the trampoline call with a direct call. Since there is no 'nest'
1846 // parameter, there is no need to adjust the argument list. Let the generic
1847 // code sort out any function type mismatches.
1848 Constant *NewCallee =
1849 NestF->getType() == PTy ? NestF :
1850 ConstantExpr::getBitCast(NestF, PTy);
1851 CS.setCalledFunction(NewCallee);
1852 return CS.getInstruction();