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 "InstCombine.h"
15 #include "llvm/IntrinsicInst.h"
16 #include "llvm/Support/CallSite.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/Transforms/Utils/BuildLibCalls.h"
22 /// getPromotedType - Return the specified type promoted as it would be to pass
23 /// though a va_arg area.
24 static const Type *getPromotedType(const Type *Ty) {
25 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
26 if (ITy->getBitWidth() < 32)
27 return Type::getInt32Ty(Ty->getContext());
32 /// EnforceKnownAlignment - If the specified pointer points to an object that
33 /// we control, modify the object's alignment to PrefAlign. This isn't
34 /// often possible though. If alignment is important, a more reliable approach
35 /// is to simply align all global variables and allocation instructions to
36 /// their preferred alignment from the beginning.
38 static unsigned EnforceKnownAlignment(Value *V,
39 unsigned Align, unsigned PrefAlign) {
41 User *U = dyn_cast<User>(V);
44 switch (Operator::getOpcode(U)) {
46 case Instruction::BitCast:
47 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
48 case Instruction::GetElementPtr: {
49 // If all indexes are zero, it is just the alignment of the base pointer.
50 bool AllZeroOperands = true;
51 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
52 if (!isa<Constant>(*i) ||
53 !cast<Constant>(*i)->isNullValue()) {
54 AllZeroOperands = false;
58 if (AllZeroOperands) {
59 // Treat this like a bitcast.
60 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
64 case Instruction::Alloca: {
65 AllocaInst *AI = cast<AllocaInst>(V);
66 // If there is a requested alignment and if this is an alloca, round up.
67 if (AI->getAlignment() >= PrefAlign)
68 return AI->getAlignment();
69 AI->setAlignment(PrefAlign);
74 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
75 // If there is a large requested alignment and we can, bump up the alignment
77 if (GV->isDeclaration()) return Align;
79 if (GV->getAlignment() >= PrefAlign)
80 return GV->getAlignment();
81 // We can only increase the alignment of the global if it has no alignment
82 // specified or if it is not assigned a section. If it is assigned a
83 // section, the global could be densely packed with other objects in the
84 // section, increasing the alignment could cause padding issues.
85 if (!GV->hasSection() || GV->getAlignment() == 0)
86 GV->setAlignment(PrefAlign);
87 return GV->getAlignment();
93 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
94 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
95 /// and it is more than the alignment of the ultimate object, see if we can
96 /// increase the alignment of the ultimate object, making this check succeed.
97 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
99 assert(V->getType()->isPointerTy() &&
100 "GetOrEnforceKnownAlignment expects a pointer!");
101 unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 64;
102 APInt Mask = APInt::getAllOnesValue(BitWidth);
103 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
104 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
105 unsigned TrailZ = KnownZero.countTrailingOnes();
107 // LLVM doesn't support alignments larger than this currently.
108 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
110 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
112 if (PrefAlign > Align)
113 Align = EnforceKnownAlignment(V, Align, PrefAlign);
115 // We don't need to make any adjustment.
119 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
120 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(0));
121 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(1));
122 unsigned MinAlign = std::min(DstAlign, SrcAlign);
123 unsigned CopyAlign = MI->getAlignment();
125 if (CopyAlign < MinAlign) {
126 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
131 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
133 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
134 if (MemOpLength == 0) return 0;
136 // Source and destination pointer types are always "i8*" for intrinsic. See
137 // if the size is something we can handle with a single primitive load/store.
138 // A single load+store correctly handles overlapping memory in the memmove
140 unsigned Size = MemOpLength->getZExtValue();
141 if (Size == 0) return MI; // Delete this mem transfer.
143 if (Size > 8 || (Size&(Size-1)))
144 return 0; // If not 1/2/4/8 bytes, exit.
146 // Use an integer load+store unless we can find something better.
148 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
150 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
152 const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
153 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
154 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
156 // Memcpy forces the use of i8* for the source and destination. That means
157 // that if you're using memcpy to move one double around, you'll get a cast
158 // from double* to i8*. We'd much rather use a double load+store rather than
159 // an i64 load+store, here because this improves the odds that the source or
160 // dest address will be promotable. See if we can find a better type than the
162 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
163 if (StrippedDest != MI->getArgOperand(0)) {
164 const Type *SrcETy = cast<PointerType>(StrippedDest->getType())
166 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
167 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
168 // down through these levels if so.
169 while (!SrcETy->isSingleValueType()) {
170 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
171 if (STy->getNumElements() == 1)
172 SrcETy = STy->getElementType(0);
175 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
176 if (ATy->getNumElements() == 1)
177 SrcETy = ATy->getElementType();
184 if (SrcETy->isSingleValueType()) {
185 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
186 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
192 // If the memcpy/memmove provides better alignment info than we can
194 SrcAlign = std::max(SrcAlign, CopyAlign);
195 DstAlign = std::max(DstAlign, CopyAlign);
197 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
198 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
199 Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign);
200 InsertNewInstBefore(L, *MI);
201 InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign),
204 // Set the size of the copy to 0, it will be deleted on the next iteration.
205 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
209 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
210 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
211 if (MI->getAlignment() < Alignment) {
212 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
217 // Extract the length and alignment and fill if they are constant.
218 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
219 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
220 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
222 uint64_t Len = LenC->getZExtValue();
223 Alignment = MI->getAlignment();
225 // If the length is zero, this is a no-op
226 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
228 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
229 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
230 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
232 Value *Dest = MI->getDest();
233 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
235 // Alignment 0 is identity for alignment 1 for memset, but not store.
236 if (Alignment == 0) Alignment = 1;
238 // Extract the fill value and store.
239 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
240 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
241 Dest, false, Alignment), *MI);
243 // Set the size of the copy to 0, it will be deleted on the next iteration.
244 MI->setLength(Constant::getNullValue(LenC->getType()));
251 /// visitCallInst - CallInst simplification. This mostly only handles folding
252 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
253 /// the heavy lifting.
255 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
257 return visitFree(CI);
259 return visitMalloc(CI);
261 // If the caller function is nounwind, mark the call as nounwind, even if the
263 if (CI.getParent()->getParent()->doesNotThrow() &&
264 !CI.doesNotThrow()) {
265 CI.setDoesNotThrow();
269 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
270 if (!II) return visitCallSite(&CI);
272 // Intrinsics cannot occur in an invoke, so handle them here instead of in
274 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
275 bool Changed = false;
277 // memmove/cpy/set of zero bytes is a noop.
278 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
279 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
281 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
282 if (CI->getZExtValue() == 1) {
283 // Replace the instruction with just byte operations. We would
284 // transform other cases to loads/stores, but we don't know if
285 // alignment is sufficient.
289 // If we have a memmove and the source operation is a constant global,
290 // then the source and dest pointers can't alias, so we can change this
291 // into a call to memcpy.
292 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
293 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
294 if (GVSrc->isConstant()) {
295 Module *M = CI.getParent()->getParent()->getParent();
296 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
297 const Type *Tys[3] = { CI.getArgOperand(0)->getType(),
298 CI.getArgOperand(1)->getType(),
299 CI.getArgOperand(2)->getType() };
300 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
305 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
306 // memmove(x,x,size) -> noop.
307 if (MTI->getSource() == MTI->getDest())
308 return EraseInstFromFunction(CI);
311 // If we can determine a pointer alignment that is bigger than currently
312 // set, update the alignment.
313 if (isa<MemTransferInst>(MI)) {
314 if (Instruction *I = SimplifyMemTransfer(MI))
316 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
317 if (Instruction *I = SimplifyMemSet(MSI))
321 if (Changed) return II;
324 switch (II->getIntrinsicID()) {
326 case Intrinsic::objectsize: {
327 // We need target data for just about everything so depend on it.
330 const Type *ReturnTy = CI.getType();
331 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
333 // Get to the real allocated thing and offset as fast as possible.
334 Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
336 // If we've stripped down to a single global variable that we
337 // can know the size of then just return that.
338 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
339 if (GV->hasDefinitiveInitializer()) {
340 Constant *C = GV->getInitializer();
341 uint64_t GlobalSize = TD->getTypeAllocSize(C->getType());
342 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize));
344 // Can't determine size of the GV.
345 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
346 return ReplaceInstUsesWith(CI, RetVal);
348 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
350 if (AI->getAllocatedType()->isSized()) {
351 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
352 if (AI->isArrayAllocation()) {
353 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
355 AllocaSize *= C->getZExtValue();
357 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize));
359 } else if (CallInst *MI = extractMallocCall(Op1)) {
360 const Type* MallocType = getMallocAllocatedType(MI);
362 if (MallocType && MallocType->isSized()) {
363 if (Value *NElems = getMallocArraySize(MI, TD, true)) {
364 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
365 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy,
366 (NElements->getZExtValue() * TD->getTypeAllocSize(MallocType))));
369 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) {
370 // Only handle constant GEPs here.
371 if (CE->getOpcode() != Instruction::GetElementPtr) break;
372 GEPOperator *GEP = cast<GEPOperator>(CE);
374 // Make sure we're not a constant offset from an external
376 Value *Operand = GEP->getPointerOperand();
377 Operand = Operand->stripPointerCasts();
378 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand))
379 if (!GV->hasDefinitiveInitializer()) break;
381 // Get what we're pointing to and its size.
382 const PointerType *BaseType =
383 cast<PointerType>(Operand->getType());
384 uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType());
386 // Get the current byte offset into the thing. Use the original
387 // operand in case we're looking through a bitcast.
388 SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end());
389 const PointerType *OffsetType =
390 cast<PointerType>(GEP->getPointerOperand()->getType());
391 uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size());
394 // Out of bound reference? Negative index normalized to large
395 // index? Just return "I don't know".
396 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
397 return ReplaceInstUsesWith(CI, RetVal);
400 Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset);
401 return ReplaceInstUsesWith(CI, RetVal);
404 // Do not return "I don't know" here. Later optimization passes could
405 // make it possible to evaluate objectsize to a constant.
408 case Intrinsic::bswap:
409 // bswap(bswap(x)) -> x
410 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
411 if (Operand->getIntrinsicID() == Intrinsic::bswap)
412 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
414 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
415 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
416 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
417 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
418 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
419 TI->getType()->getPrimitiveSizeInBits();
420 Value *CV = ConstantInt::get(Operand->getType(), C);
421 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
422 return new TruncInst(V, TI->getType());
427 case Intrinsic::powi:
428 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
431 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
434 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
435 // powi(x, -1) -> 1/x
436 if (Power->isAllOnesValue())
437 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
438 II->getArgOperand(0));
441 case Intrinsic::cttz: {
442 // If all bits below the first known one are known zero,
443 // this value is constant.
444 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
445 uint32_t BitWidth = IT->getBitWidth();
446 APInt KnownZero(BitWidth, 0);
447 APInt KnownOne(BitWidth, 0);
448 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
449 KnownZero, KnownOne);
450 unsigned TrailingZeros = KnownOne.countTrailingZeros();
451 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
452 if ((Mask & KnownZero) == Mask)
453 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
454 APInt(BitWidth, TrailingZeros)));
458 case Intrinsic::ctlz: {
459 // If all bits above the first known one are known zero,
460 // this value is constant.
461 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
462 uint32_t BitWidth = IT->getBitWidth();
463 APInt KnownZero(BitWidth, 0);
464 APInt KnownOne(BitWidth, 0);
465 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
466 KnownZero, KnownOne);
467 unsigned LeadingZeros = KnownOne.countLeadingZeros();
468 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
469 if ((Mask & KnownZero) == Mask)
470 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
471 APInt(BitWidth, LeadingZeros)));
475 case Intrinsic::uadd_with_overflow: {
476 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
477 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
478 uint32_t BitWidth = IT->getBitWidth();
479 APInt Mask = APInt::getSignBit(BitWidth);
480 APInt LHSKnownZero(BitWidth, 0);
481 APInt LHSKnownOne(BitWidth, 0);
482 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
483 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
484 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
486 if (LHSKnownNegative || LHSKnownPositive) {
487 APInt RHSKnownZero(BitWidth, 0);
488 APInt RHSKnownOne(BitWidth, 0);
489 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
490 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
491 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
492 if (LHSKnownNegative && RHSKnownNegative) {
493 // The sign bit is set in both cases: this MUST overflow.
494 // Create a simple add instruction, and insert it into the struct.
495 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
498 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
500 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
501 return InsertValueInst::Create(Struct, Add, 0);
504 if (LHSKnownPositive && RHSKnownPositive) {
505 // The sign bit is clear in both cases: this CANNOT overflow.
506 // Create a simple add instruction, and insert it into the struct.
507 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
510 UndefValue::get(LHS->getType()),
511 ConstantInt::getFalse(II->getContext())
513 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
514 return InsertValueInst::Create(Struct, Add, 0);
518 // FALL THROUGH uadd into sadd
519 case Intrinsic::sadd_with_overflow:
520 // Canonicalize constants into the RHS.
521 if (isa<Constant>(II->getArgOperand(0)) &&
522 !isa<Constant>(II->getArgOperand(1))) {
523 Value *LHS = II->getArgOperand(0);
524 II->setArgOperand(0, II->getArgOperand(1));
525 II->setArgOperand(1, LHS);
529 // X + undef -> undef
530 if (isa<UndefValue>(II->getArgOperand(1)))
531 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
533 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
534 // X + 0 -> {X, false}
537 UndefValue::get(II->getCalledValue()->getType()),
538 ConstantInt::getFalse(II->getContext())
540 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
541 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
545 case Intrinsic::usub_with_overflow:
546 case Intrinsic::ssub_with_overflow:
547 // undef - X -> undef
548 // X - undef -> undef
549 if (isa<UndefValue>(II->getArgOperand(0)) ||
550 isa<UndefValue>(II->getArgOperand(1)))
551 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
553 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
554 // X - 0 -> {X, false}
557 UndefValue::get(II->getArgOperand(0)->getType()),
558 ConstantInt::getFalse(II->getContext())
560 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
561 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
565 case Intrinsic::umul_with_overflow:
566 case Intrinsic::smul_with_overflow:
567 // Canonicalize constants into the RHS.
568 if (isa<Constant>(II->getArgOperand(0)) &&
569 !isa<Constant>(II->getArgOperand(1))) {
570 Value *LHS = II->getArgOperand(0);
571 II->setArgOperand(0, II->getArgOperand(1));
572 II->setArgOperand(1, LHS);
576 // X * undef -> undef
577 if (isa<UndefValue>(II->getArgOperand(1)))
578 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
580 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
583 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
585 // X * 1 -> {X, false}
586 if (RHSI->equalsInt(1)) {
588 UndefValue::get(II->getArgOperand(0)->getType()),
589 ConstantInt::getFalse(II->getContext())
591 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
592 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
596 case Intrinsic::ppc_altivec_lvx:
597 case Intrinsic::ppc_altivec_lvxl:
598 case Intrinsic::x86_sse_loadu_ps:
599 case Intrinsic::x86_sse2_loadu_pd:
600 case Intrinsic::x86_sse2_loadu_dq:
601 // Turn PPC lvx -> load if the pointer is known aligned.
602 // Turn X86 loadups -> load if the pointer is known aligned.
603 if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
604 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
605 PointerType::getUnqual(II->getType()));
606 return new LoadInst(Ptr);
609 case Intrinsic::ppc_altivec_stvx:
610 case Intrinsic::ppc_altivec_stvxl:
611 // Turn stvx -> store if the pointer is known aligned.
612 if (GetOrEnforceKnownAlignment(II->getArgOperand(1), 16) >= 16) {
613 const Type *OpPtrTy =
614 PointerType::getUnqual(II->getArgOperand(0)->getType());
615 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
616 return new StoreInst(II->getArgOperand(0), Ptr);
619 case Intrinsic::x86_sse_storeu_ps:
620 case Intrinsic::x86_sse2_storeu_pd:
621 case Intrinsic::x86_sse2_storeu_dq:
622 // Turn X86 storeu -> store if the pointer is known aligned.
623 if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
624 const Type *OpPtrTy =
625 PointerType::getUnqual(II->getArgOperand(1)->getType());
626 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
627 return new StoreInst(II->getArgOperand(1), Ptr);
631 case Intrinsic::x86_sse_cvttss2si: {
632 // These intrinsics only demands the 0th element of its input vector. If
633 // we can simplify the input based on that, do so now.
635 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
636 APInt DemandedElts(VWidth, 1);
637 APInt UndefElts(VWidth, 0);
638 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
639 DemandedElts, UndefElts)) {
640 II->setArgOperand(0, V);
646 case Intrinsic::ppc_altivec_vperm:
647 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
648 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
649 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
651 // Check that all of the elements are integer constants or undefs.
652 bool AllEltsOk = true;
653 for (unsigned i = 0; i != 16; ++i) {
654 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
655 !isa<UndefValue>(Mask->getOperand(i))) {
662 // Cast the input vectors to byte vectors.
663 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
665 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
667 Value *Result = UndefValue::get(Op0->getType());
669 // Only extract each element once.
670 Value *ExtractedElts[32];
671 memset(ExtractedElts, 0, sizeof(ExtractedElts));
673 for (unsigned i = 0; i != 16; ++i) {
674 if (isa<UndefValue>(Mask->getOperand(i)))
676 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
677 Idx &= 31; // Match the hardware behavior.
679 if (ExtractedElts[Idx] == 0) {
681 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
682 ConstantInt::get(Type::getInt32Ty(II->getContext()),
683 Idx&15, false), "tmp");
686 // Insert this value into the result vector.
687 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
688 ConstantInt::get(Type::getInt32Ty(II->getContext()),
691 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
696 case Intrinsic::stackrestore: {
697 // If the save is right next to the restore, remove the restore. This can
698 // happen when variable allocas are DCE'd.
699 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
700 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
701 BasicBlock::iterator BI = SS;
703 return EraseInstFromFunction(CI);
707 // Scan down this block to see if there is another stack restore in the
708 // same block without an intervening call/alloca.
709 BasicBlock::iterator BI = II;
710 TerminatorInst *TI = II->getParent()->getTerminator();
711 bool CannotRemove = false;
712 for (++BI; &*BI != TI; ++BI) {
713 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
717 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
718 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
719 // If there is a stackrestore below this one, remove this one.
720 if (II->getIntrinsicID() == Intrinsic::stackrestore)
721 return EraseInstFromFunction(CI);
722 // Otherwise, ignore the intrinsic.
724 // If we found a non-intrinsic call, we can't remove the stack
732 // If the stack restore is in a return/unwind block and if there are no
733 // allocas or calls between the restore and the return, nuke the restore.
734 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
735 return EraseInstFromFunction(CI);
740 return visitCallSite(II);
743 // InvokeInst simplification
745 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
746 return visitCallSite(&II);
749 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
750 /// passed through the varargs area, we can eliminate the use of the cast.
751 static bool isSafeToEliminateVarargsCast(const CallSite CS,
752 const CastInst * const CI,
753 const TargetData * const TD,
755 if (!CI->isLosslessCast())
758 // The size of ByVal arguments is derived from the type, so we
759 // can't change to a type with a different size. If the size were
760 // passed explicitly we could avoid this check.
761 if (!CS.paramHasAttr(ix, Attribute::ByVal))
765 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
766 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
767 if (!SrcTy->isSized() || !DstTy->isSized())
769 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
775 class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
778 void replaceCall(Value *With) {
779 NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
781 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
782 if (ConstantInt *SizeCI =
783 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
784 if (SizeCI->isAllOnesValue())
787 return SizeCI->getZExtValue() >=
788 GetStringLength(CI->getArgOperand(SizeArgOp));
789 if (ConstantInt *Arg = dyn_cast<ConstantInt>(
790 CI->getArgOperand(SizeArgOp)))
791 return SizeCI->getZExtValue() >= Arg->getZExtValue();
796 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
797 Instruction *NewInstruction;
799 } // end anonymous namespace
801 // Try to fold some different type of calls here.
802 // Currently we're only working with the checking functions, memcpy_chk,
803 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
804 // strcat_chk and strncat_chk.
805 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
806 if (CI->getCalledFunction() == 0) return 0;
808 InstCombineFortifiedLibCalls Simplifier(this);
809 Simplifier.fold(CI, TD);
810 return Simplifier.NewInstruction;
813 // visitCallSite - Improvements for call and invoke instructions.
815 Instruction *InstCombiner::visitCallSite(CallSite CS) {
816 bool Changed = false;
818 // If the callee is a constexpr cast of a function, attempt to move the cast
819 // to the arguments of the call/invoke.
820 if (transformConstExprCastCall(CS)) return 0;
822 Value *Callee = CS.getCalledValue();
824 if (Function *CalleeF = dyn_cast<Function>(Callee))
825 // If the call and callee calling conventions don't match, this call must
826 // be unreachable, as the call is undefined.
827 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
828 // Only do this for calls to a function with a body. A prototype may
829 // not actually end up matching the implementation's calling conv for a
830 // variety of reasons (e.g. it may be written in assembly).
831 !CalleeF->isDeclaration()) {
832 Instruction *OldCall = CS.getInstruction();
833 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
834 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
836 // If OldCall dues not return void then replaceAllUsesWith undef.
837 // This allows ValueHandlers and custom metadata to adjust itself.
838 if (!OldCall->getType()->isVoidTy())
839 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
840 if (isa<CallInst>(OldCall))
841 return EraseInstFromFunction(*OldCall);
843 // We cannot remove an invoke, because it would change the CFG, just
844 // change the callee to a null pointer.
845 cast<InvokeInst>(OldCall)->setCalledFunction(
846 Constant::getNullValue(CalleeF->getType()));
850 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
851 // This instruction is not reachable, just remove it. We insert a store to
852 // undef so that we know that this code is not reachable, despite the fact
853 // that we can't modify the CFG here.
854 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
855 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
856 CS.getInstruction());
858 // If CS does not return void then replaceAllUsesWith undef.
859 // This allows ValueHandlers and custom metadata to adjust itself.
860 if (!CS.getInstruction()->getType()->isVoidTy())
861 CS.getInstruction()->
862 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
864 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
865 // Don't break the CFG, insert a dummy cond branch.
866 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
867 ConstantInt::getTrue(Callee->getContext()), II);
869 return EraseInstFromFunction(*CS.getInstruction());
872 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
873 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
874 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
875 return transformCallThroughTrampoline(CS);
877 const PointerType *PTy = cast<PointerType>(Callee->getType());
878 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
879 if (FTy->isVarArg()) {
880 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
881 // See if we can optimize any arguments passed through the varargs area of
883 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
884 E = CS.arg_end(); I != E; ++I, ++ix) {
885 CastInst *CI = dyn_cast<CastInst>(*I);
886 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
887 *I = CI->getOperand(0);
893 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
894 // Inline asm calls cannot throw - mark them 'nounwind'.
895 CS.setDoesNotThrow();
899 // Try to optimize the call if possible, we require TargetData for most of
900 // this. None of these calls are seen as possibly dead so go ahead and
901 // delete the instruction now.
902 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
903 Instruction *I = tryOptimizeCall(CI, TD);
904 // If we changed something return the result, etc. Otherwise let
905 // the fallthrough check.
906 if (I) return EraseInstFromFunction(*I);
909 return Changed ? CS.getInstruction() : 0;
912 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
913 // attempt to move the cast to the arguments of the call/invoke.
915 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
916 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
917 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
918 if (CE->getOpcode() != Instruction::BitCast ||
919 !isa<Function>(CE->getOperand(0)))
921 Function *Callee = cast<Function>(CE->getOperand(0));
922 Instruction *Caller = CS.getInstruction();
923 const AttrListPtr &CallerPAL = CS.getAttributes();
925 // Okay, this is a cast from a function to a different type. Unless doing so
926 // would cause a type conversion of one of our arguments, change this call to
927 // be a direct call with arguments casted to the appropriate types.
929 const FunctionType *FT = Callee->getFunctionType();
930 const Type *OldRetTy = Caller->getType();
931 const Type *NewRetTy = FT->getReturnType();
933 if (NewRetTy->isStructTy())
934 return false; // TODO: Handle multiple return values.
936 // Check to see if we are changing the return type...
937 if (OldRetTy != NewRetTy) {
938 if (Callee->isDeclaration() &&
939 // Conversion is ok if changing from one pointer type to another or from
940 // a pointer to an integer of the same size.
941 !((OldRetTy->isPointerTy() || !TD ||
942 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
943 (NewRetTy->isPointerTy() || !TD ||
944 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
945 return false; // Cannot transform this return value.
947 if (!Caller->use_empty() &&
948 // void -> non-void is handled specially
949 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
950 return false; // Cannot transform this return value.
952 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
953 Attributes RAttrs = CallerPAL.getRetAttributes();
954 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
955 return false; // Attribute not compatible with transformed value.
958 // If the callsite is an invoke instruction, and the return value is used by
959 // a PHI node in a successor, we cannot change the return type of the call
960 // because there is no place to put the cast instruction (without breaking
961 // the critical edge). Bail out in this case.
962 if (!Caller->use_empty())
963 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
964 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
966 if (PHINode *PN = dyn_cast<PHINode>(*UI))
967 if (PN->getParent() == II->getNormalDest() ||
968 PN->getParent() == II->getUnwindDest())
972 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
973 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
975 CallSite::arg_iterator AI = CS.arg_begin();
976 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
977 const Type *ParamTy = FT->getParamType(i);
978 const Type *ActTy = (*AI)->getType();
980 if (!CastInst::isCastable(ActTy, ParamTy))
981 return false; // Cannot transform this parameter value.
983 if (CallerPAL.getParamAttributes(i + 1)
984 & Attribute::typeIncompatible(ParamTy))
985 return false; // Attribute not compatible with transformed value.
987 // Converting from one pointer type to another or between a pointer and an
988 // integer of the same size is safe even if we do not have a body.
989 bool isConvertible = ActTy == ParamTy ||
990 (TD && ((ParamTy->isPointerTy() ||
991 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
992 (ActTy->isPointerTy() ||
993 ActTy == TD->getIntPtrType(Caller->getContext()))));
994 if (Callee->isDeclaration() && !isConvertible) return false;
997 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
998 Callee->isDeclaration())
999 return false; // Do not delete arguments unless we have a function body.
1001 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1002 !CallerPAL.isEmpty())
1003 // In this case we have more arguments than the new function type, but we
1004 // won't be dropping them. Check that these extra arguments have attributes
1005 // that are compatible with being a vararg call argument.
1006 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1007 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1009 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1010 if (PAttrs & Attribute::VarArgsIncompatible)
1014 // Okay, we decided that this is a safe thing to do: go ahead and start
1015 // inserting cast instructions as necessary...
1016 std::vector<Value*> Args;
1017 Args.reserve(NumActualArgs);
1018 SmallVector<AttributeWithIndex, 8> attrVec;
1019 attrVec.reserve(NumCommonArgs);
1021 // Get any return attributes.
1022 Attributes RAttrs = CallerPAL.getRetAttributes();
1024 // If the return value is not being used, the type may not be compatible
1025 // with the existing attributes. Wipe out any problematic attributes.
1026 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1028 // Add the new return attributes.
1030 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1032 AI = CS.arg_begin();
1033 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1034 const Type *ParamTy = FT->getParamType(i);
1035 if ((*AI)->getType() == ParamTy) {
1036 Args.push_back(*AI);
1038 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1039 false, ParamTy, false);
1040 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1043 // Add any parameter attributes.
1044 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1045 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1048 // If the function takes more arguments than the call was taking, add them
1050 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1051 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1053 // If we are removing arguments to the function, emit an obnoxious warning.
1054 if (FT->getNumParams() < NumActualArgs) {
1055 if (!FT->isVarArg()) {
1056 errs() << "WARNING: While resolving call to function '"
1057 << Callee->getName() << "' arguments were dropped!\n";
1059 // Add all of the arguments in their promoted form to the arg list.
1060 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1061 const Type *PTy = getPromotedType((*AI)->getType());
1062 if (PTy != (*AI)->getType()) {
1063 // Must promote to pass through va_arg area!
1064 Instruction::CastOps opcode =
1065 CastInst::getCastOpcode(*AI, false, PTy, false);
1066 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1068 Args.push_back(*AI);
1071 // Add any parameter attributes.
1072 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1073 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1078 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1079 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1081 if (NewRetTy->isVoidTy())
1082 Caller->setName(""); // Void type should not have a name.
1084 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1088 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1089 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
1090 Args.begin(), Args.end(),
1091 Caller->getName(), Caller);
1092 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1093 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1095 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
1096 Caller->getName(), Caller);
1097 CallInst *CI = cast<CallInst>(Caller);
1098 if (CI->isTailCall())
1099 cast<CallInst>(NC)->setTailCall();
1100 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1101 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1104 // Insert a cast of the return type as necessary.
1106 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1107 if (!NV->getType()->isVoidTy()) {
1108 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
1110 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1112 // If this is an invoke instruction, we should insert it after the first
1113 // non-phi, instruction in the normal successor block.
1114 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1115 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1116 InsertNewInstBefore(NC, *I);
1118 // Otherwise, it's a call, just insert cast right after the call instr
1119 InsertNewInstBefore(NC, *Caller);
1121 Worklist.AddUsersToWorkList(*Caller);
1123 NV = UndefValue::get(Caller->getType());
1128 if (!Caller->use_empty())
1129 Caller->replaceAllUsesWith(NV);
1131 EraseInstFromFunction(*Caller);
1135 // transformCallThroughTrampoline - Turn a call to a function created by the
1136 // init_trampoline intrinsic into a direct call to the underlying function.
1138 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1139 Value *Callee = CS.getCalledValue();
1140 const PointerType *PTy = cast<PointerType>(Callee->getType());
1141 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1142 const AttrListPtr &Attrs = CS.getAttributes();
1144 // If the call already has the 'nest' attribute somewhere then give up -
1145 // otherwise 'nest' would occur twice after splicing in the chain.
1146 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1149 IntrinsicInst *Tramp =
1150 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1152 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1153 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1154 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1156 const AttrListPtr &NestAttrs = NestF->getAttributes();
1157 if (!NestAttrs.isEmpty()) {
1158 unsigned NestIdx = 1;
1159 const Type *NestTy = 0;
1160 Attributes NestAttr = Attribute::None;
1162 // Look for a parameter marked with the 'nest' attribute.
1163 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1164 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1165 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1166 // Record the parameter type and any other attributes.
1168 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1173 Instruction *Caller = CS.getInstruction();
1174 std::vector<Value*> NewArgs;
1175 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1177 SmallVector<AttributeWithIndex, 8> NewAttrs;
1178 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1180 // Insert the nest argument into the call argument list, which may
1181 // mean appending it. Likewise for attributes.
1183 // Add any result attributes.
1184 if (Attributes Attr = Attrs.getRetAttributes())
1185 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1189 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1191 if (Idx == NestIdx) {
1192 // Add the chain argument and attributes.
1193 Value *NestVal = Tramp->getArgOperand(2);
1194 if (NestVal->getType() != NestTy)
1195 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
1196 NewArgs.push_back(NestVal);
1197 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1203 // Add the original argument and attributes.
1204 NewArgs.push_back(*I);
1205 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1207 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1213 // Add any function attributes.
1214 if (Attributes Attr = Attrs.getFnAttributes())
1215 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1217 // The trampoline may have been bitcast to a bogus type (FTy).
1218 // Handle this by synthesizing a new function type, equal to FTy
1219 // with the chain parameter inserted.
1221 std::vector<const Type*> NewTypes;
1222 NewTypes.reserve(FTy->getNumParams()+1);
1224 // Insert the chain's type into the list of parameter types, which may
1225 // mean appending it.
1228 FunctionType::param_iterator I = FTy->param_begin(),
1229 E = FTy->param_end();
1233 // Add the chain's type.
1234 NewTypes.push_back(NestTy);
1239 // Add the original type.
1240 NewTypes.push_back(*I);
1246 // Replace the trampoline call with a direct call. Let the generic
1247 // code sort out any function type mismatches.
1248 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1250 Constant *NewCallee =
1251 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1252 NestF : ConstantExpr::getBitCast(NestF,
1253 PointerType::getUnqual(NewFTy));
1254 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1257 Instruction *NewCaller;
1258 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1259 NewCaller = InvokeInst::Create(NewCallee,
1260 II->getNormalDest(), II->getUnwindDest(),
1261 NewArgs.begin(), NewArgs.end(),
1262 Caller->getName(), Caller);
1263 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1264 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1266 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
1267 Caller->getName(), Caller);
1268 if (cast<CallInst>(Caller)->isTailCall())
1269 cast<CallInst>(NewCaller)->setTailCall();
1270 cast<CallInst>(NewCaller)->
1271 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1272 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1274 if (!Caller->getType()->isVoidTy())
1275 Caller->replaceAllUsesWith(NewCaller);
1276 Caller->eraseFromParent();
1277 Worklist.Remove(Caller);
1282 // Replace the trampoline call with a direct call. Since there is no 'nest'
1283 // parameter, there is no need to adjust the argument list. Let the generic
1284 // code sort out any function type mismatches.
1285 Constant *NewCallee =
1286 NestF->getType() == PTy ? NestF :
1287 ConstantExpr::getBitCast(NestF, PTy);
1288 CS.setCalledFunction(NewCallee);
1289 return CS.getInstruction();