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"
20 #include "llvm/Transforms/Utils/Local.h"
23 /// getPromotedType - Return the specified type promoted as it would be to pass
24 /// though a va_arg area.
25 static const Type *getPromotedType(const Type *Ty) {
26 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
27 if (ITy->getBitWidth() < 32)
28 return Type::getInt32Ty(Ty->getContext());
34 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
35 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD);
36 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
37 unsigned MinAlign = std::min(DstAlign, SrcAlign);
38 unsigned CopyAlign = MI->getAlignment();
40 if (CopyAlign < MinAlign) {
41 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
46 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
48 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
49 if (MemOpLength == 0) return 0;
51 // Source and destination pointer types are always "i8*" for intrinsic. See
52 // if the size is something we can handle with a single primitive load/store.
53 // A single load+store correctly handles overlapping memory in the memmove
55 unsigned Size = MemOpLength->getZExtValue();
56 if (Size == 0) return MI; // Delete this mem transfer.
58 if (Size > 8 || (Size&(Size-1)))
59 return 0; // If not 1/2/4/8 bytes, exit.
61 // Use an integer load+store unless we can find something better.
63 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
65 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
67 const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
68 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
69 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
71 // Memcpy forces the use of i8* for the source and destination. That means
72 // that if you're using memcpy to move one double around, you'll get a cast
73 // from double* to i8*. We'd much rather use a double load+store rather than
74 // an i64 load+store, here because this improves the odds that the source or
75 // dest address will be promotable. See if we can find a better type than the
77 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
78 if (StrippedDest != MI->getArgOperand(0)) {
79 const Type *SrcETy = cast<PointerType>(StrippedDest->getType())
81 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
82 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
83 // down through these levels if so.
84 while (!SrcETy->isSingleValueType()) {
85 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
86 if (STy->getNumElements() == 1)
87 SrcETy = STy->getElementType(0);
90 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
91 if (ATy->getNumElements() == 1)
92 SrcETy = ATy->getElementType();
99 if (SrcETy->isSingleValueType()) {
100 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
101 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
107 // If the memcpy/memmove provides better alignment info than we can
109 SrcAlign = std::max(SrcAlign, CopyAlign);
110 DstAlign = std::max(DstAlign, CopyAlign);
112 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
113 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
114 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
115 L->setAlignment(SrcAlign);
116 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
117 S->setAlignment(DstAlign);
119 // Set the size of the copy to 0, it will be deleted on the next iteration.
120 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
124 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
125 unsigned Alignment = getKnownAlignment(MI->getDest(), TD);
126 if (MI->getAlignment() < Alignment) {
127 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
132 // Extract the length and alignment and fill if they are constant.
133 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
134 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
135 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
137 uint64_t Len = LenC->getZExtValue();
138 Alignment = MI->getAlignment();
140 // If the length is zero, this is a no-op
141 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
143 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
144 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
145 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
147 Value *Dest = MI->getDest();
148 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
149 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
150 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
152 // Alignment 0 is identity for alignment 1 for memset, but not store.
153 if (Alignment == 0) Alignment = 1;
155 // Extract the fill value and store.
156 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
157 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
159 S->setAlignment(Alignment);
161 // Set the size of the copy to 0, it will be deleted on the next iteration.
162 MI->setLength(Constant::getNullValue(LenC->getType()));
169 /// visitCallInst - CallInst simplification. This mostly only handles folding
170 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
171 /// the heavy lifting.
173 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
175 return visitFree(CI);
177 return visitMalloc(CI);
179 // If the caller function is nounwind, mark the call as nounwind, even if the
181 if (CI.getParent()->getParent()->doesNotThrow() &&
182 !CI.doesNotThrow()) {
183 CI.setDoesNotThrow();
187 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
188 if (!II) return visitCallSite(&CI);
190 // Intrinsics cannot occur in an invoke, so handle them here instead of in
192 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
193 bool Changed = false;
195 // memmove/cpy/set of zero bytes is a noop.
196 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
197 if (NumBytes->isNullValue())
198 return EraseInstFromFunction(CI);
200 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
201 if (CI->getZExtValue() == 1) {
202 // Replace the instruction with just byte operations. We would
203 // transform other cases to loads/stores, but we don't know if
204 // alignment is sufficient.
208 // No other transformations apply to volatile transfers.
209 if (MI->isVolatile())
212 // If we have a memmove and the source operation is a constant global,
213 // then the source and dest pointers can't alias, so we can change this
214 // into a call to memcpy.
215 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
216 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
217 if (GVSrc->isConstant()) {
218 Module *M = CI.getParent()->getParent()->getParent();
219 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
220 const Type *Tys[3] = { CI.getArgOperand(0)->getType(),
221 CI.getArgOperand(1)->getType(),
222 CI.getArgOperand(2)->getType() };
223 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
228 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
229 // memmove(x,x,size) -> noop.
230 if (MTI->getSource() == MTI->getDest())
231 return EraseInstFromFunction(CI);
234 // If we can determine a pointer alignment that is bigger than currently
235 // set, update the alignment.
236 if (isa<MemTransferInst>(MI)) {
237 if (Instruction *I = SimplifyMemTransfer(MI))
239 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
240 if (Instruction *I = SimplifyMemSet(MSI))
244 if (Changed) return II;
247 switch (II->getIntrinsicID()) {
249 case Intrinsic::objectsize: {
250 // We need target data for just about everything so depend on it.
253 const Type *ReturnTy = CI.getType();
254 uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL;
256 // Get to the real allocated thing and offset as fast as possible.
257 Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
260 uint64_t Size = -1ULL;
262 // Try to look through constant GEPs.
263 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) {
264 if (!GEP->hasAllConstantIndices()) break;
266 // Get the current byte offset into the thing. Use the original
267 // operand in case we're looking through a bitcast.
268 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
269 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(),
270 Ops.data(), Ops.size());
272 Op1 = GEP->getPointerOperand()->stripPointerCasts();
274 // Make sure we're not a constant offset from an external
276 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1))
277 if (!GV->hasDefinitiveInitializer()) break;
280 // If we've stripped down to a single global variable that we
281 // can know the size of then just return that.
282 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
283 if (GV->hasDefinitiveInitializer()) {
284 Constant *C = GV->getInitializer();
285 Size = TD->getTypeAllocSize(C->getType());
287 // Can't determine size of the GV.
288 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow);
289 return ReplaceInstUsesWith(CI, RetVal);
291 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
293 if (AI->getAllocatedType()->isSized()) {
294 Size = TD->getTypeAllocSize(AI->getAllocatedType());
295 if (AI->isArrayAllocation()) {
296 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
298 Size *= C->getZExtValue();
301 } else if (CallInst *MI = extractMallocCall(Op1)) {
302 // Get allocation size.
303 const Type* MallocType = getMallocAllocatedType(MI);
304 if (MallocType && MallocType->isSized())
305 if (Value *NElems = getMallocArraySize(MI, TD, true))
306 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
307 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
310 // Do not return "I don't know" here. Later optimization passes could
311 // make it possible to evaluate objectsize to a constant.
316 // Out of bound reference? Negative index normalized to large
317 // index? Just return "I don't know".
318 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow));
320 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
322 case Intrinsic::bswap:
323 // bswap(bswap(x)) -> x
324 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
325 if (Operand->getIntrinsicID() == Intrinsic::bswap)
326 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
328 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
329 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
330 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
331 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
332 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
333 TI->getType()->getPrimitiveSizeInBits();
334 Value *CV = ConstantInt::get(Operand->getType(), C);
335 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
336 return new TruncInst(V, TI->getType());
341 case Intrinsic::powi:
342 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
345 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
348 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
349 // powi(x, -1) -> 1/x
350 if (Power->isAllOnesValue())
351 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
352 II->getArgOperand(0));
355 case Intrinsic::cttz: {
356 // If all bits below the first known one are known zero,
357 // this value is constant.
358 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
359 uint32_t BitWidth = IT->getBitWidth();
360 APInt KnownZero(BitWidth, 0);
361 APInt KnownOne(BitWidth, 0);
362 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
363 KnownZero, KnownOne);
364 unsigned TrailingZeros = KnownOne.countTrailingZeros();
365 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
366 if ((Mask & KnownZero) == Mask)
367 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
368 APInt(BitWidth, TrailingZeros)));
372 case Intrinsic::ctlz: {
373 // If all bits above the first known one are known zero,
374 // this value is constant.
375 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
376 uint32_t BitWidth = IT->getBitWidth();
377 APInt KnownZero(BitWidth, 0);
378 APInt KnownOne(BitWidth, 0);
379 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
380 KnownZero, KnownOne);
381 unsigned LeadingZeros = KnownOne.countLeadingZeros();
382 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
383 if ((Mask & KnownZero) == Mask)
384 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
385 APInt(BitWidth, LeadingZeros)));
389 case Intrinsic::uadd_with_overflow: {
390 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
391 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
392 uint32_t BitWidth = IT->getBitWidth();
393 APInt Mask = APInt::getSignBit(BitWidth);
394 APInt LHSKnownZero(BitWidth, 0);
395 APInt LHSKnownOne(BitWidth, 0);
396 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
397 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
398 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
400 if (LHSKnownNegative || LHSKnownPositive) {
401 APInt RHSKnownZero(BitWidth, 0);
402 APInt RHSKnownOne(BitWidth, 0);
403 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
404 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
405 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
406 if (LHSKnownNegative && RHSKnownNegative) {
407 // The sign bit is set in both cases: this MUST overflow.
408 // Create a simple add instruction, and insert it into the struct.
409 Value *Add = Builder->CreateAdd(LHS, RHS);
412 UndefValue::get(LHS->getType()),
413 ConstantInt::getTrue(II->getContext())
415 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
416 return InsertValueInst::Create(Struct, Add, 0);
419 if (LHSKnownPositive && RHSKnownPositive) {
420 // The sign bit is clear in both cases: this CANNOT overflow.
421 // Create a simple add instruction, and insert it into the struct.
422 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
425 UndefValue::get(LHS->getType()),
426 ConstantInt::getFalse(II->getContext())
428 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
429 return InsertValueInst::Create(Struct, Add, 0);
433 // FALL THROUGH uadd into sadd
434 case Intrinsic::sadd_with_overflow:
435 // Canonicalize constants into the RHS.
436 if (isa<Constant>(II->getArgOperand(0)) &&
437 !isa<Constant>(II->getArgOperand(1))) {
438 Value *LHS = II->getArgOperand(0);
439 II->setArgOperand(0, II->getArgOperand(1));
440 II->setArgOperand(1, LHS);
444 // X + undef -> undef
445 if (isa<UndefValue>(II->getArgOperand(1)))
446 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
448 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
449 // X + 0 -> {X, false}
452 UndefValue::get(II->getArgOperand(0)->getType()),
453 ConstantInt::getFalse(II->getContext())
455 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
456 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
460 case Intrinsic::usub_with_overflow:
461 case Intrinsic::ssub_with_overflow:
462 // undef - X -> undef
463 // X - undef -> undef
464 if (isa<UndefValue>(II->getArgOperand(0)) ||
465 isa<UndefValue>(II->getArgOperand(1)))
466 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
468 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
469 // X - 0 -> {X, false}
472 UndefValue::get(II->getArgOperand(0)->getType()),
473 ConstantInt::getFalse(II->getContext())
475 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
476 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
480 case Intrinsic::umul_with_overflow: {
481 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
482 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
483 APInt Mask = APInt::getAllOnesValue(BitWidth);
485 APInt LHSKnownZero(BitWidth, 0);
486 APInt LHSKnownOne(BitWidth, 0);
487 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
488 APInt RHSKnownZero(BitWidth, 0);
489 APInt RHSKnownOne(BitWidth, 0);
490 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
492 // Get the largest possible values for each operand.
493 APInt LHSMax = ~LHSKnownZero;
494 APInt RHSMax = ~RHSKnownZero;
496 // If multiplying the maximum values does not overflow then we can turn
497 // this into a plain NUW mul.
499 LHSMax.umul_ov(RHSMax, Overflow);
501 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
503 UndefValue::get(LHS->getType()),
506 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
507 return InsertValueInst::Create(Struct, Mul, 0);
510 case Intrinsic::smul_with_overflow:
511 // Canonicalize constants into the RHS.
512 if (isa<Constant>(II->getArgOperand(0)) &&
513 !isa<Constant>(II->getArgOperand(1))) {
514 Value *LHS = II->getArgOperand(0);
515 II->setArgOperand(0, II->getArgOperand(1));
516 II->setArgOperand(1, LHS);
520 // X * undef -> undef
521 if (isa<UndefValue>(II->getArgOperand(1)))
522 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
524 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
527 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
529 // X * 1 -> {X, false}
530 if (RHSI->equalsInt(1)) {
532 UndefValue::get(II->getArgOperand(0)->getType()),
533 ConstantInt::getFalse(II->getContext())
535 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
536 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
539 // [su]mul.with.overflow(X, 2) -> [su]add.with.overflow(X, X)
540 if (RHSI->equalsInt(2)) {
542 II->getIntrinsicID() == Intrinsic::smul_with_overflow ?
543 Intrinsic::sadd_with_overflow : Intrinsic::uadd_with_overflow;
545 Module *M = II->getParent()->getParent()->getParent();
546 const Type *Ty = RHSI->getType();
547 Function *F = Intrinsic::getDeclaration(M, Add, &Ty, 1);
549 Value *Ops[] = { II->getArgOperand(0), II->getArgOperand(0) };
550 return CallInst::Create(F, Ops, Ops+2);
554 case Intrinsic::ppc_altivec_lvx:
555 case Intrinsic::ppc_altivec_lvxl:
556 // Turn PPC lvx -> load if the pointer is known aligned.
557 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
558 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
559 PointerType::getUnqual(II->getType()));
560 return new LoadInst(Ptr);
563 case Intrinsic::ppc_altivec_stvx:
564 case Intrinsic::ppc_altivec_stvxl:
565 // Turn stvx -> store if the pointer is known aligned.
566 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
567 const Type *OpPtrTy =
568 PointerType::getUnqual(II->getArgOperand(0)->getType());
569 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
570 return new StoreInst(II->getArgOperand(0), Ptr);
573 case Intrinsic::x86_sse_storeu_ps:
574 case Intrinsic::x86_sse2_storeu_pd:
575 case Intrinsic::x86_sse2_storeu_dq:
576 // Turn X86 storeu -> store if the pointer is known aligned.
577 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
578 const Type *OpPtrTy =
579 PointerType::getUnqual(II->getArgOperand(1)->getType());
580 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
581 return new StoreInst(II->getArgOperand(1), Ptr);
585 case Intrinsic::x86_sse_cvtss2si:
586 case Intrinsic::x86_sse_cvtss2si64:
587 case Intrinsic::x86_sse_cvttss2si:
588 case Intrinsic::x86_sse_cvttss2si64:
589 case Intrinsic::x86_sse2_cvtsd2si:
590 case Intrinsic::x86_sse2_cvtsd2si64:
591 case Intrinsic::x86_sse2_cvttsd2si:
592 case Intrinsic::x86_sse2_cvttsd2si64: {
593 // These intrinsics only demand the 0th element of their input vectors. If
594 // we can simplify the input based on that, do so now.
596 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
597 APInt DemandedElts(VWidth, 1);
598 APInt UndefElts(VWidth, 0);
599 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
600 DemandedElts, UndefElts)) {
601 II->setArgOperand(0, V);
608 case Intrinsic::x86_sse41_pmovsxbw:
609 case Intrinsic::x86_sse41_pmovsxwd:
610 case Intrinsic::x86_sse41_pmovsxdq:
611 case Intrinsic::x86_sse41_pmovzxbw:
612 case Intrinsic::x86_sse41_pmovzxwd:
613 case Intrinsic::x86_sse41_pmovzxdq: {
614 // pmov{s|z}x ignores the upper half of their input vectors.
616 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
617 unsigned LowHalfElts = VWidth / 2;
618 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
619 APInt UndefElts(VWidth, 0);
620 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
623 II->setArgOperand(0, TmpV);
629 case Intrinsic::ppc_altivec_vperm:
630 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
631 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
632 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
634 // Check that all of the elements are integer constants or undefs.
635 bool AllEltsOk = true;
636 for (unsigned i = 0; i != 16; ++i) {
637 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
638 !isa<UndefValue>(Mask->getOperand(i))) {
645 // Cast the input vectors to byte vectors.
646 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
648 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
650 Value *Result = UndefValue::get(Op0->getType());
652 // Only extract each element once.
653 Value *ExtractedElts[32];
654 memset(ExtractedElts, 0, sizeof(ExtractedElts));
656 for (unsigned i = 0; i != 16; ++i) {
657 if (isa<UndefValue>(Mask->getOperand(i)))
659 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
660 Idx &= 31; // Match the hardware behavior.
662 if (ExtractedElts[Idx] == 0) {
664 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
665 ConstantInt::get(Type::getInt32Ty(II->getContext()),
666 Idx&15, false), "tmp");
669 // Insert this value into the result vector.
670 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
671 ConstantInt::get(Type::getInt32Ty(II->getContext()),
674 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
679 case Intrinsic::arm_neon_vld1:
680 case Intrinsic::arm_neon_vld2:
681 case Intrinsic::arm_neon_vld3:
682 case Intrinsic::arm_neon_vld4:
683 case Intrinsic::arm_neon_vld2lane:
684 case Intrinsic::arm_neon_vld3lane:
685 case Intrinsic::arm_neon_vld4lane:
686 case Intrinsic::arm_neon_vst1:
687 case Intrinsic::arm_neon_vst2:
688 case Intrinsic::arm_neon_vst3:
689 case Intrinsic::arm_neon_vst4:
690 case Intrinsic::arm_neon_vst2lane:
691 case Intrinsic::arm_neon_vst3lane:
692 case Intrinsic::arm_neon_vst4lane: {
693 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
694 unsigned AlignArg = II->getNumArgOperands() - 1;
695 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
696 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
697 II->setArgOperand(AlignArg,
698 ConstantInt::get(Type::getInt32Ty(II->getContext()),
705 case Intrinsic::stackrestore: {
706 // If the save is right next to the restore, remove the restore. This can
707 // happen when variable allocas are DCE'd.
708 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
709 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
710 BasicBlock::iterator BI = SS;
712 return EraseInstFromFunction(CI);
716 // Scan down this block to see if there is another stack restore in the
717 // same block without an intervening call/alloca.
718 BasicBlock::iterator BI = II;
719 TerminatorInst *TI = II->getParent()->getTerminator();
720 bool CannotRemove = false;
721 for (++BI; &*BI != TI; ++BI) {
722 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
726 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
727 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
728 // If there is a stackrestore below this one, remove this one.
729 if (II->getIntrinsicID() == Intrinsic::stackrestore)
730 return EraseInstFromFunction(CI);
731 // Otherwise, ignore the intrinsic.
733 // If we found a non-intrinsic call, we can't remove the stack
741 // If the stack restore is in a return/unwind block and if there are no
742 // allocas or calls between the restore and the return, nuke the restore.
743 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
744 return EraseInstFromFunction(CI);
749 return visitCallSite(II);
752 // InvokeInst simplification
754 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
755 return visitCallSite(&II);
758 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
759 /// passed through the varargs area, we can eliminate the use of the cast.
760 static bool isSafeToEliminateVarargsCast(const CallSite CS,
761 const CastInst * const CI,
762 const TargetData * const TD,
764 if (!CI->isLosslessCast())
767 // The size of ByVal arguments is derived from the type, so we
768 // can't change to a type with a different size. If the size were
769 // passed explicitly we could avoid this check.
770 if (!CS.paramHasAttr(ix, Attribute::ByVal))
774 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
775 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
776 if (!SrcTy->isSized() || !DstTy->isSized())
778 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
784 class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
787 void replaceCall(Value *With) {
788 NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
790 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
791 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
793 if (ConstantInt *SizeCI =
794 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
795 if (SizeCI->isAllOnesValue())
798 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
799 // If the length is 0 we don't know how long it is and so we can't
801 if (Len == 0) return false;
802 return SizeCI->getZExtValue() >= Len;
804 if (ConstantInt *Arg = dyn_cast<ConstantInt>(
805 CI->getArgOperand(SizeArgOp)))
806 return SizeCI->getZExtValue() >= Arg->getZExtValue();
811 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
812 Instruction *NewInstruction;
814 } // end anonymous namespace
816 // Try to fold some different type of calls here.
817 // Currently we're only working with the checking functions, memcpy_chk,
818 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
819 // strcat_chk and strncat_chk.
820 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
821 if (CI->getCalledFunction() == 0) return 0;
823 InstCombineFortifiedLibCalls Simplifier(this);
824 Simplifier.fold(CI, TD);
825 return Simplifier.NewInstruction;
828 // visitCallSite - Improvements for call and invoke instructions.
830 Instruction *InstCombiner::visitCallSite(CallSite CS) {
831 bool Changed = false;
833 // If the callee is a pointer to a function, attempt to move any casts to the
834 // arguments of the call/invoke.
835 Value *Callee = CS.getCalledValue();
836 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
839 if (Function *CalleeF = dyn_cast<Function>(Callee))
840 // If the call and callee calling conventions don't match, this call must
841 // be unreachable, as the call is undefined.
842 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
843 // Only do this for calls to a function with a body. A prototype may
844 // not actually end up matching the implementation's calling conv for a
845 // variety of reasons (e.g. it may be written in assembly).
846 !CalleeF->isDeclaration()) {
847 Instruction *OldCall = CS.getInstruction();
848 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
849 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
851 // If OldCall dues not return void then replaceAllUsesWith undef.
852 // This allows ValueHandlers and custom metadata to adjust itself.
853 if (!OldCall->getType()->isVoidTy())
854 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
855 if (isa<CallInst>(OldCall))
856 return EraseInstFromFunction(*OldCall);
858 // We cannot remove an invoke, because it would change the CFG, just
859 // change the callee to a null pointer.
860 cast<InvokeInst>(OldCall)->setCalledFunction(
861 Constant::getNullValue(CalleeF->getType()));
865 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
866 // This instruction is not reachable, just remove it. We insert a store to
867 // undef so that we know that this code is not reachable, despite the fact
868 // that we can't modify the CFG here.
869 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
870 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
871 CS.getInstruction());
873 // If CS does not return void then replaceAllUsesWith undef.
874 // This allows ValueHandlers and custom metadata to adjust itself.
875 if (!CS.getInstruction()->getType()->isVoidTy())
876 ReplaceInstUsesWith(*CS.getInstruction(),
877 UndefValue::get(CS.getInstruction()->getType()));
879 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
880 // Don't break the CFG, insert a dummy cond branch.
881 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
882 ConstantInt::getTrue(Callee->getContext()), II);
884 return EraseInstFromFunction(*CS.getInstruction());
887 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
888 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
889 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
890 return transformCallThroughTrampoline(CS);
892 const PointerType *PTy = cast<PointerType>(Callee->getType());
893 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
894 if (FTy->isVarArg()) {
895 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
896 // See if we can optimize any arguments passed through the varargs area of
898 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
899 E = CS.arg_end(); I != E; ++I, ++ix) {
900 CastInst *CI = dyn_cast<CastInst>(*I);
901 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
902 *I = CI->getOperand(0);
908 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
909 // Inline asm calls cannot throw - mark them 'nounwind'.
910 CS.setDoesNotThrow();
914 // Try to optimize the call if possible, we require TargetData for most of
915 // this. None of these calls are seen as possibly dead so go ahead and
916 // delete the instruction now.
917 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
918 Instruction *I = tryOptimizeCall(CI, TD);
919 // If we changed something return the result, etc. Otherwise let
920 // the fallthrough check.
921 if (I) return EraseInstFromFunction(*I);
924 return Changed ? CS.getInstruction() : 0;
927 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
928 // attempt to move the cast to the arguments of the call/invoke.
930 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
932 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
935 Instruction *Caller = CS.getInstruction();
936 const AttrListPtr &CallerPAL = CS.getAttributes();
938 // Okay, this is a cast from a function to a different type. Unless doing so
939 // would cause a type conversion of one of our arguments, change this call to
940 // be a direct call with arguments casted to the appropriate types.
942 const FunctionType *FT = Callee->getFunctionType();
943 const Type *OldRetTy = Caller->getType();
944 const Type *NewRetTy = FT->getReturnType();
946 if (NewRetTy->isStructTy())
947 return false; // TODO: Handle multiple return values.
949 // Check to see if we are changing the return type...
950 if (OldRetTy != NewRetTy) {
951 if (Callee->isDeclaration() &&
952 // Conversion is ok if changing from one pointer type to another or from
953 // a pointer to an integer of the same size.
954 !((OldRetTy->isPointerTy() || !TD ||
955 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
956 (NewRetTy->isPointerTy() || !TD ||
957 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
958 return false; // Cannot transform this return value.
960 if (!Caller->use_empty() &&
961 // void -> non-void is handled specially
962 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
963 return false; // Cannot transform this return value.
965 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
966 Attributes RAttrs = CallerPAL.getRetAttributes();
967 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
968 return false; // Attribute not compatible with transformed value.
971 // If the callsite is an invoke instruction, and the return value is used by
972 // a PHI node in a successor, we cannot change the return type of the call
973 // because there is no place to put the cast instruction (without breaking
974 // the critical edge). Bail out in this case.
975 if (!Caller->use_empty())
976 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
977 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
979 if (PHINode *PN = dyn_cast<PHINode>(*UI))
980 if (PN->getParent() == II->getNormalDest() ||
981 PN->getParent() == II->getUnwindDest())
985 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
986 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
988 CallSite::arg_iterator AI = CS.arg_begin();
989 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
990 const Type *ParamTy = FT->getParamType(i);
991 const Type *ActTy = (*AI)->getType();
993 if (!CastInst::isCastable(ActTy, ParamTy))
994 return false; // Cannot transform this parameter value.
996 unsigned Attrs = CallerPAL.getParamAttributes(i + 1);
997 if (Attrs & Attribute::typeIncompatible(ParamTy))
998 return false; // Attribute not compatible with transformed value.
1000 // If the parameter is passed as a byval argument, then we have to have a
1001 // sized type and the sized type has to have the same size as the old type.
1002 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
1003 const PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1004 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1007 const Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1008 if (TD->getTypeAllocSize(CurElTy) !=
1009 TD->getTypeAllocSize(ParamPTy->getElementType()))
1013 // Converting from one pointer type to another or between a pointer and an
1014 // integer of the same size is safe even if we do not have a body.
1015 bool isConvertible = ActTy == ParamTy ||
1016 (TD && ((ParamTy->isPointerTy() ||
1017 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1018 (ActTy->isPointerTy() ||
1019 ActTy == TD->getIntPtrType(Caller->getContext()))));
1020 if (Callee->isDeclaration() && !isConvertible) return false;
1023 if (Callee->isDeclaration()) {
1024 // Do not delete arguments unless we have a function body.
1025 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1028 // If the callee is just a declaration, don't change the varargsness of the
1029 // call. We don't want to introduce a varargs call where one doesn't
1031 const PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1032 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1036 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1037 !CallerPAL.isEmpty())
1038 // In this case we have more arguments than the new function type, but we
1039 // won't be dropping them. Check that these extra arguments have attributes
1040 // that are compatible with being a vararg call argument.
1041 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1042 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1044 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1045 if (PAttrs & Attribute::VarArgsIncompatible)
1050 // Okay, we decided that this is a safe thing to do: go ahead and start
1051 // inserting cast instructions as necessary.
1052 std::vector<Value*> Args;
1053 Args.reserve(NumActualArgs);
1054 SmallVector<AttributeWithIndex, 8> attrVec;
1055 attrVec.reserve(NumCommonArgs);
1057 // Get any return attributes.
1058 Attributes RAttrs = CallerPAL.getRetAttributes();
1060 // If the return value is not being used, the type may not be compatible
1061 // with the existing attributes. Wipe out any problematic attributes.
1062 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1064 // Add the new return attributes.
1066 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1068 AI = CS.arg_begin();
1069 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1070 const Type *ParamTy = FT->getParamType(i);
1071 if ((*AI)->getType() == ParamTy) {
1072 Args.push_back(*AI);
1074 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1075 false, ParamTy, false);
1076 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1079 // Add any parameter attributes.
1080 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1081 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1084 // If the function takes more arguments than the call was taking, add them
1086 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1087 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1089 // If we are removing arguments to the function, emit an obnoxious warning.
1090 if (FT->getNumParams() < NumActualArgs) {
1091 if (!FT->isVarArg()) {
1092 errs() << "WARNING: While resolving call to function '"
1093 << Callee->getName() << "' arguments were dropped!\n";
1095 // Add all of the arguments in their promoted form to the arg list.
1096 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1097 const Type *PTy = getPromotedType((*AI)->getType());
1098 if (PTy != (*AI)->getType()) {
1099 // Must promote to pass through va_arg area!
1100 Instruction::CastOps opcode =
1101 CastInst::getCastOpcode(*AI, false, PTy, false);
1102 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1104 Args.push_back(*AI);
1107 // Add any parameter attributes.
1108 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1109 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1114 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1115 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1117 if (NewRetTy->isVoidTy())
1118 Caller->setName(""); // Void type should not have a name.
1120 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1124 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1125 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1126 II->getUnwindDest(), Args.begin(), Args.end());
1128 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1129 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1131 CallInst *CI = cast<CallInst>(Caller);
1132 NC = Builder->CreateCall(Callee, Args.begin(), Args.end());
1134 if (CI->isTailCall())
1135 cast<CallInst>(NC)->setTailCall();
1136 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1137 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1140 // Insert a cast of the return type as necessary.
1142 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1143 if (!NV->getType()->isVoidTy()) {
1144 Instruction::CastOps opcode =
1145 CastInst::getCastOpcode(NC, false, OldRetTy, false);
1146 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1148 // If this is an invoke instruction, we should insert it after the first
1149 // non-phi, instruction in the normal successor block.
1150 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1151 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1152 InsertNewInstBefore(NC, *I);
1154 // Otherwise, it's a call, just insert cast right after the call.
1155 InsertNewInstBefore(NC, *Caller);
1157 Worklist.AddUsersToWorkList(*Caller);
1159 NV = UndefValue::get(Caller->getType());
1163 if (!Caller->use_empty())
1164 ReplaceInstUsesWith(*Caller, NV);
1166 EraseInstFromFunction(*Caller);
1170 // transformCallThroughTrampoline - Turn a call to a function created by the
1171 // init_trampoline intrinsic into a direct call to the underlying function.
1173 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1174 Value *Callee = CS.getCalledValue();
1175 const PointerType *PTy = cast<PointerType>(Callee->getType());
1176 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1177 const AttrListPtr &Attrs = CS.getAttributes();
1179 // If the call already has the 'nest' attribute somewhere then give up -
1180 // otherwise 'nest' would occur twice after splicing in the chain.
1181 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1184 IntrinsicInst *Tramp =
1185 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1187 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1188 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1189 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1191 const AttrListPtr &NestAttrs = NestF->getAttributes();
1192 if (!NestAttrs.isEmpty()) {
1193 unsigned NestIdx = 1;
1194 const Type *NestTy = 0;
1195 Attributes NestAttr = Attribute::None;
1197 // Look for a parameter marked with the 'nest' attribute.
1198 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1199 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1200 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1201 // Record the parameter type and any other attributes.
1203 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1208 Instruction *Caller = CS.getInstruction();
1209 std::vector<Value*> NewArgs;
1210 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1212 SmallVector<AttributeWithIndex, 8> NewAttrs;
1213 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1215 // Insert the nest argument into the call argument list, which may
1216 // mean appending it. Likewise for attributes.
1218 // Add any result attributes.
1219 if (Attributes Attr = Attrs.getRetAttributes())
1220 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1224 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1226 if (Idx == NestIdx) {
1227 // Add the chain argument and attributes.
1228 Value *NestVal = Tramp->getArgOperand(2);
1229 if (NestVal->getType() != NestTy)
1230 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1231 NewArgs.push_back(NestVal);
1232 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1238 // Add the original argument and attributes.
1239 NewArgs.push_back(*I);
1240 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1242 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1248 // Add any function attributes.
1249 if (Attributes Attr = Attrs.getFnAttributes())
1250 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1252 // The trampoline may have been bitcast to a bogus type (FTy).
1253 // Handle this by synthesizing a new function type, equal to FTy
1254 // with the chain parameter inserted.
1256 std::vector<const Type*> NewTypes;
1257 NewTypes.reserve(FTy->getNumParams()+1);
1259 // Insert the chain's type into the list of parameter types, which may
1260 // mean appending it.
1263 FunctionType::param_iterator I = FTy->param_begin(),
1264 E = FTy->param_end();
1268 // Add the chain's type.
1269 NewTypes.push_back(NestTy);
1274 // Add the original type.
1275 NewTypes.push_back(*I);
1281 // Replace the trampoline call with a direct call. Let the generic
1282 // code sort out any function type mismatches.
1283 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1285 Constant *NewCallee =
1286 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1287 NestF : ConstantExpr::getBitCast(NestF,
1288 PointerType::getUnqual(NewFTy));
1289 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1292 Instruction *NewCaller;
1293 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1294 NewCaller = InvokeInst::Create(NewCallee,
1295 II->getNormalDest(), II->getUnwindDest(),
1296 NewArgs.begin(), NewArgs.end());
1297 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1298 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1300 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end());
1301 if (cast<CallInst>(Caller)->isTailCall())
1302 cast<CallInst>(NewCaller)->setTailCall();
1303 cast<CallInst>(NewCaller)->
1304 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1305 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1312 // Replace the trampoline call with a direct call. Since there is no 'nest'
1313 // parameter, there is no need to adjust the argument list. Let the generic
1314 // code sort out any function type mismatches.
1315 Constant *NewCallee =
1316 NestF->getType() == PTy ? NestF :
1317 ConstantExpr::getBitCast(NestF, PTy);
1318 CS.setCalledFunction(NewCallee);
1319 return CS.getInstruction();