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 Type *getPromotedType(Type *Ty) {
26 if (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 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 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 (StructType *STy = dyn_cast<StructType>(SrcETy)) {
86 if (STy->getNumElements() == 1)
87 SrcETy = STy->getElementType(0);
90 } else if (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 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 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));
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 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(), Ops);
271 Op1 = GEP->getPointerOperand()->stripPointerCasts();
273 // Make sure we're not a constant offset from an external
275 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1))
276 if (!GV->hasDefinitiveInitializer()) break;
279 // If we've stripped down to a single global variable that we
280 // can know the size of then just return that.
281 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
282 if (GV->hasDefinitiveInitializer()) {
283 Constant *C = GV->getInitializer();
284 Size = TD->getTypeAllocSize(C->getType());
286 // Can't determine size of the GV.
287 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow);
288 return ReplaceInstUsesWith(CI, RetVal);
290 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
292 if (AI->getAllocatedType()->isSized()) {
293 Size = TD->getTypeAllocSize(AI->getAllocatedType());
294 if (AI->isArrayAllocation()) {
295 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
297 Size *= C->getZExtValue();
300 } else if (CallInst *MI = extractMallocCall(Op1)) {
301 // Get allocation size.
302 Type* MallocType = getMallocAllocatedType(MI);
303 if (MallocType && MallocType->isSized())
304 if (Value *NElems = getMallocArraySize(MI, TD, true))
305 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
306 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
309 // Do not return "I don't know" here. Later optimization passes could
310 // make it possible to evaluate objectsize to a constant.
315 // Out of bound reference? Negative index normalized to large
316 // index? Just return "I don't know".
317 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow));
319 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
321 case Intrinsic::bswap:
322 // bswap(bswap(x)) -> x
323 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
324 if (Operand->getIntrinsicID() == Intrinsic::bswap)
325 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
327 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
328 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
329 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
330 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
331 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
332 TI->getType()->getPrimitiveSizeInBits();
333 Value *CV = ConstantInt::get(Operand->getType(), C);
334 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
335 return new TruncInst(V, TI->getType());
340 case Intrinsic::powi:
341 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
344 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
347 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
348 // powi(x, -1) -> 1/x
349 if (Power->isAllOnesValue())
350 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
351 II->getArgOperand(0));
354 case Intrinsic::cttz: {
355 // If all bits below the first known one are known zero,
356 // this value is constant.
357 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
358 // FIXME: Try to simplify vectors of integers.
360 uint32_t BitWidth = IT->getBitWidth();
361 APInt KnownZero(BitWidth, 0);
362 APInt KnownOne(BitWidth, 0);
363 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
364 KnownZero, KnownOne);
365 unsigned TrailingZeros = KnownOne.countTrailingZeros();
366 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
367 if ((Mask & KnownZero) == Mask)
368 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
369 APInt(BitWidth, TrailingZeros)));
373 case Intrinsic::ctlz: {
374 // If all bits above the first known one are known zero,
375 // this value is constant.
376 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
377 // FIXME: Try to simplify vectors of integers.
379 uint32_t BitWidth = IT->getBitWidth();
380 APInt KnownZero(BitWidth, 0);
381 APInt KnownOne(BitWidth, 0);
382 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
383 KnownZero, KnownOne);
384 unsigned LeadingZeros = KnownOne.countLeadingZeros();
385 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
386 if ((Mask & KnownZero) == Mask)
387 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
388 APInt(BitWidth, LeadingZeros)));
392 case Intrinsic::uadd_with_overflow: {
393 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
394 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
395 uint32_t BitWidth = IT->getBitWidth();
396 APInt Mask = APInt::getSignBit(BitWidth);
397 APInt LHSKnownZero(BitWidth, 0);
398 APInt LHSKnownOne(BitWidth, 0);
399 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
400 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
401 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
403 if (LHSKnownNegative || LHSKnownPositive) {
404 APInt RHSKnownZero(BitWidth, 0);
405 APInt RHSKnownOne(BitWidth, 0);
406 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
407 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
408 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
409 if (LHSKnownNegative && RHSKnownNegative) {
410 // The sign bit is set in both cases: this MUST overflow.
411 // Create a simple add instruction, and insert it into the struct.
412 Value *Add = Builder->CreateAdd(LHS, RHS);
415 UndefValue::get(LHS->getType()),
416 ConstantInt::getTrue(II->getContext())
418 StructType *ST = cast<StructType>(II->getType());
419 Constant *Struct = ConstantStruct::get(ST, V);
420 return InsertValueInst::Create(Struct, Add, 0);
423 if (LHSKnownPositive && RHSKnownPositive) {
424 // The sign bit is clear in both cases: this CANNOT overflow.
425 // Create a simple add instruction, and insert it into the struct.
426 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
429 UndefValue::get(LHS->getType()),
430 ConstantInt::getFalse(II->getContext())
432 StructType *ST = cast<StructType>(II->getType());
433 Constant *Struct = ConstantStruct::get(ST, V);
434 return InsertValueInst::Create(Struct, Add, 0);
438 // FALL THROUGH uadd into sadd
439 case Intrinsic::sadd_with_overflow:
440 // Canonicalize constants into the RHS.
441 if (isa<Constant>(II->getArgOperand(0)) &&
442 !isa<Constant>(II->getArgOperand(1))) {
443 Value *LHS = II->getArgOperand(0);
444 II->setArgOperand(0, II->getArgOperand(1));
445 II->setArgOperand(1, LHS);
449 // X + undef -> undef
450 if (isa<UndefValue>(II->getArgOperand(1)))
451 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
453 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
454 // X + 0 -> {X, false}
457 UndefValue::get(II->getArgOperand(0)->getType()),
458 ConstantInt::getFalse(II->getContext())
461 ConstantStruct::get(cast<StructType>(II->getType()), V);
462 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
466 case Intrinsic::usub_with_overflow:
467 case Intrinsic::ssub_with_overflow:
468 // undef - X -> undef
469 // X - undef -> undef
470 if (isa<UndefValue>(II->getArgOperand(0)) ||
471 isa<UndefValue>(II->getArgOperand(1)))
472 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
474 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
475 // X - 0 -> {X, false}
478 UndefValue::get(II->getArgOperand(0)->getType()),
479 ConstantInt::getFalse(II->getContext())
482 ConstantStruct::get(cast<StructType>(II->getType()), V);
483 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
487 case Intrinsic::umul_with_overflow: {
488 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
489 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
490 APInt Mask = APInt::getAllOnesValue(BitWidth);
492 APInt LHSKnownZero(BitWidth, 0);
493 APInt LHSKnownOne(BitWidth, 0);
494 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
495 APInt RHSKnownZero(BitWidth, 0);
496 APInt RHSKnownOne(BitWidth, 0);
497 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
499 // Get the largest possible values for each operand.
500 APInt LHSMax = ~LHSKnownZero;
501 APInt RHSMax = ~RHSKnownZero;
503 // If multiplying the maximum values does not overflow then we can turn
504 // this into a plain NUW mul.
506 LHSMax.umul_ov(RHSMax, Overflow);
508 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
510 UndefValue::get(LHS->getType()),
513 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
514 return InsertValueInst::Create(Struct, Mul, 0);
517 case Intrinsic::smul_with_overflow:
518 // Canonicalize constants into the RHS.
519 if (isa<Constant>(II->getArgOperand(0)) &&
520 !isa<Constant>(II->getArgOperand(1))) {
521 Value *LHS = II->getArgOperand(0);
522 II->setArgOperand(0, II->getArgOperand(1));
523 II->setArgOperand(1, LHS);
527 // X * undef -> undef
528 if (isa<UndefValue>(II->getArgOperand(1)))
529 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
531 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
534 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
536 // X * 1 -> {X, false}
537 if (RHSI->equalsInt(1)) {
539 UndefValue::get(II->getArgOperand(0)->getType()),
540 ConstantInt::getFalse(II->getContext())
543 ConstantStruct::get(cast<StructType>(II->getType()), V);
544 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
548 case Intrinsic::ppc_altivec_lvx:
549 case Intrinsic::ppc_altivec_lvxl:
550 // Turn PPC lvx -> load if the pointer is known aligned.
551 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
552 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
553 PointerType::getUnqual(II->getType()));
554 return new LoadInst(Ptr);
557 case Intrinsic::ppc_altivec_stvx:
558 case Intrinsic::ppc_altivec_stvxl:
559 // Turn stvx -> store if the pointer is known aligned.
560 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
562 PointerType::getUnqual(II->getArgOperand(0)->getType());
563 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
564 return new StoreInst(II->getArgOperand(0), Ptr);
567 case Intrinsic::x86_sse_storeu_ps:
568 case Intrinsic::x86_sse2_storeu_pd:
569 case Intrinsic::x86_sse2_storeu_dq:
570 // Turn X86 storeu -> store if the pointer is known aligned.
571 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
573 PointerType::getUnqual(II->getArgOperand(1)->getType());
574 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
575 return new StoreInst(II->getArgOperand(1), Ptr);
579 case Intrinsic::x86_sse_cvtss2si:
580 case Intrinsic::x86_sse_cvtss2si64:
581 case Intrinsic::x86_sse_cvttss2si:
582 case Intrinsic::x86_sse_cvttss2si64:
583 case Intrinsic::x86_sse2_cvtsd2si:
584 case Intrinsic::x86_sse2_cvtsd2si64:
585 case Intrinsic::x86_sse2_cvttsd2si:
586 case Intrinsic::x86_sse2_cvttsd2si64: {
587 // These intrinsics only demand the 0th element of their input vectors. If
588 // we can simplify the input based on that, do so now.
590 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
591 APInt DemandedElts(VWidth, 1);
592 APInt UndefElts(VWidth, 0);
593 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
594 DemandedElts, UndefElts)) {
595 II->setArgOperand(0, V);
602 case Intrinsic::x86_sse41_pmovsxbw:
603 case Intrinsic::x86_sse41_pmovsxwd:
604 case Intrinsic::x86_sse41_pmovsxdq:
605 case Intrinsic::x86_sse41_pmovzxbw:
606 case Intrinsic::x86_sse41_pmovzxwd:
607 case Intrinsic::x86_sse41_pmovzxdq: {
608 // pmov{s|z}x ignores the upper half of their input vectors.
610 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
611 unsigned LowHalfElts = VWidth / 2;
612 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
613 APInt UndefElts(VWidth, 0);
614 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
617 II->setArgOperand(0, TmpV);
623 case Intrinsic::ppc_altivec_vperm:
624 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
625 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
626 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
628 // Check that all of the elements are integer constants or undefs.
629 bool AllEltsOk = true;
630 for (unsigned i = 0; i != 16; ++i) {
631 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
632 !isa<UndefValue>(Mask->getOperand(i))) {
639 // Cast the input vectors to byte vectors.
640 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
642 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
644 Value *Result = UndefValue::get(Op0->getType());
646 // Only extract each element once.
647 Value *ExtractedElts[32];
648 memset(ExtractedElts, 0, sizeof(ExtractedElts));
650 for (unsigned i = 0; i != 16; ++i) {
651 if (isa<UndefValue>(Mask->getOperand(i)))
653 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
654 Idx &= 31; // Match the hardware behavior.
656 if (ExtractedElts[Idx] == 0) {
658 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
659 ConstantInt::get(Type::getInt32Ty(II->getContext()),
660 Idx&15, false), "tmp");
663 // Insert this value into the result vector.
664 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
665 ConstantInt::get(Type::getInt32Ty(II->getContext()),
668 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
673 case Intrinsic::arm_neon_vld1:
674 case Intrinsic::arm_neon_vld2:
675 case Intrinsic::arm_neon_vld3:
676 case Intrinsic::arm_neon_vld4:
677 case Intrinsic::arm_neon_vld2lane:
678 case Intrinsic::arm_neon_vld3lane:
679 case Intrinsic::arm_neon_vld4lane:
680 case Intrinsic::arm_neon_vst1:
681 case Intrinsic::arm_neon_vst2:
682 case Intrinsic::arm_neon_vst3:
683 case Intrinsic::arm_neon_vst4:
684 case Intrinsic::arm_neon_vst2lane:
685 case Intrinsic::arm_neon_vst3lane:
686 case Intrinsic::arm_neon_vst4lane: {
687 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
688 unsigned AlignArg = II->getNumArgOperands() - 1;
689 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
690 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
691 II->setArgOperand(AlignArg,
692 ConstantInt::get(Type::getInt32Ty(II->getContext()),
699 case Intrinsic::stackrestore: {
700 // If the save is right next to the restore, remove the restore. This can
701 // happen when variable allocas are DCE'd.
702 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
703 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
704 BasicBlock::iterator BI = SS;
706 return EraseInstFromFunction(CI);
710 // Scan down this block to see if there is another stack restore in the
711 // same block without an intervening call/alloca.
712 BasicBlock::iterator BI = II;
713 TerminatorInst *TI = II->getParent()->getTerminator();
714 bool CannotRemove = false;
715 for (++BI; &*BI != TI; ++BI) {
716 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
720 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
721 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
722 // If there is a stackrestore below this one, remove this one.
723 if (II->getIntrinsicID() == Intrinsic::stackrestore)
724 return EraseInstFromFunction(CI);
725 // Otherwise, ignore the intrinsic.
727 // If we found a non-intrinsic call, we can't remove the stack
735 // If the stack restore is in a return, resume, or unwind block and if there
736 // are no allocas or calls between the restore and the return, nuke the
738 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI) ||
739 isa<UnwindInst>(TI)))
740 return EraseInstFromFunction(CI);
745 return visitCallSite(II);
748 // InvokeInst simplification
750 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
751 return visitCallSite(&II);
754 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
755 /// passed through the varargs area, we can eliminate the use of the cast.
756 static bool isSafeToEliminateVarargsCast(const CallSite CS,
757 const CastInst * const CI,
758 const TargetData * const TD,
760 if (!CI->isLosslessCast())
763 // The size of ByVal arguments is derived from the type, so we
764 // can't change to a type with a different size. If the size were
765 // passed explicitly we could avoid this check.
766 if (!CS.paramHasAttr(ix, Attribute::ByVal))
770 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
771 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
772 if (!SrcTy->isSized() || !DstTy->isSized())
774 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
780 class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
783 void replaceCall(Value *With) {
784 NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
786 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
787 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
789 if (ConstantInt *SizeCI =
790 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
791 if (SizeCI->isAllOnesValue())
794 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
795 // If the length is 0 we don't know how long it is and so we can't
797 if (Len == 0) return false;
798 return SizeCI->getZExtValue() >= Len;
800 if (ConstantInt *Arg = dyn_cast<ConstantInt>(
801 CI->getArgOperand(SizeArgOp)))
802 return SizeCI->getZExtValue() >= Arg->getZExtValue();
807 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
808 Instruction *NewInstruction;
810 } // end anonymous namespace
812 // Try to fold some different type of calls here.
813 // Currently we're only working with the checking functions, memcpy_chk,
814 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
815 // strcat_chk and strncat_chk.
816 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
817 if (CI->getCalledFunction() == 0) return 0;
819 InstCombineFortifiedLibCalls Simplifier(this);
820 Simplifier.fold(CI, TD);
821 return Simplifier.NewInstruction;
824 // visitCallSite - Improvements for call and invoke instructions.
826 Instruction *InstCombiner::visitCallSite(CallSite CS) {
827 bool Changed = false;
829 // If the callee is a pointer to a function, attempt to move any casts to the
830 // arguments of the call/invoke.
831 Value *Callee = CS.getCalledValue();
832 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
835 if (Function *CalleeF = dyn_cast<Function>(Callee))
836 // If the call and callee calling conventions don't match, this call must
837 // be unreachable, as the call is undefined.
838 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
839 // Only do this for calls to a function with a body. A prototype may
840 // not actually end up matching the implementation's calling conv for a
841 // variety of reasons (e.g. it may be written in assembly).
842 !CalleeF->isDeclaration()) {
843 Instruction *OldCall = CS.getInstruction();
844 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
845 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
847 // If OldCall dues not return void then replaceAllUsesWith undef.
848 // This allows ValueHandlers and custom metadata to adjust itself.
849 if (!OldCall->getType()->isVoidTy())
850 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
851 if (isa<CallInst>(OldCall))
852 return EraseInstFromFunction(*OldCall);
854 // We cannot remove an invoke, because it would change the CFG, just
855 // change the callee to a null pointer.
856 cast<InvokeInst>(OldCall)->setCalledFunction(
857 Constant::getNullValue(CalleeF->getType()));
861 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
862 // This instruction is not reachable, just remove it. We insert a store to
863 // undef so that we know that this code is not reachable, despite the fact
864 // that we can't modify the CFG here.
865 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
866 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
867 CS.getInstruction());
869 // If CS does not return void then replaceAllUsesWith undef.
870 // This allows ValueHandlers and custom metadata to adjust itself.
871 if (!CS.getInstruction()->getType()->isVoidTy())
872 ReplaceInstUsesWith(*CS.getInstruction(),
873 UndefValue::get(CS.getInstruction()->getType()));
875 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
876 // Don't break the CFG, insert a dummy cond branch.
877 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
878 ConstantInt::getTrue(Callee->getContext()), II);
880 return EraseInstFromFunction(*CS.getInstruction());
883 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
884 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
885 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
886 return transformCallThroughTrampoline(CS);
888 PointerType *PTy = cast<PointerType>(Callee->getType());
889 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
890 if (FTy->isVarArg()) {
891 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
892 // See if we can optimize any arguments passed through the varargs area of
894 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
895 E = CS.arg_end(); I != E; ++I, ++ix) {
896 CastInst *CI = dyn_cast<CastInst>(*I);
897 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
898 *I = CI->getOperand(0);
904 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
905 // Inline asm calls cannot throw - mark them 'nounwind'.
906 CS.setDoesNotThrow();
910 // Try to optimize the call if possible, we require TargetData for most of
911 // this. None of these calls are seen as possibly dead so go ahead and
912 // delete the instruction now.
913 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
914 Instruction *I = tryOptimizeCall(CI, TD);
915 // If we changed something return the result, etc. Otherwise let
916 // the fallthrough check.
917 if (I) return EraseInstFromFunction(*I);
920 return Changed ? CS.getInstruction() : 0;
923 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
924 // attempt to move the cast to the arguments of the call/invoke.
926 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
928 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
931 Instruction *Caller = CS.getInstruction();
932 const AttrListPtr &CallerPAL = CS.getAttributes();
934 // Okay, this is a cast from a function to a different type. Unless doing so
935 // would cause a type conversion of one of our arguments, change this call to
936 // be a direct call with arguments casted to the appropriate types.
938 FunctionType *FT = Callee->getFunctionType();
939 Type *OldRetTy = Caller->getType();
940 Type *NewRetTy = FT->getReturnType();
942 if (NewRetTy->isStructTy())
943 return false; // TODO: Handle multiple return values.
945 // Check to see if we are changing the return type...
946 if (OldRetTy != NewRetTy) {
947 if (Callee->isDeclaration() &&
948 // Conversion is ok if changing from one pointer type to another or from
949 // a pointer to an integer of the same size.
950 !((OldRetTy->isPointerTy() || !TD ||
951 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
952 (NewRetTy->isPointerTy() || !TD ||
953 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
954 return false; // Cannot transform this return value.
956 if (!Caller->use_empty() &&
957 // void -> non-void is handled specially
958 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
959 return false; // Cannot transform this return value.
961 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
962 Attributes RAttrs = CallerPAL.getRetAttributes();
963 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
964 return false; // Attribute not compatible with transformed value.
967 // If the callsite is an invoke instruction, and the return value is used by
968 // a PHI node in a successor, we cannot change the return type of the call
969 // because there is no place to put the cast instruction (without breaking
970 // the critical edge). Bail out in this case.
971 if (!Caller->use_empty())
972 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
973 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
975 if (PHINode *PN = dyn_cast<PHINode>(*UI))
976 if (PN->getParent() == II->getNormalDest() ||
977 PN->getParent() == II->getUnwindDest())
981 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
982 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
984 CallSite::arg_iterator AI = CS.arg_begin();
985 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
986 Type *ParamTy = FT->getParamType(i);
987 Type *ActTy = (*AI)->getType();
989 if (!CastInst::isCastable(ActTy, ParamTy))
990 return false; // Cannot transform this parameter value.
992 unsigned Attrs = CallerPAL.getParamAttributes(i + 1);
993 if (Attrs & Attribute::typeIncompatible(ParamTy))
994 return false; // Attribute not compatible with transformed value.
996 // If the parameter is passed as a byval argument, then we have to have a
997 // sized type and the sized type has to have the same size as the old type.
998 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
999 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1000 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1003 Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1004 if (TD->getTypeAllocSize(CurElTy) !=
1005 TD->getTypeAllocSize(ParamPTy->getElementType()))
1009 // Converting from one pointer type to another or between a pointer and an
1010 // integer of the same size is safe even if we do not have a body.
1011 bool isConvertible = ActTy == ParamTy ||
1012 (TD && ((ParamTy->isPointerTy() ||
1013 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1014 (ActTy->isPointerTy() ||
1015 ActTy == TD->getIntPtrType(Caller->getContext()))));
1016 if (Callee->isDeclaration() && !isConvertible) return false;
1019 if (Callee->isDeclaration()) {
1020 // Do not delete arguments unless we have a function body.
1021 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1024 // If the callee is just a declaration, don't change the varargsness of the
1025 // call. We don't want to introduce a varargs call where one doesn't
1027 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1028 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1032 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1033 !CallerPAL.isEmpty())
1034 // In this case we have more arguments than the new function type, but we
1035 // won't be dropping them. Check that these extra arguments have attributes
1036 // that are compatible with being a vararg call argument.
1037 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1038 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1040 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1041 if (PAttrs & Attribute::VarArgsIncompatible)
1046 // Okay, we decided that this is a safe thing to do: go ahead and start
1047 // inserting cast instructions as necessary.
1048 std::vector<Value*> Args;
1049 Args.reserve(NumActualArgs);
1050 SmallVector<AttributeWithIndex, 8> attrVec;
1051 attrVec.reserve(NumCommonArgs);
1053 // Get any return attributes.
1054 Attributes RAttrs = CallerPAL.getRetAttributes();
1056 // If the return value is not being used, the type may not be compatible
1057 // with the existing attributes. Wipe out any problematic attributes.
1058 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1060 // Add the new return attributes.
1062 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1064 AI = CS.arg_begin();
1065 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1066 Type *ParamTy = FT->getParamType(i);
1067 if ((*AI)->getType() == ParamTy) {
1068 Args.push_back(*AI);
1070 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1071 false, ParamTy, false);
1072 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1075 // Add any parameter attributes.
1076 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1077 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1080 // If the function takes more arguments than the call was taking, add them
1082 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1083 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1085 // If we are removing arguments to the function, emit an obnoxious warning.
1086 if (FT->getNumParams() < NumActualArgs) {
1087 if (!FT->isVarArg()) {
1088 errs() << "WARNING: While resolving call to function '"
1089 << Callee->getName() << "' arguments were dropped!\n";
1091 // Add all of the arguments in their promoted form to the arg list.
1092 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1093 Type *PTy = getPromotedType((*AI)->getType());
1094 if (PTy != (*AI)->getType()) {
1095 // Must promote to pass through va_arg area!
1096 Instruction::CastOps opcode =
1097 CastInst::getCastOpcode(*AI, false, PTy, false);
1098 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1100 Args.push_back(*AI);
1103 // Add any parameter attributes.
1104 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1105 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1110 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1111 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1113 if (NewRetTy->isVoidTy())
1114 Caller->setName(""); // Void type should not have a name.
1116 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1120 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1121 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1122 II->getUnwindDest(), Args);
1124 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1125 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1127 CallInst *CI = cast<CallInst>(Caller);
1128 NC = Builder->CreateCall(Callee, Args);
1130 if (CI->isTailCall())
1131 cast<CallInst>(NC)->setTailCall();
1132 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1133 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1136 // Insert a cast of the return type as necessary.
1138 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1139 if (!NV->getType()->isVoidTy()) {
1140 Instruction::CastOps opcode =
1141 CastInst::getCastOpcode(NC, false, OldRetTy, false);
1142 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1143 NC->setDebugLoc(Caller->getDebugLoc());
1145 // If this is an invoke instruction, we should insert it after the first
1146 // non-phi, instruction in the normal successor block.
1147 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1148 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1149 InsertNewInstBefore(NC, *I);
1151 // Otherwise, it's a call, just insert cast right after the call.
1152 InsertNewInstBefore(NC, *Caller);
1154 Worklist.AddUsersToWorkList(*Caller);
1156 NV = UndefValue::get(Caller->getType());
1160 if (!Caller->use_empty())
1161 ReplaceInstUsesWith(*Caller, NV);
1163 EraseInstFromFunction(*Caller);
1167 // transformCallThroughTrampoline - Turn a call to a function created by the
1168 // init_trampoline intrinsic into a direct call to the underlying function.
1170 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1171 Value *Callee = CS.getCalledValue();
1172 PointerType *PTy = cast<PointerType>(Callee->getType());
1173 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1174 const AttrListPtr &Attrs = CS.getAttributes();
1176 // If the call already has the 'nest' attribute somewhere then give up -
1177 // otherwise 'nest' would occur twice after splicing in the chain.
1178 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1181 IntrinsicInst *Tramp =
1182 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1184 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1185 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1186 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1188 const AttrListPtr &NestAttrs = NestF->getAttributes();
1189 if (!NestAttrs.isEmpty()) {
1190 unsigned NestIdx = 1;
1192 Attributes NestAttr = Attribute::None;
1194 // Look for a parameter marked with the 'nest' attribute.
1195 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1196 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1197 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1198 // Record the parameter type and any other attributes.
1200 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1205 Instruction *Caller = CS.getInstruction();
1206 std::vector<Value*> NewArgs;
1207 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1209 SmallVector<AttributeWithIndex, 8> NewAttrs;
1210 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1212 // Insert the nest argument into the call argument list, which may
1213 // mean appending it. Likewise for attributes.
1215 // Add any result attributes.
1216 if (Attributes Attr = Attrs.getRetAttributes())
1217 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1221 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1223 if (Idx == NestIdx) {
1224 // Add the chain argument and attributes.
1225 Value *NestVal = Tramp->getArgOperand(2);
1226 if (NestVal->getType() != NestTy)
1227 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1228 NewArgs.push_back(NestVal);
1229 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1235 // Add the original argument and attributes.
1236 NewArgs.push_back(*I);
1237 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1239 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1245 // Add any function attributes.
1246 if (Attributes Attr = Attrs.getFnAttributes())
1247 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1249 // The trampoline may have been bitcast to a bogus type (FTy).
1250 // Handle this by synthesizing a new function type, equal to FTy
1251 // with the chain parameter inserted.
1253 std::vector<Type*> NewTypes;
1254 NewTypes.reserve(FTy->getNumParams()+1);
1256 // Insert the chain's type into the list of parameter types, which may
1257 // mean appending it.
1260 FunctionType::param_iterator I = FTy->param_begin(),
1261 E = FTy->param_end();
1265 // Add the chain's type.
1266 NewTypes.push_back(NestTy);
1271 // Add the original type.
1272 NewTypes.push_back(*I);
1278 // Replace the trampoline call with a direct call. Let the generic
1279 // code sort out any function type mismatches.
1280 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1282 Constant *NewCallee =
1283 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1284 NestF : ConstantExpr::getBitCast(NestF,
1285 PointerType::getUnqual(NewFTy));
1286 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1289 Instruction *NewCaller;
1290 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1291 NewCaller = InvokeInst::Create(NewCallee,
1292 II->getNormalDest(), II->getUnwindDest(),
1294 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1295 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1297 NewCaller = CallInst::Create(NewCallee, NewArgs);
1298 if (cast<CallInst>(Caller)->isTailCall())
1299 cast<CallInst>(NewCaller)->setTailCall();
1300 cast<CallInst>(NewCaller)->
1301 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1302 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1309 // Replace the trampoline call with a direct call. Since there is no 'nest'
1310 // parameter, there is no need to adjust the argument list. Let the generic
1311 // code sort out any function type mismatches.
1312 Constant *NewCallee =
1313 NestF->getType() == PTy ? NestF :
1314 ConstantExpr::getBitCast(NestF, PTy);
1315 CS.setCalledFunction(NewCallee);
1316 return CS.getInstruction();