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 Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign);
115 InsertNewInstBefore(L, *MI);
116 InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), 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 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
158 Dest, false, Alignment), *MI);
160 // Set the size of the copy to 0, it will be deleted on the next iteration.
161 MI->setLength(Constant::getNullValue(LenC->getType()));
168 /// visitCallInst - CallInst simplification. This mostly only handles folding
169 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
170 /// the heavy lifting.
172 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
174 return visitFree(CI);
176 return visitMalloc(CI);
178 // If the caller function is nounwind, mark the call as nounwind, even if the
180 if (CI.getParent()->getParent()->doesNotThrow() &&
181 !CI.doesNotThrow()) {
182 CI.setDoesNotThrow();
186 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
187 if (!II) return visitCallSite(&CI);
189 // Intrinsics cannot occur in an invoke, so handle them here instead of in
191 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
192 bool Changed = false;
194 // memmove/cpy/set of zero bytes is a noop.
195 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
196 if (NumBytes->isNullValue())
197 return EraseInstFromFunction(CI);
199 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
200 if (CI->getZExtValue() == 1) {
201 // Replace the instruction with just byte operations. We would
202 // transform other cases to loads/stores, but we don't know if
203 // alignment is sufficient.
207 // No other transformations apply to volatile transfers.
208 if (MI->isVolatile())
211 // If we have a memmove and the source operation is a constant global,
212 // then the source and dest pointers can't alias, so we can change this
213 // into a call to memcpy.
214 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
215 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
216 if (GVSrc->isConstant()) {
217 Module *M = CI.getParent()->getParent()->getParent();
218 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
219 const Type *Tys[3] = { CI.getArgOperand(0)->getType(),
220 CI.getArgOperand(1)->getType(),
221 CI.getArgOperand(2)->getType() };
222 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
227 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
228 // memmove(x,x,size) -> noop.
229 if (MTI->getSource() == MTI->getDest())
230 return EraseInstFromFunction(CI);
233 // If we can determine a pointer alignment that is bigger than currently
234 // set, update the alignment.
235 if (isa<MemTransferInst>(MI)) {
236 if (Instruction *I = SimplifyMemTransfer(MI))
238 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
239 if (Instruction *I = SimplifyMemSet(MSI))
243 if (Changed) return II;
246 switch (II->getIntrinsicID()) {
248 case Intrinsic::objectsize: {
249 // We need target data for just about everything so depend on it.
252 const Type *ReturnTy = CI.getType();
253 uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL;
255 // Get to the real allocated thing and offset as fast as possible.
256 Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
259 uint64_t Size = -1ULL;
261 // Try to look through constant GEPs.
262 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) {
263 if (!GEP->hasAllConstantIndices()) break;
265 // Get the current byte offset into the thing. Use the original
266 // operand in case we're looking through a bitcast.
267 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
268 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(),
269 Ops.data(), Ops.size());
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 const 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 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
358 uint32_t BitWidth = IT->getBitWidth();
359 APInt KnownZero(BitWidth, 0);
360 APInt KnownOne(BitWidth, 0);
361 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
362 KnownZero, KnownOne);
363 unsigned TrailingZeros = KnownOne.countTrailingZeros();
364 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
365 if ((Mask & KnownZero) == Mask)
366 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
367 APInt(BitWidth, TrailingZeros)));
371 case Intrinsic::ctlz: {
372 // If all bits above the first known one are known zero,
373 // this value is constant.
374 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
375 uint32_t BitWidth = IT->getBitWidth();
376 APInt KnownZero(BitWidth, 0);
377 APInt KnownOne(BitWidth, 0);
378 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
379 KnownZero, KnownOne);
380 unsigned LeadingZeros = KnownOne.countLeadingZeros();
381 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
382 if ((Mask & KnownZero) == Mask)
383 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
384 APInt(BitWidth, LeadingZeros)));
388 case Intrinsic::uadd_with_overflow: {
389 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
390 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
391 uint32_t BitWidth = IT->getBitWidth();
392 APInt Mask = APInt::getSignBit(BitWidth);
393 APInt LHSKnownZero(BitWidth, 0);
394 APInt LHSKnownOne(BitWidth, 0);
395 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
396 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
397 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
399 if (LHSKnownNegative || LHSKnownPositive) {
400 APInt RHSKnownZero(BitWidth, 0);
401 APInt RHSKnownOne(BitWidth, 0);
402 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
403 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
404 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
405 if (LHSKnownNegative && RHSKnownNegative) {
406 // The sign bit is set in both cases: this MUST overflow.
407 // Create a simple add instruction, and insert it into the struct.
408 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
411 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
413 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
414 return InsertValueInst::Create(Struct, Add, 0);
417 if (LHSKnownPositive && RHSKnownPositive) {
418 // The sign bit is clear in both cases: this CANNOT overflow.
419 // Create a simple add instruction, and insert it into the struct.
420 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
423 UndefValue::get(LHS->getType()),
424 ConstantInt::getFalse(II->getContext())
426 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
427 return InsertValueInst::Create(Struct, Add, 0);
431 // FALL THROUGH uadd into sadd
432 case Intrinsic::sadd_with_overflow:
433 // Canonicalize constants into the RHS.
434 if (isa<Constant>(II->getArgOperand(0)) &&
435 !isa<Constant>(II->getArgOperand(1))) {
436 Value *LHS = II->getArgOperand(0);
437 II->setArgOperand(0, II->getArgOperand(1));
438 II->setArgOperand(1, LHS);
442 // X + undef -> undef
443 if (isa<UndefValue>(II->getArgOperand(1)))
444 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
446 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
447 // X + 0 -> {X, false}
450 UndefValue::get(II->getArgOperand(0)->getType()),
451 ConstantInt::getFalse(II->getContext())
453 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
454 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
458 case Intrinsic::usub_with_overflow:
459 case Intrinsic::ssub_with_overflow:
460 // undef - X -> undef
461 // X - undef -> undef
462 if (isa<UndefValue>(II->getArgOperand(0)) ||
463 isa<UndefValue>(II->getArgOperand(1)))
464 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
466 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
467 // X - 0 -> {X, false}
470 UndefValue::get(II->getArgOperand(0)->getType()),
471 ConstantInt::getFalse(II->getContext())
473 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
474 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
478 case Intrinsic::umul_with_overflow: {
479 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
480 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
481 APInt Mask = APInt::getAllOnesValue(BitWidth);
483 APInt LHSKnownZero(BitWidth, 0);
484 APInt LHSKnownOne(BitWidth, 0);
485 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
486 APInt RHSKnownZero(BitWidth, 0);
487 APInt RHSKnownOne(BitWidth, 0);
488 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
490 // Get the largest possible values for each operand.
491 APInt LHSMax = ~LHSKnownZero;
492 APInt RHSMax = ~RHSKnownZero;
494 // If multiplying the maximum values does not overflow then we can turn
495 // this into a plain NUW mul.
497 LHSMax.umul_ov(RHSMax, Overflow);
499 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
501 UndefValue::get(LHS->getType()),
504 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
505 return InsertValueInst::Create(Struct, Mul, 0);
508 case Intrinsic::smul_with_overflow:
509 // Canonicalize constants into the RHS.
510 if (isa<Constant>(II->getArgOperand(0)) &&
511 !isa<Constant>(II->getArgOperand(1))) {
512 Value *LHS = II->getArgOperand(0);
513 II->setArgOperand(0, II->getArgOperand(1));
514 II->setArgOperand(1, LHS);
518 // X * undef -> undef
519 if (isa<UndefValue>(II->getArgOperand(1)))
520 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
522 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
525 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
527 // X * 1 -> {X, false}
528 if (RHSI->equalsInt(1)) {
530 UndefValue::get(II->getArgOperand(0)->getType()),
531 ConstantInt::getFalse(II->getContext())
533 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
534 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
538 case Intrinsic::ppc_altivec_lvx:
539 case Intrinsic::ppc_altivec_lvxl:
540 // Turn PPC lvx -> load if the pointer is known aligned.
541 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
542 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
543 PointerType::getUnqual(II->getType()));
544 return new LoadInst(Ptr);
547 case Intrinsic::ppc_altivec_stvx:
548 case Intrinsic::ppc_altivec_stvxl:
549 // Turn stvx -> store if the pointer is known aligned.
550 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
551 const Type *OpPtrTy =
552 PointerType::getUnqual(II->getArgOperand(0)->getType());
553 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
554 return new StoreInst(II->getArgOperand(0), Ptr);
557 case Intrinsic::x86_sse_storeu_ps:
558 case Intrinsic::x86_sse2_storeu_pd:
559 case Intrinsic::x86_sse2_storeu_dq:
560 // Turn X86 storeu -> store if the pointer is known aligned.
561 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
562 const Type *OpPtrTy =
563 PointerType::getUnqual(II->getArgOperand(1)->getType());
564 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
565 return new StoreInst(II->getArgOperand(1), Ptr);
569 case Intrinsic::x86_sse_cvtss2si:
570 case Intrinsic::x86_sse_cvtss2si64:
571 case Intrinsic::x86_sse_cvttss2si:
572 case Intrinsic::x86_sse_cvttss2si64:
573 case Intrinsic::x86_sse2_cvtsd2si:
574 case Intrinsic::x86_sse2_cvtsd2si64:
575 case Intrinsic::x86_sse2_cvttsd2si:
576 case Intrinsic::x86_sse2_cvttsd2si64: {
577 // These intrinsics only demand the 0th element of their input vectors. If
578 // we can simplify the input based on that, do so now.
580 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
581 APInt DemandedElts(VWidth, 1);
582 APInt UndefElts(VWidth, 0);
583 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
584 DemandedElts, UndefElts)) {
585 II->setArgOperand(0, V);
592 case Intrinsic::x86_sse41_pmovsxbw:
593 case Intrinsic::x86_sse41_pmovsxwd:
594 case Intrinsic::x86_sse41_pmovsxdq:
595 case Intrinsic::x86_sse41_pmovzxbw:
596 case Intrinsic::x86_sse41_pmovzxwd:
597 case Intrinsic::x86_sse41_pmovzxdq: {
599 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
600 unsigned LowHalfElts = VWidth / 2;
601 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
602 APInt UndefElts(VWidth, 0);
603 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
606 II->setArgOperand(0, TmpV);
612 case Intrinsic::ppc_altivec_vperm:
613 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
614 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
615 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
617 // Check that all of the elements are integer constants or undefs.
618 bool AllEltsOk = true;
619 for (unsigned i = 0; i != 16; ++i) {
620 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
621 !isa<UndefValue>(Mask->getOperand(i))) {
628 // Cast the input vectors to byte vectors.
629 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
631 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
633 Value *Result = UndefValue::get(Op0->getType());
635 // Only extract each element once.
636 Value *ExtractedElts[32];
637 memset(ExtractedElts, 0, sizeof(ExtractedElts));
639 for (unsigned i = 0; i != 16; ++i) {
640 if (isa<UndefValue>(Mask->getOperand(i)))
642 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
643 Idx &= 31; // Match the hardware behavior.
645 if (ExtractedElts[Idx] == 0) {
647 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
648 ConstantInt::get(Type::getInt32Ty(II->getContext()),
649 Idx&15, false), "tmp");
652 // Insert this value into the result vector.
653 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
654 ConstantInt::get(Type::getInt32Ty(II->getContext()),
657 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
662 case Intrinsic::arm_neon_vld1:
663 case Intrinsic::arm_neon_vld2:
664 case Intrinsic::arm_neon_vld3:
665 case Intrinsic::arm_neon_vld4:
666 case Intrinsic::arm_neon_vld2lane:
667 case Intrinsic::arm_neon_vld3lane:
668 case Intrinsic::arm_neon_vld4lane:
669 case Intrinsic::arm_neon_vst1:
670 case Intrinsic::arm_neon_vst2:
671 case Intrinsic::arm_neon_vst3:
672 case Intrinsic::arm_neon_vst4:
673 case Intrinsic::arm_neon_vst2lane:
674 case Intrinsic::arm_neon_vst3lane:
675 case Intrinsic::arm_neon_vst4lane: {
676 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
677 unsigned AlignArg = II->getNumArgOperands() - 1;
678 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
679 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
680 II->setArgOperand(AlignArg,
681 ConstantInt::get(Type::getInt32Ty(II->getContext()),
688 case Intrinsic::stackrestore: {
689 // If the save is right next to the restore, remove the restore. This can
690 // happen when variable allocas are DCE'd.
691 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
692 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
693 BasicBlock::iterator BI = SS;
695 return EraseInstFromFunction(CI);
699 // Scan down this block to see if there is another stack restore in the
700 // same block without an intervening call/alloca.
701 BasicBlock::iterator BI = II;
702 TerminatorInst *TI = II->getParent()->getTerminator();
703 bool CannotRemove = false;
704 for (++BI; &*BI != TI; ++BI) {
705 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
709 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
710 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
711 // If there is a stackrestore below this one, remove this one.
712 if (II->getIntrinsicID() == Intrinsic::stackrestore)
713 return EraseInstFromFunction(CI);
714 // Otherwise, ignore the intrinsic.
716 // If we found a non-intrinsic call, we can't remove the stack
724 // If the stack restore is in a return/unwind block and if there are no
725 // allocas or calls between the restore and the return, nuke the restore.
726 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
727 return EraseInstFromFunction(CI);
732 return visitCallSite(II);
735 // InvokeInst simplification
737 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
738 return visitCallSite(&II);
741 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
742 /// passed through the varargs area, we can eliminate the use of the cast.
743 static bool isSafeToEliminateVarargsCast(const CallSite CS,
744 const CastInst * const CI,
745 const TargetData * const TD,
747 if (!CI->isLosslessCast())
750 // The size of ByVal arguments is derived from the type, so we
751 // can't change to a type with a different size. If the size were
752 // passed explicitly we could avoid this check.
753 if (!CS.paramHasAttr(ix, Attribute::ByVal))
757 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
758 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
759 if (!SrcTy->isSized() || !DstTy->isSized())
761 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
767 class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
770 void replaceCall(Value *With) {
771 NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
773 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
774 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
776 if (ConstantInt *SizeCI =
777 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
778 if (SizeCI->isAllOnesValue())
781 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
782 // If the length is 0 we don't know how long it is and so we can't
784 if (Len == 0) return false;
785 return SizeCI->getZExtValue() >= Len;
787 if (ConstantInt *Arg = dyn_cast<ConstantInt>(
788 CI->getArgOperand(SizeArgOp)))
789 return SizeCI->getZExtValue() >= Arg->getZExtValue();
794 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
795 Instruction *NewInstruction;
797 } // end anonymous namespace
799 // Try to fold some different type of calls here.
800 // Currently we're only working with the checking functions, memcpy_chk,
801 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
802 // strcat_chk and strncat_chk.
803 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
804 if (CI->getCalledFunction() == 0) return 0;
806 InstCombineFortifiedLibCalls Simplifier(this);
807 Simplifier.fold(CI, TD);
808 return Simplifier.NewInstruction;
811 // visitCallSite - Improvements for call and invoke instructions.
813 Instruction *InstCombiner::visitCallSite(CallSite CS) {
814 bool Changed = false;
816 // If the callee is a pointer to a function, attempt to move any casts to the
817 // arguments of the call/invoke.
818 Value *Callee = CS.getCalledValue();
819 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
822 if (Function *CalleeF = dyn_cast<Function>(Callee))
823 // If the call and callee calling conventions don't match, this call must
824 // be unreachable, as the call is undefined.
825 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
826 // Only do this for calls to a function with a body. A prototype may
827 // not actually end up matching the implementation's calling conv for a
828 // variety of reasons (e.g. it may be written in assembly).
829 !CalleeF->isDeclaration()) {
830 Instruction *OldCall = CS.getInstruction();
831 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
832 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
834 // If OldCall dues not return void then replaceAllUsesWith undef.
835 // This allows ValueHandlers and custom metadata to adjust itself.
836 if (!OldCall->getType()->isVoidTy())
837 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
838 if (isa<CallInst>(OldCall))
839 return EraseInstFromFunction(*OldCall);
841 // We cannot remove an invoke, because it would change the CFG, just
842 // change the callee to a null pointer.
843 cast<InvokeInst>(OldCall)->setCalledFunction(
844 Constant::getNullValue(CalleeF->getType()));
848 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
849 // This instruction is not reachable, just remove it. We insert a store to
850 // undef so that we know that this code is not reachable, despite the fact
851 // that we can't modify the CFG here.
852 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
853 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
854 CS.getInstruction());
856 // If CS does not return void then replaceAllUsesWith undef.
857 // This allows ValueHandlers and custom metadata to adjust itself.
858 if (!CS.getInstruction()->getType()->isVoidTy())
859 ReplaceInstUsesWith(*CS.getInstruction(),
860 UndefValue::get(CS.getInstruction()->getType()));
862 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
863 // Don't break the CFG, insert a dummy cond branch.
864 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
865 ConstantInt::getTrue(Callee->getContext()), II);
867 return EraseInstFromFunction(*CS.getInstruction());
870 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
871 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
872 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
873 return transformCallThroughTrampoline(CS);
875 const PointerType *PTy = cast<PointerType>(Callee->getType());
876 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
877 if (FTy->isVarArg()) {
878 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
879 // See if we can optimize any arguments passed through the varargs area of
881 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
882 E = CS.arg_end(); I != E; ++I, ++ix) {
883 CastInst *CI = dyn_cast<CastInst>(*I);
884 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
885 *I = CI->getOperand(0);
891 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
892 // Inline asm calls cannot throw - mark them 'nounwind'.
893 CS.setDoesNotThrow();
897 // Try to optimize the call if possible, we require TargetData for most of
898 // this. None of these calls are seen as possibly dead so go ahead and
899 // delete the instruction now.
900 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
901 Instruction *I = tryOptimizeCall(CI, TD);
902 // If we changed something return the result, etc. Otherwise let
903 // the fallthrough check.
904 if (I) return EraseInstFromFunction(*I);
907 return Changed ? CS.getInstruction() : 0;
910 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
911 // attempt to move the cast to the arguments of the call/invoke.
913 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
915 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
918 Instruction *Caller = CS.getInstruction();
919 const AttrListPtr &CallerPAL = CS.getAttributes();
921 // Okay, this is a cast from a function to a different type. Unless doing so
922 // would cause a type conversion of one of our arguments, change this call to
923 // be a direct call with arguments casted to the appropriate types.
925 const FunctionType *FT = Callee->getFunctionType();
926 const Type *OldRetTy = Caller->getType();
927 const Type *NewRetTy = FT->getReturnType();
929 if (NewRetTy->isStructTy())
930 return false; // TODO: Handle multiple return values.
932 // Check to see if we are changing the return type...
933 if (OldRetTy != NewRetTy) {
934 if (Callee->isDeclaration() &&
935 // Conversion is ok if changing from one pointer type to another or from
936 // a pointer to an integer of the same size.
937 !((OldRetTy->isPointerTy() || !TD ||
938 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
939 (NewRetTy->isPointerTy() || !TD ||
940 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
941 return false; // Cannot transform this return value.
943 if (!Caller->use_empty() &&
944 // void -> non-void is handled specially
945 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
946 return false; // Cannot transform this return value.
948 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
949 Attributes RAttrs = CallerPAL.getRetAttributes();
950 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
951 return false; // Attribute not compatible with transformed value.
954 // If the callsite is an invoke instruction, and the return value is used by
955 // a PHI node in a successor, we cannot change the return type of the call
956 // because there is no place to put the cast instruction (without breaking
957 // the critical edge). Bail out in this case.
958 if (!Caller->use_empty())
959 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
960 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
962 if (PHINode *PN = dyn_cast<PHINode>(*UI))
963 if (PN->getParent() == II->getNormalDest() ||
964 PN->getParent() == II->getUnwindDest())
968 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
969 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
971 CallSite::arg_iterator AI = CS.arg_begin();
972 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
973 const Type *ParamTy = FT->getParamType(i);
974 const Type *ActTy = (*AI)->getType();
976 if (!CastInst::isCastable(ActTy, ParamTy))
977 return false; // Cannot transform this parameter value.
979 unsigned Attrs = CallerPAL.getParamAttributes(i + 1);
980 if (Attrs & Attribute::typeIncompatible(ParamTy))
981 return false; // Attribute not compatible with transformed value.
983 // If the parameter is passed as a byval argument, then we have to have a
984 // sized type and the sized type has to have the same size as the old type.
985 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
986 const PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
987 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
990 const Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
991 if (TD->getTypeAllocSize(CurElTy) !=
992 TD->getTypeAllocSize(ParamPTy->getElementType()))
996 // Converting from one pointer type to another or between a pointer and an
997 // integer of the same size is safe even if we do not have a body.
998 bool isConvertible = ActTy == ParamTy ||
999 (TD && ((ParamTy->isPointerTy() ||
1000 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1001 (ActTy->isPointerTy() ||
1002 ActTy == TD->getIntPtrType(Caller->getContext()))));
1003 if (Callee->isDeclaration() && !isConvertible) return false;
1006 if (Callee->isDeclaration()) {
1007 // Do not delete arguments unless we have a function body.
1008 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1011 // If the callee is just a declaration, don't change the varargsness of the
1012 // call. We don't want to introduce a varargs call where one doesn't
1014 const PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1015 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1019 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1020 !CallerPAL.isEmpty())
1021 // In this case we have more arguments than the new function type, but we
1022 // won't be dropping them. Check that these extra arguments have attributes
1023 // that are compatible with being a vararg call argument.
1024 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1025 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1027 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1028 if (PAttrs & Attribute::VarArgsIncompatible)
1033 // Okay, we decided that this is a safe thing to do: go ahead and start
1034 // inserting cast instructions as necessary.
1035 std::vector<Value*> Args;
1036 Args.reserve(NumActualArgs);
1037 SmallVector<AttributeWithIndex, 8> attrVec;
1038 attrVec.reserve(NumCommonArgs);
1040 // Get any return attributes.
1041 Attributes RAttrs = CallerPAL.getRetAttributes();
1043 // If the return value is not being used, the type may not be compatible
1044 // with the existing attributes. Wipe out any problematic attributes.
1045 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1047 // Add the new return attributes.
1049 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1051 AI = CS.arg_begin();
1052 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1053 const Type *ParamTy = FT->getParamType(i);
1054 if ((*AI)->getType() == ParamTy) {
1055 Args.push_back(*AI);
1057 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1058 false, ParamTy, false);
1059 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1062 // Add any parameter attributes.
1063 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1064 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1067 // If the function takes more arguments than the call was taking, add them
1069 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1070 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1072 // If we are removing arguments to the function, emit an obnoxious warning.
1073 if (FT->getNumParams() < NumActualArgs) {
1074 if (!FT->isVarArg()) {
1075 errs() << "WARNING: While resolving call to function '"
1076 << Callee->getName() << "' arguments were dropped!\n";
1078 // Add all of the arguments in their promoted form to the arg list.
1079 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1080 const Type *PTy = getPromotedType((*AI)->getType());
1081 if (PTy != (*AI)->getType()) {
1082 // Must promote to pass through va_arg area!
1083 Instruction::CastOps opcode =
1084 CastInst::getCastOpcode(*AI, false, PTy, false);
1085 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1087 Args.push_back(*AI);
1090 // Add any parameter attributes.
1091 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1092 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1097 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1098 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1100 if (NewRetTy->isVoidTy())
1101 Caller->setName(""); // Void type should not have a name.
1103 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1107 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1108 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1109 II->getUnwindDest(), Args.begin(), Args.end());
1111 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1112 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1114 CallInst *CI = cast<CallInst>(Caller);
1115 NC = Builder->CreateCall(Callee, Args.begin(), Args.end());
1117 if (CI->isTailCall())
1118 cast<CallInst>(NC)->setTailCall();
1119 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1120 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1123 // Insert a cast of the return type as necessary.
1125 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1126 if (!NV->getType()->isVoidTy()) {
1127 Instruction::CastOps opcode =
1128 CastInst::getCastOpcode(NC, false, OldRetTy, false);
1129 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1131 // If this is an invoke instruction, we should insert it after the first
1132 // non-phi, instruction in the normal successor block.
1133 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1134 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1135 InsertNewInstBefore(NC, *I);
1137 // Otherwise, it's a call, just insert cast right after the call.
1138 InsertNewInstBefore(NC, *Caller);
1140 Worklist.AddUsersToWorkList(*Caller);
1142 NV = UndefValue::get(Caller->getType());
1146 if (!Caller->use_empty())
1147 ReplaceInstUsesWith(*Caller, NV);
1149 EraseInstFromFunction(*Caller);
1153 // transformCallThroughTrampoline - Turn a call to a function created by the
1154 // init_trampoline intrinsic into a direct call to the underlying function.
1156 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1157 Value *Callee = CS.getCalledValue();
1158 const PointerType *PTy = cast<PointerType>(Callee->getType());
1159 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1160 const AttrListPtr &Attrs = CS.getAttributes();
1162 // If the call already has the 'nest' attribute somewhere then give up -
1163 // otherwise 'nest' would occur twice after splicing in the chain.
1164 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1167 IntrinsicInst *Tramp =
1168 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1170 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1171 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1172 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1174 const AttrListPtr &NestAttrs = NestF->getAttributes();
1175 if (!NestAttrs.isEmpty()) {
1176 unsigned NestIdx = 1;
1177 const Type *NestTy = 0;
1178 Attributes NestAttr = Attribute::None;
1180 // Look for a parameter marked with the 'nest' attribute.
1181 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1182 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1183 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1184 // Record the parameter type and any other attributes.
1186 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1191 Instruction *Caller = CS.getInstruction();
1192 std::vector<Value*> NewArgs;
1193 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1195 SmallVector<AttributeWithIndex, 8> NewAttrs;
1196 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1198 // Insert the nest argument into the call argument list, which may
1199 // mean appending it. Likewise for attributes.
1201 // Add any result attributes.
1202 if (Attributes Attr = Attrs.getRetAttributes())
1203 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1207 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1209 if (Idx == NestIdx) {
1210 // Add the chain argument and attributes.
1211 Value *NestVal = Tramp->getArgOperand(2);
1212 if (NestVal->getType() != NestTy)
1213 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
1214 NewArgs.push_back(NestVal);
1215 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1221 // Add the original argument and attributes.
1222 NewArgs.push_back(*I);
1223 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1225 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1231 // Add any function attributes.
1232 if (Attributes Attr = Attrs.getFnAttributes())
1233 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1235 // The trampoline may have been bitcast to a bogus type (FTy).
1236 // Handle this by synthesizing a new function type, equal to FTy
1237 // with the chain parameter inserted.
1239 std::vector<const Type*> NewTypes;
1240 NewTypes.reserve(FTy->getNumParams()+1);
1242 // Insert the chain's type into the list of parameter types, which may
1243 // mean appending it.
1246 FunctionType::param_iterator I = FTy->param_begin(),
1247 E = FTy->param_end();
1251 // Add the chain's type.
1252 NewTypes.push_back(NestTy);
1257 // Add the original type.
1258 NewTypes.push_back(*I);
1264 // Replace the trampoline call with a direct call. Let the generic
1265 // code sort out any function type mismatches.
1266 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1268 Constant *NewCallee =
1269 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1270 NestF : ConstantExpr::getBitCast(NestF,
1271 PointerType::getUnqual(NewFTy));
1272 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1275 Instruction *NewCaller;
1276 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1277 NewCaller = InvokeInst::Create(NewCallee,
1278 II->getNormalDest(), II->getUnwindDest(),
1279 NewArgs.begin(), NewArgs.end(),
1280 Caller->getName(), Caller);
1281 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1282 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1284 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
1285 Caller->getName(), Caller);
1286 if (cast<CallInst>(Caller)->isTailCall())
1287 cast<CallInst>(NewCaller)->setTailCall();
1288 cast<CallInst>(NewCaller)->
1289 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1290 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1292 if (!Caller->getType()->isVoidTy())
1293 ReplaceInstUsesWith(*Caller, NewCaller);
1294 Caller->eraseFromParent();
1295 Worklist.Remove(Caller);
1300 // Replace the trampoline call with a direct call. Since there is no 'nest'
1301 // parameter, there is no need to adjust the argument list. Let the generic
1302 // code sort out any function type mismatches.
1303 Constant *NewCallee =
1304 NestF->getType() == PTy ? NestF :
1305 ConstantExpr::getBitCast(NestF, PTy);
1306 CS.setCalledFunction(NewCallee);
1307 return CS.getInstruction();