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/Support/CallSite.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/Transforms/Utils/BuildLibCalls.h"
19 #include "llvm/Transforms/Utils/Local.h"
22 /// getPromotedType - Return the specified type promoted as it would be to pass
23 /// though a va_arg area.
24 static Type *getPromotedType(Type *Ty) {
25 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
26 if (ITy->getBitWidth() < 32)
27 return Type::getInt32Ty(Ty->getContext());
33 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
34 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD);
35 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
36 unsigned MinAlign = std::min(DstAlign, SrcAlign);
37 unsigned CopyAlign = MI->getAlignment();
39 if (CopyAlign < MinAlign) {
40 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
45 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
47 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
48 if (MemOpLength == 0) return 0;
50 // Source and destination pointer types are always "i8*" for intrinsic. See
51 // if the size is something we can handle with a single primitive load/store.
52 // A single load+store correctly handles overlapping memory in the memmove
54 unsigned Size = MemOpLength->getZExtValue();
55 if (Size == 0) return MI; // Delete this mem transfer.
57 if (Size > 8 || (Size&(Size-1)))
58 return 0; // If not 1/2/4/8 bytes, exit.
60 // Use an integer load+store unless we can find something better.
62 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
64 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
66 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
67 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
68 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
70 // Memcpy forces the use of i8* for the source and destination. That means
71 // that if you're using memcpy to move one double around, you'll get a cast
72 // from double* to i8*. We'd much rather use a double load+store rather than
73 // an i64 load+store, here because this improves the odds that the source or
74 // dest address will be promotable. See if we can find a better type than the
76 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
77 if (StrippedDest != MI->getArgOperand(0)) {
78 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
80 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
81 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
82 // down through these levels if so.
83 while (!SrcETy->isSingleValueType()) {
84 if (StructType *STy = dyn_cast<StructType>(SrcETy)) {
85 if (STy->getNumElements() == 1)
86 SrcETy = STy->getElementType(0);
89 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
90 if (ATy->getNumElements() == 1)
91 SrcETy = ATy->getElementType();
98 if (SrcETy->isSingleValueType()) {
99 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
100 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
106 // If the memcpy/memmove provides better alignment info than we can
108 SrcAlign = std::max(SrcAlign, CopyAlign);
109 DstAlign = std::max(DstAlign, CopyAlign);
111 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
112 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
113 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
114 L->setAlignment(SrcAlign);
115 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
116 S->setAlignment(DstAlign);
118 // Set the size of the copy to 0, it will be deleted on the next iteration.
119 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
123 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
124 unsigned Alignment = getKnownAlignment(MI->getDest(), TD);
125 if (MI->getAlignment() < Alignment) {
126 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
131 // Extract the length and alignment and fill if they are constant.
132 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
133 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
134 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
136 uint64_t Len = LenC->getZExtValue();
137 Alignment = MI->getAlignment();
139 // If the length is zero, this is a no-op
140 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
142 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
143 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
144 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
146 Value *Dest = MI->getDest();
147 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
148 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
149 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
151 // Alignment 0 is identity for alignment 1 for memset, but not store.
152 if (Alignment == 0) Alignment = 1;
154 // Extract the fill value and store.
155 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
156 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
158 S->setAlignment(Alignment);
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 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));
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 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(), Ops);
270 Op1 = GEP->getPointerOperand()->stripPointerCasts();
272 // Make sure we're not a constant offset from an external
274 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1))
275 if (!GV->hasDefinitiveInitializer()) break;
278 // If we've stripped down to a single global variable that we
279 // can know the size of then just return that.
280 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
281 if (GV->hasDefinitiveInitializer()) {
282 Constant *C = GV->getInitializer();
283 Size = TD->getTypeAllocSize(C->getType());
285 // Can't determine size of the GV.
286 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow);
287 return ReplaceInstUsesWith(CI, RetVal);
289 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
291 if (AI->getAllocatedType()->isSized()) {
292 Size = TD->getTypeAllocSize(AI->getAllocatedType());
293 if (AI->isArrayAllocation()) {
294 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
296 Size *= C->getZExtValue();
299 } else if (CallInst *MI = extractMallocCall(Op1)) {
300 // Get allocation size.
301 Type* MallocType = getMallocAllocatedType(MI);
302 if (MallocType && MallocType->isSized())
303 if (Value *NElems = getMallocArraySize(MI, TD, true))
304 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
305 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
308 // Do not return "I don't know" here. Later optimization passes could
309 // make it possible to evaluate objectsize to a constant.
314 // Out of bound reference? Negative index normalized to large
315 // index? Just return "I don't know".
316 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow));
318 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
320 case Intrinsic::bswap:
321 // bswap(bswap(x)) -> x
322 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
323 if (Operand->getIntrinsicID() == Intrinsic::bswap)
324 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
326 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
327 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
328 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
329 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
330 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
331 TI->getType()->getPrimitiveSizeInBits();
332 Value *CV = ConstantInt::get(Operand->getType(), C);
333 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
334 return new TruncInst(V, TI->getType());
339 case Intrinsic::powi:
340 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
343 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
346 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
347 // powi(x, -1) -> 1/x
348 if (Power->isAllOnesValue())
349 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
350 II->getArgOperand(0));
353 case Intrinsic::cttz: {
354 // If all bits below the first known one are known zero,
355 // this value is constant.
356 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
357 // FIXME: Try to simplify vectors of integers.
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 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
376 // FIXME: Try to simplify vectors of integers.
378 uint32_t BitWidth = IT->getBitWidth();
379 APInt KnownZero(BitWidth, 0);
380 APInt KnownOne(BitWidth, 0);
381 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
382 KnownZero, KnownOne);
383 unsigned LeadingZeros = KnownOne.countLeadingZeros();
384 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
385 if ((Mask & KnownZero) == Mask)
386 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
387 APInt(BitWidth, LeadingZeros)));
391 case Intrinsic::uadd_with_overflow: {
392 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
393 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
394 uint32_t BitWidth = IT->getBitWidth();
395 APInt Mask = APInt::getSignBit(BitWidth);
396 APInt LHSKnownZero(BitWidth, 0);
397 APInt LHSKnownOne(BitWidth, 0);
398 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
399 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
400 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
402 if (LHSKnownNegative || LHSKnownPositive) {
403 APInt RHSKnownZero(BitWidth, 0);
404 APInt RHSKnownOne(BitWidth, 0);
405 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
406 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
407 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
408 if (LHSKnownNegative && RHSKnownNegative) {
409 // The sign bit is set in both cases: this MUST overflow.
410 // Create a simple add instruction, and insert it into the struct.
411 Value *Add = Builder->CreateAdd(LHS, RHS);
414 UndefValue::get(LHS->getType()),
415 ConstantInt::getTrue(II->getContext())
417 StructType *ST = cast<StructType>(II->getType());
418 Constant *Struct = ConstantStruct::get(ST, V);
419 return InsertValueInst::Create(Struct, Add, 0);
422 if (LHSKnownPositive && RHSKnownPositive) {
423 // The sign bit is clear in both cases: this CANNOT overflow.
424 // Create a simple add instruction, and insert it into the struct.
425 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
428 UndefValue::get(LHS->getType()),
429 ConstantInt::getFalse(II->getContext())
431 StructType *ST = cast<StructType>(II->getType());
432 Constant *Struct = ConstantStruct::get(ST, V);
433 return InsertValueInst::Create(Struct, Add, 0);
437 // FALL THROUGH uadd into sadd
438 case Intrinsic::sadd_with_overflow:
439 // Canonicalize constants into the RHS.
440 if (isa<Constant>(II->getArgOperand(0)) &&
441 !isa<Constant>(II->getArgOperand(1))) {
442 Value *LHS = II->getArgOperand(0);
443 II->setArgOperand(0, II->getArgOperand(1));
444 II->setArgOperand(1, LHS);
448 // X + undef -> undef
449 if (isa<UndefValue>(II->getArgOperand(1)))
450 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
452 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
453 // X + 0 -> {X, false}
456 UndefValue::get(II->getArgOperand(0)->getType()),
457 ConstantInt::getFalse(II->getContext())
460 ConstantStruct::get(cast<StructType>(II->getType()), V);
461 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
465 case Intrinsic::usub_with_overflow:
466 case Intrinsic::ssub_with_overflow:
467 // undef - X -> undef
468 // X - undef -> undef
469 if (isa<UndefValue>(II->getArgOperand(0)) ||
470 isa<UndefValue>(II->getArgOperand(1)))
471 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
473 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
474 // X - 0 -> {X, false}
477 UndefValue::get(II->getArgOperand(0)->getType()),
478 ConstantInt::getFalse(II->getContext())
481 ConstantStruct::get(cast<StructType>(II->getType()), V);
482 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
486 case Intrinsic::umul_with_overflow: {
487 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
488 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
489 APInt Mask = APInt::getAllOnesValue(BitWidth);
491 APInt LHSKnownZero(BitWidth, 0);
492 APInt LHSKnownOne(BitWidth, 0);
493 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
494 APInt RHSKnownZero(BitWidth, 0);
495 APInt RHSKnownOne(BitWidth, 0);
496 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
498 // Get the largest possible values for each operand.
499 APInt LHSMax = ~LHSKnownZero;
500 APInt RHSMax = ~RHSKnownZero;
502 // If multiplying the maximum values does not overflow then we can turn
503 // this into a plain NUW mul.
505 LHSMax.umul_ov(RHSMax, Overflow);
507 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
509 UndefValue::get(LHS->getType()),
512 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
513 return InsertValueInst::Create(Struct, Mul, 0);
516 case Intrinsic::smul_with_overflow:
517 // Canonicalize constants into the RHS.
518 if (isa<Constant>(II->getArgOperand(0)) &&
519 !isa<Constant>(II->getArgOperand(1))) {
520 Value *LHS = II->getArgOperand(0);
521 II->setArgOperand(0, II->getArgOperand(1));
522 II->setArgOperand(1, LHS);
526 // X * undef -> undef
527 if (isa<UndefValue>(II->getArgOperand(1)))
528 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
530 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
533 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
535 // X * 1 -> {X, false}
536 if (RHSI->equalsInt(1)) {
538 UndefValue::get(II->getArgOperand(0)->getType()),
539 ConstantInt::getFalse(II->getContext())
542 ConstantStruct::get(cast<StructType>(II->getType()), V);
543 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
547 case Intrinsic::ppc_altivec_lvx:
548 case Intrinsic::ppc_altivec_lvxl:
549 // Turn PPC lvx -> load if the pointer is known aligned.
550 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
551 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
552 PointerType::getUnqual(II->getType()));
553 return new LoadInst(Ptr);
556 case Intrinsic::ppc_altivec_stvx:
557 case Intrinsic::ppc_altivec_stvxl:
558 // Turn stvx -> store if the pointer is known aligned.
559 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
561 PointerType::getUnqual(II->getArgOperand(0)->getType());
562 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
563 return new StoreInst(II->getArgOperand(0), Ptr);
566 case Intrinsic::x86_sse_storeu_ps:
567 case Intrinsic::x86_sse2_storeu_pd:
568 case Intrinsic::x86_sse2_storeu_dq:
569 // Turn X86 storeu -> store if the pointer is known aligned.
570 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
572 PointerType::getUnqual(II->getArgOperand(1)->getType());
573 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
574 return new StoreInst(II->getArgOperand(1), Ptr);
578 case Intrinsic::x86_sse_cvtss2si:
579 case Intrinsic::x86_sse_cvtss2si64:
580 case Intrinsic::x86_sse_cvttss2si:
581 case Intrinsic::x86_sse_cvttss2si64:
582 case Intrinsic::x86_sse2_cvtsd2si:
583 case Intrinsic::x86_sse2_cvtsd2si64:
584 case Intrinsic::x86_sse2_cvttsd2si:
585 case Intrinsic::x86_sse2_cvttsd2si64: {
586 // These intrinsics only demand the 0th element of their input vectors. If
587 // we can simplify the input based on that, do so now.
589 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
590 APInt DemandedElts(VWidth, 1);
591 APInt UndefElts(VWidth, 0);
592 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
593 DemandedElts, UndefElts)) {
594 II->setArgOperand(0, V);
601 case Intrinsic::x86_sse41_pmovsxbw:
602 case Intrinsic::x86_sse41_pmovsxwd:
603 case Intrinsic::x86_sse41_pmovsxdq:
604 case Intrinsic::x86_sse41_pmovzxbw:
605 case Intrinsic::x86_sse41_pmovzxwd:
606 case Intrinsic::x86_sse41_pmovzxdq: {
607 // pmov{s|z}x ignores the upper half of their input vectors.
609 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
610 unsigned LowHalfElts = VWidth / 2;
611 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
612 APInt UndefElts(VWidth, 0);
613 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
616 II->setArgOperand(0, TmpV);
622 case Intrinsic::ppc_altivec_vperm:
623 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
624 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
625 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
627 // Check that all of the elements are integer constants or undefs.
628 bool AllEltsOk = true;
629 for (unsigned i = 0; i != 16; ++i) {
630 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
631 !isa<UndefValue>(Mask->getOperand(i))) {
638 // Cast the input vectors to byte vectors.
639 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
641 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
643 Value *Result = UndefValue::get(Op0->getType());
645 // Only extract each element once.
646 Value *ExtractedElts[32];
647 memset(ExtractedElts, 0, sizeof(ExtractedElts));
649 for (unsigned i = 0; i != 16; ++i) {
650 if (isa<UndefValue>(Mask->getOperand(i)))
652 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
653 Idx &= 31; // Match the hardware behavior.
655 if (ExtractedElts[Idx] == 0) {
657 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
658 Builder->getInt32(Idx&15));
661 // Insert this value into the result vector.
662 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
663 Builder->getInt32(i));
665 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
670 case Intrinsic::arm_neon_vld1:
671 case Intrinsic::arm_neon_vld2:
672 case Intrinsic::arm_neon_vld3:
673 case Intrinsic::arm_neon_vld4:
674 case Intrinsic::arm_neon_vld2lane:
675 case Intrinsic::arm_neon_vld3lane:
676 case Intrinsic::arm_neon_vld4lane:
677 case Intrinsic::arm_neon_vst1:
678 case Intrinsic::arm_neon_vst2:
679 case Intrinsic::arm_neon_vst3:
680 case Intrinsic::arm_neon_vst4:
681 case Intrinsic::arm_neon_vst2lane:
682 case Intrinsic::arm_neon_vst3lane:
683 case Intrinsic::arm_neon_vst4lane: {
684 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
685 unsigned AlignArg = II->getNumArgOperands() - 1;
686 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
687 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
688 II->setArgOperand(AlignArg,
689 ConstantInt::get(Type::getInt32Ty(II->getContext()),
696 case Intrinsic::stackrestore: {
697 // If the save is right next to the restore, remove the restore. This can
698 // happen when variable allocas are DCE'd.
699 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
700 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
701 BasicBlock::iterator BI = SS;
703 return EraseInstFromFunction(CI);
707 // Scan down this block to see if there is another stack restore in the
708 // same block without an intervening call/alloca.
709 BasicBlock::iterator BI = II;
710 TerminatorInst *TI = II->getParent()->getTerminator();
711 bool CannotRemove = false;
712 for (++BI; &*BI != TI; ++BI) {
713 if (isa<AllocaInst>(BI) || isMalloc(BI)) {
717 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
718 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
719 // If there is a stackrestore below this one, remove this one.
720 if (II->getIntrinsicID() == Intrinsic::stackrestore)
721 return EraseInstFromFunction(CI);
722 // Otherwise, ignore the intrinsic.
724 // If we found a non-intrinsic call, we can't remove the stack
732 // If the stack restore is in a return, resume, or unwind block and if there
733 // are no allocas or calls between the restore and the return, nuke the
735 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI) ||
736 isa<UnwindInst>(TI)))
737 return EraseInstFromFunction(CI);
742 return visitCallSite(II);
745 // InvokeInst simplification
747 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
748 return visitCallSite(&II);
751 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
752 /// passed through the varargs area, we can eliminate the use of the cast.
753 static bool isSafeToEliminateVarargsCast(const CallSite CS,
754 const CastInst * const CI,
755 const TargetData * const TD,
757 if (!CI->isLosslessCast())
760 // The size of ByVal arguments is derived from the type, so we
761 // can't change to a type with a different size. If the size were
762 // passed explicitly we could avoid this check.
763 if (!CS.isByValArgument(ix))
767 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
768 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
769 if (!SrcTy->isSized() || !DstTy->isSized())
771 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
777 class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
780 void replaceCall(Value *With) {
781 NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
783 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
784 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
786 if (ConstantInt *SizeCI =
787 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
788 if (SizeCI->isAllOnesValue())
791 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
792 // If the length is 0 we don't know how long it is and so we can't
794 if (Len == 0) return false;
795 return SizeCI->getZExtValue() >= Len;
797 if (ConstantInt *Arg = dyn_cast<ConstantInt>(
798 CI->getArgOperand(SizeArgOp)))
799 return SizeCI->getZExtValue() >= Arg->getZExtValue();
804 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
805 Instruction *NewInstruction;
807 } // end anonymous namespace
809 // Try to fold some different type of calls here.
810 // Currently we're only working with the checking functions, memcpy_chk,
811 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
812 // strcat_chk and strncat_chk.
813 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
814 if (CI->getCalledFunction() == 0) return 0;
816 InstCombineFortifiedLibCalls Simplifier(this);
817 Simplifier.fold(CI, TD);
818 return Simplifier.NewInstruction;
821 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
822 // Strip off at most one level of pointer casts, looking for an alloca. This
823 // is good enough in practice and simpler than handling any number of casts.
824 Value *Underlying = TrampMem->stripPointerCasts();
825 if (Underlying != TrampMem &&
826 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
828 if (!isa<AllocaInst>(Underlying))
831 IntrinsicInst *InitTrampoline = 0;
832 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
834 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
837 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
839 // More than one init_trampoline writes to this value. Give up.
844 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
845 // Allow any number of calls to adjust.trampoline.
850 // No call to init.trampoline found.
854 // Check that the alloca is being used in the expected way.
855 if (InitTrampoline->getOperand(0) != TrampMem)
858 return InitTrampoline;
861 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
863 // Visit all the previous instructions in the basic block, and try to find a
864 // init.trampoline which has a direct path to the adjust.trampoline.
865 for (BasicBlock::iterator I = AdjustTramp,
866 E = AdjustTramp->getParent()->begin(); I != E; ) {
867 Instruction *Inst = --I;
868 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
869 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
870 II->getOperand(0) == TrampMem)
872 if (Inst->mayWriteToMemory())
878 // Given a call to llvm.adjust.trampoline, find and return the corresponding
879 // call to llvm.init.trampoline if the call to the trampoline can be optimized
880 // to a direct call to a function. Otherwise return NULL.
882 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
883 Callee = Callee->stripPointerCasts();
884 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
886 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
889 Value *TrampMem = AdjustTramp->getOperand(0);
891 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
893 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
898 // visitCallSite - Improvements for call and invoke instructions.
900 Instruction *InstCombiner::visitCallSite(CallSite CS) {
901 bool Changed = false;
903 // If the callee is a pointer to a function, attempt to move any casts to the
904 // arguments of the call/invoke.
905 Value *Callee = CS.getCalledValue();
906 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
909 if (Function *CalleeF = dyn_cast<Function>(Callee))
910 // If the call and callee calling conventions don't match, this call must
911 // be unreachable, as the call is undefined.
912 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
913 // Only do this for calls to a function with a body. A prototype may
914 // not actually end up matching the implementation's calling conv for a
915 // variety of reasons (e.g. it may be written in assembly).
916 !CalleeF->isDeclaration()) {
917 Instruction *OldCall = CS.getInstruction();
918 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
919 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
921 // If OldCall dues not return void then replaceAllUsesWith undef.
922 // This allows ValueHandlers and custom metadata to adjust itself.
923 if (!OldCall->getType()->isVoidTy())
924 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
925 if (isa<CallInst>(OldCall))
926 return EraseInstFromFunction(*OldCall);
928 // We cannot remove an invoke, because it would change the CFG, just
929 // change the callee to a null pointer.
930 cast<InvokeInst>(OldCall)->setCalledFunction(
931 Constant::getNullValue(CalleeF->getType()));
935 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
936 // This instruction is not reachable, just remove it. We insert a store to
937 // undef so that we know that this code is not reachable, despite the fact
938 // that we can't modify the CFG here.
939 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
940 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
941 CS.getInstruction());
943 // If CS does not return void then replaceAllUsesWith undef.
944 // This allows ValueHandlers and custom metadata to adjust itself.
945 if (!CS.getInstruction()->getType()->isVoidTy())
946 ReplaceInstUsesWith(*CS.getInstruction(),
947 UndefValue::get(CS.getInstruction()->getType()));
949 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
950 // Don't break the CFG, insert a dummy cond branch.
951 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
952 ConstantInt::getTrue(Callee->getContext()), II);
954 return EraseInstFromFunction(*CS.getInstruction());
957 if (IntrinsicInst *II = FindInitTrampoline(Callee))
958 return transformCallThroughTrampoline(CS, II);
960 PointerType *PTy = cast<PointerType>(Callee->getType());
961 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
962 if (FTy->isVarArg()) {
963 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 2 : 0);
964 // See if we can optimize any arguments passed through the varargs area of
966 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
967 E = CS.arg_end(); I != E; ++I, ++ix) {
968 CastInst *CI = dyn_cast<CastInst>(*I);
969 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
970 *I = CI->getOperand(0);
976 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
977 // Inline asm calls cannot throw - mark them 'nounwind'.
978 CS.setDoesNotThrow();
982 // Try to optimize the call if possible, we require TargetData for most of
983 // this. None of these calls are seen as possibly dead so go ahead and
984 // delete the instruction now.
985 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
986 Instruction *I = tryOptimizeCall(CI, TD);
987 // If we changed something return the result, etc. Otherwise let
988 // the fallthrough check.
989 if (I) return EraseInstFromFunction(*I);
992 return Changed ? CS.getInstruction() : 0;
995 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
996 // attempt to move the cast to the arguments of the call/invoke.
998 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1000 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1003 Instruction *Caller = CS.getInstruction();
1004 const AttrListPtr &CallerPAL = CS.getAttributes();
1006 // Okay, this is a cast from a function to a different type. Unless doing so
1007 // would cause a type conversion of one of our arguments, change this call to
1008 // be a direct call with arguments casted to the appropriate types.
1010 FunctionType *FT = Callee->getFunctionType();
1011 Type *OldRetTy = Caller->getType();
1012 Type *NewRetTy = FT->getReturnType();
1014 if (NewRetTy->isStructTy())
1015 return false; // TODO: Handle multiple return values.
1017 // Check to see if we are changing the return type...
1018 if (OldRetTy != NewRetTy) {
1019 if (Callee->isDeclaration() &&
1020 // Conversion is ok if changing from one pointer type to another or from
1021 // a pointer to an integer of the same size.
1022 !((OldRetTy->isPointerTy() || !TD ||
1023 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
1024 (NewRetTy->isPointerTy() || !TD ||
1025 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
1026 return false; // Cannot transform this return value.
1028 if (!Caller->use_empty() &&
1029 // void -> non-void is handled specially
1030 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
1031 return false; // Cannot transform this return value.
1033 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1034 Attributes RAttrs = CallerPAL.getRetAttributes();
1035 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
1036 return false; // Attribute not compatible with transformed value.
1039 // If the callsite is an invoke instruction, and the return value is used by
1040 // a PHI node in a successor, we cannot change the return type of the call
1041 // because there is no place to put the cast instruction (without breaking
1042 // the critical edge). Bail out in this case.
1043 if (!Caller->use_empty())
1044 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1045 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1047 if (PHINode *PN = dyn_cast<PHINode>(*UI))
1048 if (PN->getParent() == II->getNormalDest() ||
1049 PN->getParent() == II->getUnwindDest())
1053 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1054 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1056 CallSite::arg_iterator AI = CS.arg_begin();
1057 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1058 Type *ParamTy = FT->getParamType(i);
1059 Type *ActTy = (*AI)->getType();
1061 if (!CastInst::isCastable(ActTy, ParamTy))
1062 return false; // Cannot transform this parameter value.
1064 unsigned Attrs = CallerPAL.getParamAttributes(i + 1);
1065 if (Attrs & Attribute::typeIncompatible(ParamTy))
1066 return false; // Attribute not compatible with transformed value.
1068 // If the parameter is passed as a byval argument, then we have to have a
1069 // sized type and the sized type has to have the same size as the old type.
1070 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
1071 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1072 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1075 Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1076 if (TD->getTypeAllocSize(CurElTy) !=
1077 TD->getTypeAllocSize(ParamPTy->getElementType()))
1081 // Converting from one pointer type to another or between a pointer and an
1082 // integer of the same size is safe even if we do not have a body.
1083 bool isConvertible = ActTy == ParamTy ||
1084 (TD && ((ParamTy->isPointerTy() ||
1085 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1086 (ActTy->isPointerTy() ||
1087 ActTy == TD->getIntPtrType(Caller->getContext()))));
1088 if (Callee->isDeclaration() && !isConvertible) return false;
1091 if (Callee->isDeclaration()) {
1092 // Do not delete arguments unless we have a function body.
1093 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1096 // If the callee is just a declaration, don't change the varargsness of the
1097 // call. We don't want to introduce a varargs call where one doesn't
1099 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1100 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1104 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1105 !CallerPAL.isEmpty())
1106 // In this case we have more arguments than the new function type, but we
1107 // won't be dropping them. Check that these extra arguments have attributes
1108 // that are compatible with being a vararg call argument.
1109 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1110 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1112 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1113 if (PAttrs & Attribute::VarArgsIncompatible)
1118 // Okay, we decided that this is a safe thing to do: go ahead and start
1119 // inserting cast instructions as necessary.
1120 std::vector<Value*> Args;
1121 Args.reserve(NumActualArgs);
1122 SmallVector<AttributeWithIndex, 8> attrVec;
1123 attrVec.reserve(NumCommonArgs);
1125 // Get any return attributes.
1126 Attributes RAttrs = CallerPAL.getRetAttributes();
1128 // If the return value is not being used, the type may not be compatible
1129 // with the existing attributes. Wipe out any problematic attributes.
1130 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1132 // Add the new return attributes.
1134 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1136 AI = CS.arg_begin();
1137 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1138 Type *ParamTy = FT->getParamType(i);
1139 if ((*AI)->getType() == ParamTy) {
1140 Args.push_back(*AI);
1142 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1143 false, ParamTy, false);
1144 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy));
1147 // Add any parameter attributes.
1148 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1149 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1152 // If the function takes more arguments than the call was taking, add them
1154 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1155 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1157 // If we are removing arguments to the function, emit an obnoxious warning.
1158 if (FT->getNumParams() < NumActualArgs) {
1159 if (!FT->isVarArg()) {
1160 errs() << "WARNING: While resolving call to function '"
1161 << Callee->getName() << "' arguments were dropped!\n";
1163 // Add all of the arguments in their promoted form to the arg list.
1164 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1165 Type *PTy = getPromotedType((*AI)->getType());
1166 if (PTy != (*AI)->getType()) {
1167 // Must promote to pass through va_arg area!
1168 Instruction::CastOps opcode =
1169 CastInst::getCastOpcode(*AI, false, PTy, false);
1170 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1172 Args.push_back(*AI);
1175 // Add any parameter attributes.
1176 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1177 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1182 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1183 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1185 if (NewRetTy->isVoidTy())
1186 Caller->setName(""); // Void type should not have a name.
1188 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1192 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1193 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1194 II->getUnwindDest(), Args);
1196 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1197 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1199 CallInst *CI = cast<CallInst>(Caller);
1200 NC = Builder->CreateCall(Callee, Args);
1202 if (CI->isTailCall())
1203 cast<CallInst>(NC)->setTailCall();
1204 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1205 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1208 // Insert a cast of the return type as necessary.
1210 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1211 if (!NV->getType()->isVoidTy()) {
1212 Instruction::CastOps opcode =
1213 CastInst::getCastOpcode(NC, false, OldRetTy, false);
1214 NV = NC = CastInst::Create(opcode, NC, OldRetTy);
1215 NC->setDebugLoc(Caller->getDebugLoc());
1217 // If this is an invoke instruction, we should insert it after the first
1218 // non-phi, instruction in the normal successor block.
1219 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1220 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1221 InsertNewInstBefore(NC, *I);
1223 // Otherwise, it's a call, just insert cast right after the call.
1224 InsertNewInstBefore(NC, *Caller);
1226 Worklist.AddUsersToWorkList(*Caller);
1228 NV = UndefValue::get(Caller->getType());
1232 if (!Caller->use_empty())
1233 ReplaceInstUsesWith(*Caller, NV);
1235 EraseInstFromFunction(*Caller);
1239 // transformCallThroughTrampoline - Turn a call to a function created by
1240 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1241 // underlying function.
1244 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1245 IntrinsicInst *Tramp) {
1246 Value *Callee = CS.getCalledValue();
1247 PointerType *PTy = cast<PointerType>(Callee->getType());
1248 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1249 const AttrListPtr &Attrs = CS.getAttributes();
1251 // If the call already has the 'nest' attribute somewhere then give up -
1252 // otherwise 'nest' would occur twice after splicing in the chain.
1253 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1257 "transformCallThroughTrampoline called with incorrect CallSite.");
1259 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1260 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1261 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1263 const AttrListPtr &NestAttrs = NestF->getAttributes();
1264 if (!NestAttrs.isEmpty()) {
1265 unsigned NestIdx = 1;
1267 Attributes NestAttr = Attribute::None;
1269 // Look for a parameter marked with the 'nest' attribute.
1270 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1271 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1272 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1273 // Record the parameter type and any other attributes.
1275 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1280 Instruction *Caller = CS.getInstruction();
1281 std::vector<Value*> NewArgs;
1282 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1284 SmallVector<AttributeWithIndex, 8> NewAttrs;
1285 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1287 // Insert the nest argument into the call argument list, which may
1288 // mean appending it. Likewise for attributes.
1290 // Add any result attributes.
1291 if (Attributes Attr = Attrs.getRetAttributes())
1292 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1296 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1298 if (Idx == NestIdx) {
1299 // Add the chain argument and attributes.
1300 Value *NestVal = Tramp->getArgOperand(2);
1301 if (NestVal->getType() != NestTy)
1302 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1303 NewArgs.push_back(NestVal);
1304 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1310 // Add the original argument and attributes.
1311 NewArgs.push_back(*I);
1312 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1314 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1320 // Add any function attributes.
1321 if (Attributes Attr = Attrs.getFnAttributes())
1322 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1324 // The trampoline may have been bitcast to a bogus type (FTy).
1325 // Handle this by synthesizing a new function type, equal to FTy
1326 // with the chain parameter inserted.
1328 std::vector<Type*> NewTypes;
1329 NewTypes.reserve(FTy->getNumParams()+1);
1331 // Insert the chain's type into the list of parameter types, which may
1332 // mean appending it.
1335 FunctionType::param_iterator I = FTy->param_begin(),
1336 E = FTy->param_end();
1340 // Add the chain's type.
1341 NewTypes.push_back(NestTy);
1346 // Add the original type.
1347 NewTypes.push_back(*I);
1353 // Replace the trampoline call with a direct call. Let the generic
1354 // code sort out any function type mismatches.
1355 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1357 Constant *NewCallee =
1358 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1359 NestF : ConstantExpr::getBitCast(NestF,
1360 PointerType::getUnqual(NewFTy));
1361 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1364 Instruction *NewCaller;
1365 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1366 NewCaller = InvokeInst::Create(NewCallee,
1367 II->getNormalDest(), II->getUnwindDest(),
1369 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1370 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1372 NewCaller = CallInst::Create(NewCallee, NewArgs);
1373 if (cast<CallInst>(Caller)->isTailCall())
1374 cast<CallInst>(NewCaller)->setTailCall();
1375 cast<CallInst>(NewCaller)->
1376 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1377 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1384 // Replace the trampoline call with a direct call. Since there is no 'nest'
1385 // parameter, there is no need to adjust the argument list. Let the generic
1386 // code sort out any function type mismatches.
1387 Constant *NewCallee =
1388 NestF->getType() == PTy ? NestF :
1389 ConstantExpr::getBitCast(NestF, PTy);
1390 CS.setCalledFunction(NewCallee);
1391 return CS.getInstruction();