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/DataLayout.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());
32 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
33 /// single scalar element, like {{{type}}} or [1 x type], return type.
34 static Type *reduceToSingleValueType(Type *T) {
35 while (!T->isSingleValueType()) {
36 if (StructType *STy = dyn_cast<StructType>(T)) {
37 if (STy->getNumElements() == 1)
38 T = STy->getElementType(0);
41 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
42 if (ATy->getNumElements() == 1)
43 T = ATy->getElementType();
53 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
54 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD);
55 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
56 unsigned MinAlign = std::min(DstAlign, SrcAlign);
57 unsigned CopyAlign = MI->getAlignment();
59 if (CopyAlign < MinAlign) {
60 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
65 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
67 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
68 if (MemOpLength == 0) return 0;
70 // Source and destination pointer types are always "i8*" for intrinsic. See
71 // if the size is something we can handle with a single primitive load/store.
72 // A single load+store correctly handles overlapping memory in the memmove
74 uint64_t Size = MemOpLength->getLimitedValue();
75 assert(Size && "0-sized memory transfering should be removed already.");
77 if (Size > 8 || (Size&(Size-1)))
78 return 0; // If not 1/2/4/8 bytes, exit.
80 // Use an integer load+store unless we can find something better.
82 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
84 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
86 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
87 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
88 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
90 // Memcpy forces the use of i8* for the source and destination. That means
91 // that if you're using memcpy to move one double around, you'll get a cast
92 // from double* to i8*. We'd much rather use a double load+store rather than
93 // an i64 load+store, here because this improves the odds that the source or
94 // dest address will be promotable. See if we can find a better type than the
96 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
98 if (StrippedDest != MI->getArgOperand(0)) {
99 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
101 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
102 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
103 // down through these levels if so.
104 SrcETy = reduceToSingleValueType(SrcETy);
106 if (SrcETy->isSingleValueType()) {
107 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
108 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
110 // If the memcpy has metadata describing the members, see if we can
111 // get the TBAA tag describing our copy.
112 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
113 if (M->getNumOperands() == 3 &&
115 isa<ConstantInt>(M->getOperand(0)) &&
116 cast<ConstantInt>(M->getOperand(0))->isNullValue() &&
118 isa<ConstantInt>(M->getOperand(1)) &&
119 cast<ConstantInt>(M->getOperand(1))->getValue() == Size &&
121 isa<MDNode>(M->getOperand(2)))
122 CopyMD = cast<MDNode>(M->getOperand(2));
128 // If the memcpy/memmove provides better alignment info than we can
130 SrcAlign = std::max(SrcAlign, CopyAlign);
131 DstAlign = std::max(DstAlign, CopyAlign);
133 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
134 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
135 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
136 L->setAlignment(SrcAlign);
138 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
139 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
140 S->setAlignment(DstAlign);
142 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
144 // Set the size of the copy to 0, it will be deleted on the next iteration.
145 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
149 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
150 unsigned Alignment = getKnownAlignment(MI->getDest(), TD);
151 if (MI->getAlignment() < Alignment) {
152 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
157 // Extract the length and alignment and fill if they are constant.
158 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
159 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
160 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
162 uint64_t Len = LenC->getLimitedValue();
163 Alignment = MI->getAlignment();
164 assert(Len && "0-sized memory setting should be removed already.");
166 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
167 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
168 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
170 Value *Dest = MI->getDest();
171 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
172 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
173 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
175 // Alignment 0 is identity for alignment 1 for memset, but not store.
176 if (Alignment == 0) Alignment = 1;
178 // Extract the fill value and store.
179 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
180 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
182 S->setAlignment(Alignment);
184 // Set the size of the copy to 0, it will be deleted on the next iteration.
185 MI->setLength(Constant::getNullValue(LenC->getType()));
192 /// visitCallInst - CallInst simplification. This mostly only handles folding
193 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
194 /// the heavy lifting.
196 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
197 if (isFreeCall(&CI, TLI))
198 return visitFree(CI);
200 // If the caller function is nounwind, mark the call as nounwind, even if the
202 if (CI.getParent()->getParent()->doesNotThrow() &&
203 !CI.doesNotThrow()) {
204 CI.setDoesNotThrow();
208 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
209 if (!II) return visitCallSite(&CI);
211 // Intrinsics cannot occur in an invoke, so handle them here instead of in
213 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
214 bool Changed = false;
216 // memmove/cpy/set of zero bytes is a noop.
217 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
218 if (NumBytes->isNullValue())
219 return EraseInstFromFunction(CI);
221 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
222 if (CI->getZExtValue() == 1) {
223 // Replace the instruction with just byte operations. We would
224 // transform other cases to loads/stores, but we don't know if
225 // alignment is sufficient.
229 // No other transformations apply to volatile transfers.
230 if (MI->isVolatile())
233 // If we have a memmove and the source operation is a constant global,
234 // then the source and dest pointers can't alias, so we can change this
235 // into a call to memcpy.
236 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
237 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
238 if (GVSrc->isConstant()) {
239 Module *M = CI.getParent()->getParent()->getParent();
240 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
241 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
242 CI.getArgOperand(1)->getType(),
243 CI.getArgOperand(2)->getType() };
244 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
249 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
250 // memmove(x,x,size) -> noop.
251 if (MTI->getSource() == MTI->getDest())
252 return EraseInstFromFunction(CI);
255 // If we can determine a pointer alignment that is bigger than currently
256 // set, update the alignment.
257 if (isa<MemTransferInst>(MI)) {
258 if (Instruction *I = SimplifyMemTransfer(MI))
260 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
261 if (Instruction *I = SimplifyMemSet(MSI))
265 if (Changed) return II;
268 switch (II->getIntrinsicID()) {
270 case Intrinsic::objectsize: {
272 if (getObjectSize(II->getArgOperand(0), Size, TD, TLI))
273 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
276 case Intrinsic::bswap:
277 // bswap(bswap(x)) -> x
278 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
279 if (Operand->getIntrinsicID() == Intrinsic::bswap)
280 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
282 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
283 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
284 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
285 if (Operand->getIntrinsicID() == Intrinsic::bswap) {
286 unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
287 TI->getType()->getPrimitiveSizeInBits();
288 Value *CV = ConstantInt::get(Operand->getType(), C);
289 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
290 return new TruncInst(V, TI->getType());
295 case Intrinsic::powi:
296 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
299 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
302 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
303 // powi(x, -1) -> 1/x
304 if (Power->isAllOnesValue())
305 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
306 II->getArgOperand(0));
309 case Intrinsic::cttz: {
310 // If all bits below the first known one are known zero,
311 // this value is constant.
312 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
313 // FIXME: Try to simplify vectors of integers.
315 uint32_t BitWidth = IT->getBitWidth();
316 APInt KnownZero(BitWidth, 0);
317 APInt KnownOne(BitWidth, 0);
318 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
319 unsigned TrailingZeros = KnownOne.countTrailingZeros();
320 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
321 if ((Mask & KnownZero) == Mask)
322 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
323 APInt(BitWidth, TrailingZeros)));
327 case Intrinsic::ctlz: {
328 // If all bits above the first known one are known zero,
329 // this value is constant.
330 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
331 // FIXME: Try to simplify vectors of integers.
333 uint32_t BitWidth = IT->getBitWidth();
334 APInt KnownZero(BitWidth, 0);
335 APInt KnownOne(BitWidth, 0);
336 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
337 unsigned LeadingZeros = KnownOne.countLeadingZeros();
338 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
339 if ((Mask & KnownZero) == Mask)
340 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
341 APInt(BitWidth, LeadingZeros)));
345 case Intrinsic::uadd_with_overflow: {
346 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
347 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
348 uint32_t BitWidth = IT->getBitWidth();
349 APInt LHSKnownZero(BitWidth, 0);
350 APInt LHSKnownOne(BitWidth, 0);
351 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
352 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
353 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
355 if (LHSKnownNegative || LHSKnownPositive) {
356 APInt RHSKnownZero(BitWidth, 0);
357 APInt RHSKnownOne(BitWidth, 0);
358 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
359 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
360 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
361 if (LHSKnownNegative && RHSKnownNegative) {
362 // The sign bit is set in both cases: this MUST overflow.
363 // Create a simple add instruction, and insert it into the struct.
364 Value *Add = Builder->CreateAdd(LHS, RHS);
367 UndefValue::get(LHS->getType()),
368 ConstantInt::getTrue(II->getContext())
370 StructType *ST = cast<StructType>(II->getType());
371 Constant *Struct = ConstantStruct::get(ST, V);
372 return InsertValueInst::Create(Struct, Add, 0);
375 if (LHSKnownPositive && RHSKnownPositive) {
376 // The sign bit is clear in both cases: this CANNOT overflow.
377 // Create a simple add instruction, and insert it into the struct.
378 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
381 UndefValue::get(LHS->getType()),
382 ConstantInt::getFalse(II->getContext())
384 StructType *ST = cast<StructType>(II->getType());
385 Constant *Struct = ConstantStruct::get(ST, V);
386 return InsertValueInst::Create(Struct, Add, 0);
390 // FALL THROUGH uadd into sadd
391 case Intrinsic::sadd_with_overflow:
392 // Canonicalize constants into the RHS.
393 if (isa<Constant>(II->getArgOperand(0)) &&
394 !isa<Constant>(II->getArgOperand(1))) {
395 Value *LHS = II->getArgOperand(0);
396 II->setArgOperand(0, II->getArgOperand(1));
397 II->setArgOperand(1, LHS);
401 // X + undef -> undef
402 if (isa<UndefValue>(II->getArgOperand(1)))
403 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
405 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
406 // X + 0 -> {X, false}
409 UndefValue::get(II->getArgOperand(0)->getType()),
410 ConstantInt::getFalse(II->getContext())
413 ConstantStruct::get(cast<StructType>(II->getType()), V);
414 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
418 case Intrinsic::usub_with_overflow:
419 case Intrinsic::ssub_with_overflow:
420 // undef - X -> undef
421 // X - undef -> undef
422 if (isa<UndefValue>(II->getArgOperand(0)) ||
423 isa<UndefValue>(II->getArgOperand(1)))
424 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
426 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
427 // X - 0 -> {X, false}
430 UndefValue::get(II->getArgOperand(0)->getType()),
431 ConstantInt::getFalse(II->getContext())
434 ConstantStruct::get(cast<StructType>(II->getType()), V);
435 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
439 case Intrinsic::umul_with_overflow: {
440 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
441 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
443 APInt LHSKnownZero(BitWidth, 0);
444 APInt LHSKnownOne(BitWidth, 0);
445 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
446 APInt RHSKnownZero(BitWidth, 0);
447 APInt RHSKnownOne(BitWidth, 0);
448 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
450 // Get the largest possible values for each operand.
451 APInt LHSMax = ~LHSKnownZero;
452 APInt RHSMax = ~RHSKnownZero;
454 // If multiplying the maximum values does not overflow then we can turn
455 // this into a plain NUW mul.
457 LHSMax.umul_ov(RHSMax, Overflow);
459 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
461 UndefValue::get(LHS->getType()),
464 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
465 return InsertValueInst::Create(Struct, Mul, 0);
468 case Intrinsic::smul_with_overflow:
469 // Canonicalize constants into the RHS.
470 if (isa<Constant>(II->getArgOperand(0)) &&
471 !isa<Constant>(II->getArgOperand(1))) {
472 Value *LHS = II->getArgOperand(0);
473 II->setArgOperand(0, II->getArgOperand(1));
474 II->setArgOperand(1, LHS);
478 // X * undef -> undef
479 if (isa<UndefValue>(II->getArgOperand(1)))
480 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
482 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
485 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
487 // X * 1 -> {X, false}
488 if (RHSI->equalsInt(1)) {
490 UndefValue::get(II->getArgOperand(0)->getType()),
491 ConstantInt::getFalse(II->getContext())
494 ConstantStruct::get(cast<StructType>(II->getType()), V);
495 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
499 case Intrinsic::ppc_altivec_lvx:
500 case Intrinsic::ppc_altivec_lvxl:
501 // Turn PPC lvx -> load if the pointer is known aligned.
502 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
503 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
504 PointerType::getUnqual(II->getType()));
505 return new LoadInst(Ptr);
508 case Intrinsic::ppc_altivec_stvx:
509 case Intrinsic::ppc_altivec_stvxl:
510 // Turn stvx -> store if the pointer is known aligned.
511 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
513 PointerType::getUnqual(II->getArgOperand(0)->getType());
514 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
515 return new StoreInst(II->getArgOperand(0), Ptr);
518 case Intrinsic::x86_sse_storeu_ps:
519 case Intrinsic::x86_sse2_storeu_pd:
520 case Intrinsic::x86_sse2_storeu_dq:
521 // Turn X86 storeu -> store if the pointer is known aligned.
522 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
524 PointerType::getUnqual(II->getArgOperand(1)->getType());
525 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
526 return new StoreInst(II->getArgOperand(1), Ptr);
530 case Intrinsic::x86_sse_cvtss2si:
531 case Intrinsic::x86_sse_cvtss2si64:
532 case Intrinsic::x86_sse_cvttss2si:
533 case Intrinsic::x86_sse_cvttss2si64:
534 case Intrinsic::x86_sse2_cvtsd2si:
535 case Intrinsic::x86_sse2_cvtsd2si64:
536 case Intrinsic::x86_sse2_cvttsd2si:
537 case Intrinsic::x86_sse2_cvttsd2si64: {
538 // These intrinsics only demand the 0th element of their input vectors. If
539 // we can simplify the input based on that, do so now.
541 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
542 APInt DemandedElts(VWidth, 1);
543 APInt UndefElts(VWidth, 0);
544 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
545 DemandedElts, UndefElts)) {
546 II->setArgOperand(0, V);
553 case Intrinsic::x86_sse41_pmovsxbw:
554 case Intrinsic::x86_sse41_pmovsxwd:
555 case Intrinsic::x86_sse41_pmovsxdq:
556 case Intrinsic::x86_sse41_pmovzxbw:
557 case Intrinsic::x86_sse41_pmovzxwd:
558 case Intrinsic::x86_sse41_pmovzxdq: {
559 // pmov{s|z}x ignores the upper half of their input vectors.
561 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
562 unsigned LowHalfElts = VWidth / 2;
563 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
564 APInt UndefElts(VWidth, 0);
565 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
568 II->setArgOperand(0, TmpV);
574 case Intrinsic::ppc_altivec_vperm:
575 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
576 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
577 assert(Mask->getType()->getVectorNumElements() == 16 &&
578 "Bad type for intrinsic!");
580 // Check that all of the elements are integer constants or undefs.
581 bool AllEltsOk = true;
582 for (unsigned i = 0; i != 16; ++i) {
583 Constant *Elt = Mask->getAggregateElement(i);
585 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
592 // Cast the input vectors to byte vectors.
593 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
595 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
597 Value *Result = UndefValue::get(Op0->getType());
599 // Only extract each element once.
600 Value *ExtractedElts[32];
601 memset(ExtractedElts, 0, sizeof(ExtractedElts));
603 for (unsigned i = 0; i != 16; ++i) {
604 if (isa<UndefValue>(Mask->getAggregateElement(i)))
607 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
608 Idx &= 31; // Match the hardware behavior.
610 if (ExtractedElts[Idx] == 0) {
612 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
613 Builder->getInt32(Idx&15));
616 // Insert this value into the result vector.
617 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
618 Builder->getInt32(i));
620 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
625 case Intrinsic::arm_neon_vld1:
626 case Intrinsic::arm_neon_vld2:
627 case Intrinsic::arm_neon_vld3:
628 case Intrinsic::arm_neon_vld4:
629 case Intrinsic::arm_neon_vld2lane:
630 case Intrinsic::arm_neon_vld3lane:
631 case Intrinsic::arm_neon_vld4lane:
632 case Intrinsic::arm_neon_vst1:
633 case Intrinsic::arm_neon_vst2:
634 case Intrinsic::arm_neon_vst3:
635 case Intrinsic::arm_neon_vst4:
636 case Intrinsic::arm_neon_vst2lane:
637 case Intrinsic::arm_neon_vst3lane:
638 case Intrinsic::arm_neon_vst4lane: {
639 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
640 unsigned AlignArg = II->getNumArgOperands() - 1;
641 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
642 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
643 II->setArgOperand(AlignArg,
644 ConstantInt::get(Type::getInt32Ty(II->getContext()),
651 case Intrinsic::arm_neon_vmulls:
652 case Intrinsic::arm_neon_vmullu: {
653 Value *Arg0 = II->getArgOperand(0);
654 Value *Arg1 = II->getArgOperand(1);
656 // Handle mul by zero first:
657 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
658 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
661 // Check for constant LHS & RHS - in this case we just simplify.
662 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu);
663 VectorType *NewVT = cast<VectorType>(II->getType());
664 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth();
665 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) {
666 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
667 VectorType* VT = cast<VectorType>(CV0->getType());
668 SmallVector<Constant*, 4> NewElems;
669 for (unsigned i = 0; i < VT->getNumElements(); ++i) {
671 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue();
672 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth);
674 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue();
675 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth);
677 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E));
679 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems));
682 // Couldn't simplify - cannonicalize constant to the RHS.
683 std::swap(Arg0, Arg1);
686 // Handle mul by one:
687 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
688 if (ConstantInt *Splat =
689 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) {
690 if (Splat->isOne()) {
692 return CastInst::CreateZExtOrBitCast(Arg0, II->getType());
694 return CastInst::CreateSExtOrBitCast(Arg0, II->getType());
702 case Intrinsic::stackrestore: {
703 // If the save is right next to the restore, remove the restore. This can
704 // happen when variable allocas are DCE'd.
705 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
706 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
707 BasicBlock::iterator BI = SS;
709 return EraseInstFromFunction(CI);
713 // Scan down this block to see if there is another stack restore in the
714 // same block without an intervening call/alloca.
715 BasicBlock::iterator BI = II;
716 TerminatorInst *TI = II->getParent()->getTerminator();
717 bool CannotRemove = false;
718 for (++BI; &*BI != TI; ++BI) {
719 if (isa<AllocaInst>(BI)) {
723 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
724 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
725 // If there is a stackrestore below this one, remove this one.
726 if (II->getIntrinsicID() == Intrinsic::stackrestore)
727 return EraseInstFromFunction(CI);
728 // Otherwise, ignore the intrinsic.
730 // If we found a non-intrinsic call, we can't remove the stack
738 // If the stack restore is in a return, resume, or unwind block and if there
739 // are no allocas or calls between the restore and the return, nuke the
741 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
742 return EraseInstFromFunction(CI);
747 return visitCallSite(II);
750 // InvokeInst simplification
752 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
753 return visitCallSite(&II);
756 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
757 /// passed through the varargs area, we can eliminate the use of the cast.
758 static bool isSafeToEliminateVarargsCast(const CallSite CS,
759 const CastInst * const CI,
760 const DataLayout * const TD,
762 if (!CI->isLosslessCast())
765 // The size of ByVal arguments is derived from the type, so we
766 // can't change to a type with a different size. If the size were
767 // passed explicitly we could avoid this check.
768 if (!CS.isByValArgument(ix))
772 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
773 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
774 if (!SrcTy->isSized() || !DstTy->isSized())
776 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
781 // Try to fold some different type of calls here.
782 // Currently we're only working with the checking functions, memcpy_chk,
783 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
784 // strcat_chk and strncat_chk.
785 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *TD) {
786 if (CI->getCalledFunction() == 0) return 0;
788 if (Value *With = Simplifier->optimizeCall(CI))
789 return ReplaceInstUsesWith(*CI, With);
794 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
795 // Strip off at most one level of pointer casts, looking for an alloca. This
796 // is good enough in practice and simpler than handling any number of casts.
797 Value *Underlying = TrampMem->stripPointerCasts();
798 if (Underlying != TrampMem &&
799 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
801 if (!isa<AllocaInst>(Underlying))
804 IntrinsicInst *InitTrampoline = 0;
805 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
807 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
810 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
812 // More than one init_trampoline writes to this value. Give up.
817 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
818 // Allow any number of calls to adjust.trampoline.
823 // No call to init.trampoline found.
827 // Check that the alloca is being used in the expected way.
828 if (InitTrampoline->getOperand(0) != TrampMem)
831 return InitTrampoline;
834 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
836 // Visit all the previous instructions in the basic block, and try to find a
837 // init.trampoline which has a direct path to the adjust.trampoline.
838 for (BasicBlock::iterator I = AdjustTramp,
839 E = AdjustTramp->getParent()->begin(); I != E; ) {
840 Instruction *Inst = --I;
841 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
842 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
843 II->getOperand(0) == TrampMem)
845 if (Inst->mayWriteToMemory())
851 // Given a call to llvm.adjust.trampoline, find and return the corresponding
852 // call to llvm.init.trampoline if the call to the trampoline can be optimized
853 // to a direct call to a function. Otherwise return NULL.
855 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
856 Callee = Callee->stripPointerCasts();
857 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
859 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
862 Value *TrampMem = AdjustTramp->getOperand(0);
864 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
866 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
871 // visitCallSite - Improvements for call and invoke instructions.
873 Instruction *InstCombiner::visitCallSite(CallSite CS) {
874 if (isAllocLikeFn(CS.getInstruction(), TLI))
875 return visitAllocSite(*CS.getInstruction());
877 bool Changed = false;
879 // If the callee is a pointer to a function, attempt to move any casts to the
880 // arguments of the call/invoke.
881 Value *Callee = CS.getCalledValue();
882 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
885 if (Function *CalleeF = dyn_cast<Function>(Callee))
886 // If the call and callee calling conventions don't match, this call must
887 // be unreachable, as the call is undefined.
888 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
889 // Only do this for calls to a function with a body. A prototype may
890 // not actually end up matching the implementation's calling conv for a
891 // variety of reasons (e.g. it may be written in assembly).
892 !CalleeF->isDeclaration()) {
893 Instruction *OldCall = CS.getInstruction();
894 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
895 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
897 // If OldCall dues not return void then replaceAllUsesWith undef.
898 // This allows ValueHandlers and custom metadata to adjust itself.
899 if (!OldCall->getType()->isVoidTy())
900 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
901 if (isa<CallInst>(OldCall))
902 return EraseInstFromFunction(*OldCall);
904 // We cannot remove an invoke, because it would change the CFG, just
905 // change the callee to a null pointer.
906 cast<InvokeInst>(OldCall)->setCalledFunction(
907 Constant::getNullValue(CalleeF->getType()));
911 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
912 // If CS does not return void then replaceAllUsesWith undef.
913 // This allows ValueHandlers and custom metadata to adjust itself.
914 if (!CS.getInstruction()->getType()->isVoidTy())
915 ReplaceInstUsesWith(*CS.getInstruction(),
916 UndefValue::get(CS.getInstruction()->getType()));
918 if (isa<InvokeInst>(CS.getInstruction())) {
919 // Can't remove an invoke because we cannot change the CFG.
923 // This instruction is not reachable, just remove it. We insert a store to
924 // undef so that we know that this code is not reachable, despite the fact
925 // that we can't modify the CFG here.
926 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
927 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
928 CS.getInstruction());
930 return EraseInstFromFunction(*CS.getInstruction());
933 if (IntrinsicInst *II = FindInitTrampoline(Callee))
934 return transformCallThroughTrampoline(CS, II);
936 PointerType *PTy = cast<PointerType>(Callee->getType());
937 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
938 if (FTy->isVarArg()) {
939 int ix = FTy->getNumParams();
940 // See if we can optimize any arguments passed through the varargs area of
942 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
943 E = CS.arg_end(); I != E; ++I, ++ix) {
944 CastInst *CI = dyn_cast<CastInst>(*I);
945 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
946 *I = CI->getOperand(0);
952 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
953 // Inline asm calls cannot throw - mark them 'nounwind'.
954 CS.setDoesNotThrow();
958 // Try to optimize the call if possible, we require DataLayout for most of
959 // this. None of these calls are seen as possibly dead so go ahead and
960 // delete the instruction now.
961 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
962 Instruction *I = tryOptimizeCall(CI, TD);
963 // If we changed something return the result, etc. Otherwise let
964 // the fallthrough check.
965 if (I) return EraseInstFromFunction(*I);
968 return Changed ? CS.getInstruction() : 0;
971 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
972 // attempt to move the cast to the arguments of the call/invoke.
974 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
976 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
979 Instruction *Caller = CS.getInstruction();
980 const AttrListPtr &CallerPAL = CS.getAttributes();
982 // Okay, this is a cast from a function to a different type. Unless doing so
983 // would cause a type conversion of one of our arguments, change this call to
984 // be a direct call with arguments casted to the appropriate types.
986 FunctionType *FT = Callee->getFunctionType();
987 Type *OldRetTy = Caller->getType();
988 Type *NewRetTy = FT->getReturnType();
990 if (NewRetTy->isStructTy())
991 return false; // TODO: Handle multiple return values.
993 // Check to see if we are changing the return type...
994 if (OldRetTy != NewRetTy) {
995 if (Callee->isDeclaration() &&
996 // Conversion is ok if changing from one pointer type to another or from
997 // a pointer to an integer of the same size.
998 !((OldRetTy->isPointerTy() || !TD ||
999 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
1000 (NewRetTy->isPointerTy() || !TD ||
1001 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
1002 return false; // Cannot transform this return value.
1004 if (!Caller->use_empty() &&
1005 // void -> non-void is handled specially
1006 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
1007 return false; // Cannot transform this return value.
1009 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1010 Attributes::Builder RAttrs = CallerPAL.getRetAttributes();
1011 if (RAttrs.hasAttributes(Attributes::typeIncompatible(NewRetTy)))
1012 return false; // Attribute not compatible with transformed value.
1015 // If the callsite is an invoke instruction, and the return value is used by
1016 // a PHI node in a successor, we cannot change the return type of the call
1017 // because there is no place to put the cast instruction (without breaking
1018 // the critical edge). Bail out in this case.
1019 if (!Caller->use_empty())
1020 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1021 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1023 if (PHINode *PN = dyn_cast<PHINode>(*UI))
1024 if (PN->getParent() == II->getNormalDest() ||
1025 PN->getParent() == II->getUnwindDest())
1029 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1030 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1032 CallSite::arg_iterator AI = CS.arg_begin();
1033 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1034 Type *ParamTy = FT->getParamType(i);
1035 Type *ActTy = (*AI)->getType();
1037 if (!CastInst::isCastable(ActTy, ParamTy))
1038 return false; // Cannot transform this parameter value.
1040 Attributes Attrs = CallerPAL.getParamAttributes(i + 1);
1041 if (Attributes::Builder(Attrs).
1042 hasAttributes(Attributes::typeIncompatible(ParamTy)))
1043 return false; // Attribute not compatible with transformed value.
1045 // If the parameter is passed as a byval argument, then we have to have a
1046 // sized type and the sized type has to have the same size as the old type.
1047 if (ParamTy != ActTy && Attrs.hasAttribute(Attributes::ByVal)) {
1048 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1049 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1052 Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1053 if (TD->getTypeAllocSize(CurElTy) !=
1054 TD->getTypeAllocSize(ParamPTy->getElementType()))
1058 // Converting from one pointer type to another or between a pointer and an
1059 // integer of the same size is safe even if we do not have a body.
1060 bool isConvertible = ActTy == ParamTy ||
1061 (TD && ((ParamTy->isPointerTy() ||
1062 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1063 (ActTy->isPointerTy() ||
1064 ActTy == TD->getIntPtrType(Caller->getContext()))));
1065 if (Callee->isDeclaration() && !isConvertible) return false;
1068 if (Callee->isDeclaration()) {
1069 // Do not delete arguments unless we have a function body.
1070 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1073 // If the callee is just a declaration, don't change the varargsness of the
1074 // call. We don't want to introduce a varargs call where one doesn't
1076 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1077 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1080 // If both the callee and the cast type are varargs, we still have to make
1081 // sure the number of fixed parameters are the same or we have the same
1082 // ABI issues as if we introduce a varargs call.
1083 if (FT->isVarArg() &&
1084 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1085 FT->getNumParams() !=
1086 cast<FunctionType>(APTy->getElementType())->getNumParams())
1090 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1091 !CallerPAL.isEmpty())
1092 // In this case we have more arguments than the new function type, but we
1093 // won't be dropping them. Check that these extra arguments have attributes
1094 // that are compatible with being a vararg call argument.
1095 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1096 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1098 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1099 if (PAttrs.hasIncompatibleWithVarArgsAttrs())
1104 // Okay, we decided that this is a safe thing to do: go ahead and start
1105 // inserting cast instructions as necessary.
1106 std::vector<Value*> Args;
1107 Args.reserve(NumActualArgs);
1108 SmallVector<AttributeWithIndex, 8> attrVec;
1109 attrVec.reserve(NumCommonArgs);
1111 // Get any return attributes.
1112 Attributes::Builder RAttrs = CallerPAL.getRetAttributes();
1114 // If the return value is not being used, the type may not be compatible
1115 // with the existing attributes. Wipe out any problematic attributes.
1116 RAttrs.removeAttributes(Attributes::typeIncompatible(NewRetTy));
1118 // Add the new return attributes.
1119 if (RAttrs.hasAttributes())
1120 attrVec.push_back(AttributeWithIndex::get(0, Attributes::get(RAttrs)));
1122 AI = CS.arg_begin();
1123 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1124 Type *ParamTy = FT->getParamType(i);
1125 if ((*AI)->getType() == ParamTy) {
1126 Args.push_back(*AI);
1128 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1129 false, ParamTy, false);
1130 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy));
1133 // Add any parameter attributes.
1134 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1135 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1138 // If the function takes more arguments than the call was taking, add them
1140 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1141 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1143 // If we are removing arguments to the function, emit an obnoxious warning.
1144 if (FT->getNumParams() < NumActualArgs) {
1145 if (!FT->isVarArg()) {
1146 errs() << "WARNING: While resolving call to function '"
1147 << Callee->getName() << "' arguments were dropped!\n";
1149 // Add all of the arguments in their promoted form to the arg list.
1150 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1151 Type *PTy = getPromotedType((*AI)->getType());
1152 if (PTy != (*AI)->getType()) {
1153 // Must promote to pass through va_arg area!
1154 Instruction::CastOps opcode =
1155 CastInst::getCastOpcode(*AI, false, PTy, false);
1156 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1158 Args.push_back(*AI);
1161 // Add any parameter attributes.
1162 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1163 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1168 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
1169 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1171 if (NewRetTy->isVoidTy())
1172 Caller->setName(""); // Void type should not have a name.
1174 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec);
1177 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1178 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1179 II->getUnwindDest(), Args);
1181 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1182 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1184 CallInst *CI = cast<CallInst>(Caller);
1185 NC = Builder->CreateCall(Callee, Args);
1187 if (CI->isTailCall())
1188 cast<CallInst>(NC)->setTailCall();
1189 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1190 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1193 // Insert a cast of the return type as necessary.
1195 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1196 if (!NV->getType()->isVoidTy()) {
1197 Instruction::CastOps opcode =
1198 CastInst::getCastOpcode(NC, false, OldRetTy, false);
1199 NV = NC = CastInst::Create(opcode, NC, OldRetTy);
1200 NC->setDebugLoc(Caller->getDebugLoc());
1202 // If this is an invoke instruction, we should insert it after the first
1203 // non-phi, instruction in the normal successor block.
1204 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1205 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1206 InsertNewInstBefore(NC, *I);
1208 // Otherwise, it's a call, just insert cast right after the call.
1209 InsertNewInstBefore(NC, *Caller);
1211 Worklist.AddUsersToWorkList(*Caller);
1213 NV = UndefValue::get(Caller->getType());
1217 if (!Caller->use_empty())
1218 ReplaceInstUsesWith(*Caller, NV);
1220 EraseInstFromFunction(*Caller);
1224 // transformCallThroughTrampoline - Turn a call to a function created by
1225 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1226 // underlying function.
1229 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1230 IntrinsicInst *Tramp) {
1231 Value *Callee = CS.getCalledValue();
1232 PointerType *PTy = cast<PointerType>(Callee->getType());
1233 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1234 const AttrListPtr &Attrs = CS.getAttributes();
1236 // If the call already has the 'nest' attribute somewhere then give up -
1237 // otherwise 'nest' would occur twice after splicing in the chain.
1238 for (unsigned I = 0, E = Attrs.getNumAttrs(); I != E; ++I)
1239 if (Attrs.getAttributesAtIndex(I).hasAttribute(Attributes::Nest))
1243 "transformCallThroughTrampoline called with incorrect CallSite.");
1245 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1246 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1247 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1249 const AttrListPtr &NestAttrs = NestF->getAttributes();
1250 if (!NestAttrs.isEmpty()) {
1251 unsigned NestIdx = 1;
1253 Attributes NestAttr;
1255 // Look for a parameter marked with the 'nest' attribute.
1256 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1257 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1258 if (NestAttrs.getParamAttributes(NestIdx).hasAttribute(Attributes::Nest)){
1259 // Record the parameter type and any other attributes.
1261 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1266 Instruction *Caller = CS.getInstruction();
1267 std::vector<Value*> NewArgs;
1268 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1270 SmallVector<AttributeWithIndex, 8> NewAttrs;
1271 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1273 // Insert the nest argument into the call argument list, which may
1274 // mean appending it. Likewise for attributes.
1276 // Add any result attributes.
1277 if (Attributes Attr = Attrs.getRetAttributes())
1278 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1282 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1284 if (Idx == NestIdx) {
1285 // Add the chain argument and attributes.
1286 Value *NestVal = Tramp->getArgOperand(2);
1287 if (NestVal->getType() != NestTy)
1288 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1289 NewArgs.push_back(NestVal);
1290 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1296 // Add the original argument and attributes.
1297 NewArgs.push_back(*I);
1298 if (Attributes Attr = Attrs.getParamAttributes(Idx))
1300 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1306 // Add any function attributes.
1307 if (Attributes Attr = Attrs.getFnAttributes())
1308 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1310 // The trampoline may have been bitcast to a bogus type (FTy).
1311 // Handle this by synthesizing a new function type, equal to FTy
1312 // with the chain parameter inserted.
1314 std::vector<Type*> NewTypes;
1315 NewTypes.reserve(FTy->getNumParams()+1);
1317 // Insert the chain's type into the list of parameter types, which may
1318 // mean appending it.
1321 FunctionType::param_iterator I = FTy->param_begin(),
1322 E = FTy->param_end();
1326 // Add the chain's type.
1327 NewTypes.push_back(NestTy);
1332 // Add the original type.
1333 NewTypes.push_back(*I);
1339 // Replace the trampoline call with a direct call. Let the generic
1340 // code sort out any function type mismatches.
1341 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1343 Constant *NewCallee =
1344 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1345 NestF : ConstantExpr::getBitCast(NestF,
1346 PointerType::getUnqual(NewFTy));
1347 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs);
1349 Instruction *NewCaller;
1350 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1351 NewCaller = InvokeInst::Create(NewCallee,
1352 II->getNormalDest(), II->getUnwindDest(),
1354 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1355 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1357 NewCaller = CallInst::Create(NewCallee, NewArgs);
1358 if (cast<CallInst>(Caller)->isTailCall())
1359 cast<CallInst>(NewCaller)->setTailCall();
1360 cast<CallInst>(NewCaller)->
1361 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1362 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1369 // Replace the trampoline call with a direct call. Since there is no 'nest'
1370 // parameter, there is no need to adjust the argument list. Let the generic
1371 // code sort out any function type mismatches.
1372 Constant *NewCallee =
1373 NestF->getType() == PTy ? NestF :
1374 ConstantExpr::getBitCast(NestF, PTy);
1375 CS.setCalledFunction(NewCallee);
1376 return CS.getInstruction();