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/ADT/Statistic.h"
16 #include "llvm/Analysis/MemoryBuiltins.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Transforms/Utils/BuildLibCalls.h"
21 #include "llvm/Transforms/Utils/Local.h"
23 using namespace PatternMatch;
25 STATISTIC(NumSimplified, "Number of library calls simplified");
27 /// getPromotedType - Return the specified type promoted as it would be to pass
28 /// though a va_arg area.
29 static Type *getPromotedType(Type *Ty) {
30 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
31 if (ITy->getBitWidth() < 32)
32 return Type::getInt32Ty(Ty->getContext());
37 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
38 /// single scalar element, like {{{type}}} or [1 x type], return type.
39 static Type *reduceToSingleValueType(Type *T) {
40 while (!T->isSingleValueType()) {
41 if (StructType *STy = dyn_cast<StructType>(T)) {
42 if (STy->getNumElements() == 1)
43 T = STy->getElementType(0);
46 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
47 if (ATy->getNumElements() == 1)
48 T = ATy->getElementType();
58 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
59 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL);
60 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL);
61 unsigned MinAlign = std::min(DstAlign, SrcAlign);
62 unsigned CopyAlign = MI->getAlignment();
64 if (CopyAlign < MinAlign) {
65 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
70 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
72 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
73 if (MemOpLength == 0) return 0;
75 // Source and destination pointer types are always "i8*" for intrinsic. See
76 // if the size is something we can handle with a single primitive load/store.
77 // A single load+store correctly handles overlapping memory in the memmove
79 uint64_t Size = MemOpLength->getLimitedValue();
80 assert(Size && "0-sized memory transferring should be removed already.");
82 if (Size > 8 || (Size&(Size-1)))
83 return 0; // If not 1/2/4/8 bytes, exit.
85 // Use an integer load+store unless we can find something better.
87 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
89 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
91 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
92 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
93 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
95 // Memcpy forces the use of i8* for the source and destination. That means
96 // that if you're using memcpy to move one double around, you'll get a cast
97 // from double* to i8*. We'd much rather use a double load+store rather than
98 // an i64 load+store, here because this improves the odds that the source or
99 // dest address will be promotable. See if we can find a better type than the
101 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
103 if (StrippedDest != MI->getArgOperand(0)) {
104 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
106 if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
107 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
108 // down through these levels if so.
109 SrcETy = reduceToSingleValueType(SrcETy);
111 if (SrcETy->isSingleValueType()) {
112 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
113 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
115 // If the memcpy has metadata describing the members, see if we can
116 // get the TBAA tag describing our copy.
117 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
118 if (M->getNumOperands() == 3 &&
120 isa<ConstantInt>(M->getOperand(0)) &&
121 cast<ConstantInt>(M->getOperand(0))->isNullValue() &&
123 isa<ConstantInt>(M->getOperand(1)) &&
124 cast<ConstantInt>(M->getOperand(1))->getValue() == Size &&
126 isa<MDNode>(M->getOperand(2)))
127 CopyMD = cast<MDNode>(M->getOperand(2));
133 // If the memcpy/memmove provides better alignment info than we can
135 SrcAlign = std::max(SrcAlign, CopyAlign);
136 DstAlign = std::max(DstAlign, CopyAlign);
138 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
139 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
140 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
141 L->setAlignment(SrcAlign);
143 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
144 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
145 S->setAlignment(DstAlign);
147 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
149 // Set the size of the copy to 0, it will be deleted on the next iteration.
150 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
154 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
155 unsigned Alignment = getKnownAlignment(MI->getDest(), DL);
156 if (MI->getAlignment() < Alignment) {
157 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
162 // Extract the length and alignment and fill if they are constant.
163 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
164 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
165 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
167 uint64_t Len = LenC->getLimitedValue();
168 Alignment = MI->getAlignment();
169 assert(Len && "0-sized memory setting should be removed already.");
171 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
172 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
173 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
175 Value *Dest = MI->getDest();
176 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
177 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
178 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
180 // Alignment 0 is identity for alignment 1 for memset, but not store.
181 if (Alignment == 0) Alignment = 1;
183 // Extract the fill value and store.
184 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
185 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
187 S->setAlignment(Alignment);
189 // Set the size of the copy to 0, it will be deleted on the next iteration.
190 MI->setLength(Constant::getNullValue(LenC->getType()));
197 /// visitCallInst - CallInst simplification. This mostly only handles folding
198 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
199 /// the heavy lifting.
201 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
202 if (isFreeCall(&CI, TLI))
203 return visitFree(CI);
205 // If the caller function is nounwind, mark the call as nounwind, even if the
207 if (CI.getParent()->getParent()->doesNotThrow() &&
208 !CI.doesNotThrow()) {
209 CI.setDoesNotThrow();
213 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
214 if (!II) return visitCallSite(&CI);
216 // Intrinsics cannot occur in an invoke, so handle them here instead of in
218 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
219 bool Changed = false;
221 // memmove/cpy/set of zero bytes is a noop.
222 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
223 if (NumBytes->isNullValue())
224 return EraseInstFromFunction(CI);
226 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
227 if (CI->getZExtValue() == 1) {
228 // Replace the instruction with just byte operations. We would
229 // transform other cases to loads/stores, but we don't know if
230 // alignment is sufficient.
234 // No other transformations apply to volatile transfers.
235 if (MI->isVolatile())
238 // If we have a memmove and the source operation is a constant global,
239 // then the source and dest pointers can't alias, so we can change this
240 // into a call to memcpy.
241 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
242 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
243 if (GVSrc->isConstant()) {
244 Module *M = CI.getParent()->getParent()->getParent();
245 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
246 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
247 CI.getArgOperand(1)->getType(),
248 CI.getArgOperand(2)->getType() };
249 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
254 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
255 // memmove(x,x,size) -> noop.
256 if (MTI->getSource() == MTI->getDest())
257 return EraseInstFromFunction(CI);
260 // If we can determine a pointer alignment that is bigger than currently
261 // set, update the alignment.
262 if (isa<MemTransferInst>(MI)) {
263 if (Instruction *I = SimplifyMemTransfer(MI))
265 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
266 if (Instruction *I = SimplifyMemSet(MSI))
270 if (Changed) return II;
273 switch (II->getIntrinsicID()) {
275 case Intrinsic::objectsize: {
277 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
278 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
281 case Intrinsic::bswap: {
282 Value *IIOperand = II->getArgOperand(0);
285 // bswap(bswap(x)) -> x
286 if (match(IIOperand, m_BSwap(m_Value(X))))
287 return ReplaceInstUsesWith(CI, X);
289 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
290 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
291 unsigned C = X->getType()->getPrimitiveSizeInBits() -
292 IIOperand->getType()->getPrimitiveSizeInBits();
293 Value *CV = ConstantInt::get(X->getType(), C);
294 Value *V = Builder->CreateLShr(X, CV);
295 return new TruncInst(V, IIOperand->getType());
300 case Intrinsic::powi:
301 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
304 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
307 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
308 // powi(x, -1) -> 1/x
309 if (Power->isAllOnesValue())
310 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
311 II->getArgOperand(0));
314 case Intrinsic::cttz: {
315 // If all bits below the first known one are known zero,
316 // this value is constant.
317 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
318 // FIXME: Try to simplify vectors of integers.
320 uint32_t BitWidth = IT->getBitWidth();
321 APInt KnownZero(BitWidth, 0);
322 APInt KnownOne(BitWidth, 0);
323 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
324 unsigned TrailingZeros = KnownOne.countTrailingZeros();
325 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
326 if ((Mask & KnownZero) == Mask)
327 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
328 APInt(BitWidth, TrailingZeros)));
332 case Intrinsic::ctlz: {
333 // If all bits above the first known one are known zero,
334 // this value is constant.
335 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
336 // FIXME: Try to simplify vectors of integers.
338 uint32_t BitWidth = IT->getBitWidth();
339 APInt KnownZero(BitWidth, 0);
340 APInt KnownOne(BitWidth, 0);
341 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
342 unsigned LeadingZeros = KnownOne.countLeadingZeros();
343 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
344 if ((Mask & KnownZero) == Mask)
345 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
346 APInt(BitWidth, LeadingZeros)));
350 case Intrinsic::uadd_with_overflow: {
351 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
352 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
353 uint32_t BitWidth = IT->getBitWidth();
354 APInt LHSKnownZero(BitWidth, 0);
355 APInt LHSKnownOne(BitWidth, 0);
356 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
357 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
358 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
360 if (LHSKnownNegative || LHSKnownPositive) {
361 APInt RHSKnownZero(BitWidth, 0);
362 APInt RHSKnownOne(BitWidth, 0);
363 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
364 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
365 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
366 if (LHSKnownNegative && RHSKnownNegative) {
367 // The sign bit is set in both cases: this MUST overflow.
368 // Create a simple add instruction, and insert it into the struct.
369 Value *Add = Builder->CreateAdd(LHS, RHS);
372 UndefValue::get(LHS->getType()),
373 ConstantInt::getTrue(II->getContext())
375 StructType *ST = cast<StructType>(II->getType());
376 Constant *Struct = ConstantStruct::get(ST, V);
377 return InsertValueInst::Create(Struct, Add, 0);
380 if (LHSKnownPositive && RHSKnownPositive) {
381 // The sign bit is clear in both cases: this CANNOT overflow.
382 // Create a simple add instruction, and insert it into the struct.
383 Value *Add = Builder->CreateNUWAdd(LHS, RHS);
386 UndefValue::get(LHS->getType()),
387 ConstantInt::getFalse(II->getContext())
389 StructType *ST = cast<StructType>(II->getType());
390 Constant *Struct = ConstantStruct::get(ST, V);
391 return InsertValueInst::Create(Struct, Add, 0);
395 // FALL THROUGH uadd into sadd
396 case Intrinsic::sadd_with_overflow:
397 // Canonicalize constants into the RHS.
398 if (isa<Constant>(II->getArgOperand(0)) &&
399 !isa<Constant>(II->getArgOperand(1))) {
400 Value *LHS = II->getArgOperand(0);
401 II->setArgOperand(0, II->getArgOperand(1));
402 II->setArgOperand(1, LHS);
406 // X + undef -> undef
407 if (isa<UndefValue>(II->getArgOperand(1)))
408 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
410 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
411 // X + 0 -> {X, false}
414 UndefValue::get(II->getArgOperand(0)->getType()),
415 ConstantInt::getFalse(II->getContext())
418 ConstantStruct::get(cast<StructType>(II->getType()), V);
419 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
423 case Intrinsic::usub_with_overflow:
424 case Intrinsic::ssub_with_overflow:
425 // undef - X -> undef
426 // X - undef -> undef
427 if (isa<UndefValue>(II->getArgOperand(0)) ||
428 isa<UndefValue>(II->getArgOperand(1)))
429 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
431 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
432 // X - 0 -> {X, false}
435 UndefValue::get(II->getArgOperand(0)->getType()),
436 ConstantInt::getFalse(II->getContext())
439 ConstantStruct::get(cast<StructType>(II->getType()), V);
440 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
444 case Intrinsic::umul_with_overflow: {
445 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
446 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
448 APInt LHSKnownZero(BitWidth, 0);
449 APInt LHSKnownOne(BitWidth, 0);
450 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
451 APInt RHSKnownZero(BitWidth, 0);
452 APInt RHSKnownOne(BitWidth, 0);
453 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
455 // Get the largest possible values for each operand.
456 APInt LHSMax = ~LHSKnownZero;
457 APInt RHSMax = ~RHSKnownZero;
459 // If multiplying the maximum values does not overflow then we can turn
460 // this into a plain NUW mul.
462 LHSMax.umul_ov(RHSMax, Overflow);
464 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
466 UndefValue::get(LHS->getType()),
469 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
470 return InsertValueInst::Create(Struct, Mul, 0);
473 case Intrinsic::smul_with_overflow:
474 // Canonicalize constants into the RHS.
475 if (isa<Constant>(II->getArgOperand(0)) &&
476 !isa<Constant>(II->getArgOperand(1))) {
477 Value *LHS = II->getArgOperand(0);
478 II->setArgOperand(0, II->getArgOperand(1));
479 II->setArgOperand(1, LHS);
483 // X * undef -> undef
484 if (isa<UndefValue>(II->getArgOperand(1)))
485 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
487 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
490 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
492 // X * 1 -> {X, false}
493 if (RHSI->equalsInt(1)) {
495 UndefValue::get(II->getArgOperand(0)->getType()),
496 ConstantInt::getFalse(II->getContext())
499 ConstantStruct::get(cast<StructType>(II->getType()), V);
500 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
504 case Intrinsic::ppc_altivec_lvx:
505 case Intrinsic::ppc_altivec_lvxl:
506 // Turn PPC lvx -> load if the pointer is known aligned.
507 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
508 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
509 PointerType::getUnqual(II->getType()));
510 return new LoadInst(Ptr);
513 case Intrinsic::ppc_altivec_stvx:
514 case Intrinsic::ppc_altivec_stvxl:
515 // Turn stvx -> store if the pointer is known aligned.
516 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL) >= 16) {
518 PointerType::getUnqual(II->getArgOperand(0)->getType());
519 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
520 return new StoreInst(II->getArgOperand(0), Ptr);
523 case Intrinsic::x86_sse_storeu_ps:
524 case Intrinsic::x86_sse2_storeu_pd:
525 case Intrinsic::x86_sse2_storeu_dq:
526 // Turn X86 storeu -> store if the pointer is known aligned.
527 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
529 PointerType::getUnqual(II->getArgOperand(1)->getType());
530 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
531 return new StoreInst(II->getArgOperand(1), Ptr);
535 case Intrinsic::x86_sse_cvtss2si:
536 case Intrinsic::x86_sse_cvtss2si64:
537 case Intrinsic::x86_sse_cvttss2si:
538 case Intrinsic::x86_sse_cvttss2si64:
539 case Intrinsic::x86_sse2_cvtsd2si:
540 case Intrinsic::x86_sse2_cvtsd2si64:
541 case Intrinsic::x86_sse2_cvttsd2si:
542 case Intrinsic::x86_sse2_cvttsd2si64: {
543 // These intrinsics only demand the 0th element of their input vectors. If
544 // we can simplify the input based on that, do so now.
546 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
547 APInt DemandedElts(VWidth, 1);
548 APInt UndefElts(VWidth, 0);
549 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
550 DemandedElts, UndefElts)) {
551 II->setArgOperand(0, V);
558 case Intrinsic::x86_sse41_pmovsxbw:
559 case Intrinsic::x86_sse41_pmovsxwd:
560 case Intrinsic::x86_sse41_pmovsxdq:
561 case Intrinsic::x86_sse41_pmovzxbw:
562 case Intrinsic::x86_sse41_pmovzxwd:
563 case Intrinsic::x86_sse41_pmovzxdq: {
564 // pmov{s|z}x ignores the upper half of their input vectors.
566 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
567 unsigned LowHalfElts = VWidth / 2;
568 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
569 APInt UndefElts(VWidth, 0);
570 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
573 II->setArgOperand(0, TmpV);
579 case Intrinsic::ppc_altivec_vperm:
580 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
581 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
582 assert(Mask->getType()->getVectorNumElements() == 16 &&
583 "Bad type for intrinsic!");
585 // Check that all of the elements are integer constants or undefs.
586 bool AllEltsOk = true;
587 for (unsigned i = 0; i != 16; ++i) {
588 Constant *Elt = Mask->getAggregateElement(i);
590 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
597 // Cast the input vectors to byte vectors.
598 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
600 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
602 Value *Result = UndefValue::get(Op0->getType());
604 // Only extract each element once.
605 Value *ExtractedElts[32];
606 memset(ExtractedElts, 0, sizeof(ExtractedElts));
608 for (unsigned i = 0; i != 16; ++i) {
609 if (isa<UndefValue>(Mask->getAggregateElement(i)))
612 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
613 Idx &= 31; // Match the hardware behavior.
615 if (ExtractedElts[Idx] == 0) {
617 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
618 Builder->getInt32(Idx&15));
621 // Insert this value into the result vector.
622 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
623 Builder->getInt32(i));
625 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
630 case Intrinsic::arm_neon_vld1:
631 case Intrinsic::arm_neon_vld2:
632 case Intrinsic::arm_neon_vld3:
633 case Intrinsic::arm_neon_vld4:
634 case Intrinsic::arm_neon_vld2lane:
635 case Intrinsic::arm_neon_vld3lane:
636 case Intrinsic::arm_neon_vld4lane:
637 case Intrinsic::arm_neon_vst1:
638 case Intrinsic::arm_neon_vst2:
639 case Intrinsic::arm_neon_vst3:
640 case Intrinsic::arm_neon_vst4:
641 case Intrinsic::arm_neon_vst2lane:
642 case Intrinsic::arm_neon_vst3lane:
643 case Intrinsic::arm_neon_vst4lane: {
644 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL);
645 unsigned AlignArg = II->getNumArgOperands() - 1;
646 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
647 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
648 II->setArgOperand(AlignArg,
649 ConstantInt::get(Type::getInt32Ty(II->getContext()),
656 case Intrinsic::arm_neon_vmulls:
657 case Intrinsic::arm_neon_vmullu: {
658 Value *Arg0 = II->getArgOperand(0);
659 Value *Arg1 = II->getArgOperand(1);
661 // Handle mul by zero first:
662 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
663 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
666 // Check for constant LHS & RHS - in this case we just simplify.
667 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu);
668 VectorType *NewVT = cast<VectorType>(II->getType());
669 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
670 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
671 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
672 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
674 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
677 // Couldn't simplify - canonicalize constant to the RHS.
678 std::swap(Arg0, Arg1);
681 // Handle mul by one:
682 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
683 if (ConstantInt *Splat =
684 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
686 return CastInst::CreateIntegerCast(Arg0, II->getType(),
692 case Intrinsic::stackrestore: {
693 // If the save is right next to the restore, remove the restore. This can
694 // happen when variable allocas are DCE'd.
695 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
696 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
697 BasicBlock::iterator BI = SS;
699 return EraseInstFromFunction(CI);
703 // Scan down this block to see if there is another stack restore in the
704 // same block without an intervening call/alloca.
705 BasicBlock::iterator BI = II;
706 TerminatorInst *TI = II->getParent()->getTerminator();
707 bool CannotRemove = false;
708 for (++BI; &*BI != TI; ++BI) {
709 if (isa<AllocaInst>(BI)) {
713 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
714 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
715 // If there is a stackrestore below this one, remove this one.
716 if (II->getIntrinsicID() == Intrinsic::stackrestore)
717 return EraseInstFromFunction(CI);
718 // Otherwise, ignore the intrinsic.
720 // If we found a non-intrinsic call, we can't remove the stack
728 // If the stack restore is in a return, resume, or unwind block and if there
729 // are no allocas or calls between the restore and the return, nuke the
731 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
732 return EraseInstFromFunction(CI);
737 return visitCallSite(II);
740 // InvokeInst simplification
742 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
743 return visitCallSite(&II);
746 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
747 /// passed through the varargs area, we can eliminate the use of the cast.
748 static bool isSafeToEliminateVarargsCast(const CallSite CS,
749 const CastInst * const CI,
750 const DataLayout * const DL,
752 if (!CI->isLosslessCast())
755 // The size of ByVal or InAlloca arguments is derived from the type, so we
756 // can't change to a type with a different size. If the size were
757 // passed explicitly we could avoid this check.
758 if (!CS.isByValOrInAllocaArgument(ix))
762 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
763 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
764 if (!SrcTy->isSized() || !DstTy->isSized())
766 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
771 // Try to fold some different type of calls here.
772 // Currently we're only working with the checking functions, memcpy_chk,
773 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
774 // strcat_chk and strncat_chk.
775 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
776 if (CI->getCalledFunction() == 0) return 0;
778 if (Value *With = Simplifier->optimizeCall(CI)) {
780 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
786 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
787 // Strip off at most one level of pointer casts, looking for an alloca. This
788 // is good enough in practice and simpler than handling any number of casts.
789 Value *Underlying = TrampMem->stripPointerCasts();
790 if (Underlying != TrampMem &&
791 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
793 if (!isa<AllocaInst>(Underlying))
796 IntrinsicInst *InitTrampoline = 0;
797 for (User *U : TrampMem->users()) {
798 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
801 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
803 // More than one init_trampoline writes to this value. Give up.
808 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
809 // Allow any number of calls to adjust.trampoline.
814 // No call to init.trampoline found.
818 // Check that the alloca is being used in the expected way.
819 if (InitTrampoline->getOperand(0) != TrampMem)
822 return InitTrampoline;
825 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
827 // Visit all the previous instructions in the basic block, and try to find a
828 // init.trampoline which has a direct path to the adjust.trampoline.
829 for (BasicBlock::iterator I = AdjustTramp,
830 E = AdjustTramp->getParent()->begin(); I != E; ) {
831 Instruction *Inst = --I;
832 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
833 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
834 II->getOperand(0) == TrampMem)
836 if (Inst->mayWriteToMemory())
842 // Given a call to llvm.adjust.trampoline, find and return the corresponding
843 // call to llvm.init.trampoline if the call to the trampoline can be optimized
844 // to a direct call to a function. Otherwise return NULL.
846 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
847 Callee = Callee->stripPointerCasts();
848 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
850 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
853 Value *TrampMem = AdjustTramp->getOperand(0);
855 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
857 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
862 // visitCallSite - Improvements for call and invoke instructions.
864 Instruction *InstCombiner::visitCallSite(CallSite CS) {
865 if (isAllocLikeFn(CS.getInstruction(), TLI))
866 return visitAllocSite(*CS.getInstruction());
868 bool Changed = false;
870 // If the callee is a pointer to a function, attempt to move any casts to the
871 // arguments of the call/invoke.
872 Value *Callee = CS.getCalledValue();
873 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
876 if (Function *CalleeF = dyn_cast<Function>(Callee))
877 // If the call and callee calling conventions don't match, this call must
878 // be unreachable, as the call is undefined.
879 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
880 // Only do this for calls to a function with a body. A prototype may
881 // not actually end up matching the implementation's calling conv for a
882 // variety of reasons (e.g. it may be written in assembly).
883 !CalleeF->isDeclaration()) {
884 Instruction *OldCall = CS.getInstruction();
885 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
886 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
888 // If OldCall does not return void then replaceAllUsesWith undef.
889 // This allows ValueHandlers and custom metadata to adjust itself.
890 if (!OldCall->getType()->isVoidTy())
891 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
892 if (isa<CallInst>(OldCall))
893 return EraseInstFromFunction(*OldCall);
895 // We cannot remove an invoke, because it would change the CFG, just
896 // change the callee to a null pointer.
897 cast<InvokeInst>(OldCall)->setCalledFunction(
898 Constant::getNullValue(CalleeF->getType()));
902 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
903 // If CS does not return void then replaceAllUsesWith undef.
904 // This allows ValueHandlers and custom metadata to adjust itself.
905 if (!CS.getInstruction()->getType()->isVoidTy())
906 ReplaceInstUsesWith(*CS.getInstruction(),
907 UndefValue::get(CS.getInstruction()->getType()));
909 if (isa<InvokeInst>(CS.getInstruction())) {
910 // Can't remove an invoke because we cannot change the CFG.
914 // This instruction is not reachable, just remove it. We insert a store to
915 // undef so that we know that this code is not reachable, despite the fact
916 // that we can't modify the CFG here.
917 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
918 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
919 CS.getInstruction());
921 return EraseInstFromFunction(*CS.getInstruction());
924 if (IntrinsicInst *II = FindInitTrampoline(Callee))
925 return transformCallThroughTrampoline(CS, II);
927 PointerType *PTy = cast<PointerType>(Callee->getType());
928 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
929 if (FTy->isVarArg()) {
930 int ix = FTy->getNumParams();
931 // See if we can optimize any arguments passed through the varargs area of
933 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
934 E = CS.arg_end(); I != E; ++I, ++ix) {
935 CastInst *CI = dyn_cast<CastInst>(*I);
936 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
937 *I = CI->getOperand(0);
943 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
944 // Inline asm calls cannot throw - mark them 'nounwind'.
945 CS.setDoesNotThrow();
949 // Try to optimize the call if possible, we require DataLayout for most of
950 // this. None of these calls are seen as possibly dead so go ahead and
951 // delete the instruction now.
952 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
953 Instruction *I = tryOptimizeCall(CI, DL);
954 // If we changed something return the result, etc. Otherwise let
955 // the fallthrough check.
956 if (I) return EraseInstFromFunction(*I);
959 return Changed ? CS.getInstruction() : 0;
962 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
963 // attempt to move the cast to the arguments of the call/invoke.
965 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
967 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
970 Instruction *Caller = CS.getInstruction();
971 const AttributeSet &CallerPAL = CS.getAttributes();
973 // Okay, this is a cast from a function to a different type. Unless doing so
974 // would cause a type conversion of one of our arguments, change this call to
975 // be a direct call with arguments casted to the appropriate types.
977 FunctionType *FT = Callee->getFunctionType();
978 Type *OldRetTy = Caller->getType();
979 Type *NewRetTy = FT->getReturnType();
981 // Check to see if we are changing the return type...
982 if (OldRetTy != NewRetTy) {
984 if (NewRetTy->isStructTy())
985 return false; // TODO: Handle multiple return values.
987 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
988 if (Callee->isDeclaration())
989 return false; // Cannot transform this return value.
991 if (!Caller->use_empty() &&
992 // void -> non-void is handled specially
993 !NewRetTy->isVoidTy())
994 return false; // Cannot transform this return value.
997 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
998 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1000 hasAttributes(AttributeFuncs::
1001 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1002 AttributeSet::ReturnIndex))
1003 return false; // Attribute not compatible with transformed value.
1006 // If the callsite is an invoke instruction, and the return value is used by
1007 // a PHI node in a successor, we cannot change the return type of the call
1008 // because there is no place to put the cast instruction (without breaking
1009 // the critical edge). Bail out in this case.
1010 if (!Caller->use_empty())
1011 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1012 for (User *U : II->users())
1013 if (PHINode *PN = dyn_cast<PHINode>(U))
1014 if (PN->getParent() == II->getNormalDest() ||
1015 PN->getParent() == II->getUnwindDest())
1019 unsigned NumActualArgs = CS.arg_size();
1020 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1022 CallSite::arg_iterator AI = CS.arg_begin();
1023 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1024 Type *ParamTy = FT->getParamType(i);
1025 Type *ActTy = (*AI)->getType();
1027 if (!CastInst::isBitCastable(ActTy, ParamTy))
1028 return false; // Cannot transform this parameter value.
1030 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1031 hasAttributes(AttributeFuncs::
1032 typeIncompatible(ParamTy, i + 1), i + 1))
1033 return false; // Attribute not compatible with transformed value.
1035 if (CS.isInAllocaArgument(i))
1036 return false; // Cannot transform to and from inalloca.
1038 // If the parameter is passed as a byval argument, then we have to have a
1039 // sized type and the sized type has to have the same size as the old type.
1040 if (ParamTy != ActTy &&
1041 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1042 Attribute::ByVal)) {
1043 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1044 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || DL == 0)
1047 Type *CurElTy = ActTy->getPointerElementType();
1048 if (DL->getTypeAllocSize(CurElTy) !=
1049 DL->getTypeAllocSize(ParamPTy->getElementType()))
1054 if (Callee->isDeclaration()) {
1055 // Do not delete arguments unless we have a function body.
1056 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1059 // If the callee is just a declaration, don't change the varargsness of the
1060 // call. We don't want to introduce a varargs call where one doesn't
1062 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1063 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1066 // If both the callee and the cast type are varargs, we still have to make
1067 // sure the number of fixed parameters are the same or we have the same
1068 // ABI issues as if we introduce a varargs call.
1069 if (FT->isVarArg() &&
1070 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1071 FT->getNumParams() !=
1072 cast<FunctionType>(APTy->getElementType())->getNumParams())
1076 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1077 !CallerPAL.isEmpty())
1078 // In this case we have more arguments than the new function type, but we
1079 // won't be dropping them. Check that these extra arguments have attributes
1080 // that are compatible with being a vararg call argument.
1081 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1082 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1083 if (Index <= FT->getNumParams())
1086 // Check if it has an attribute that's incompatible with varargs.
1087 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1088 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1093 // Okay, we decided that this is a safe thing to do: go ahead and start
1094 // inserting cast instructions as necessary.
1095 std::vector<Value*> Args;
1096 Args.reserve(NumActualArgs);
1097 SmallVector<AttributeSet, 8> attrVec;
1098 attrVec.reserve(NumCommonArgs);
1100 // Get any return attributes.
1101 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1103 // If the return value is not being used, the type may not be compatible
1104 // with the existing attributes. Wipe out any problematic attributes.
1106 removeAttributes(AttributeFuncs::
1107 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1108 AttributeSet::ReturnIndex);
1110 // Add the new return attributes.
1111 if (RAttrs.hasAttributes())
1112 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1113 AttributeSet::ReturnIndex, RAttrs));
1115 AI = CS.arg_begin();
1116 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1117 Type *ParamTy = FT->getParamType(i);
1119 if ((*AI)->getType() == ParamTy) {
1120 Args.push_back(*AI);
1122 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1125 // Add any parameter attributes.
1126 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1127 if (PAttrs.hasAttributes())
1128 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1132 // If the function takes more arguments than the call was taking, add them
1134 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1135 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1137 // If we are removing arguments to the function, emit an obnoxious warning.
1138 if (FT->getNumParams() < NumActualArgs) {
1139 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1140 if (FT->isVarArg()) {
1141 // Add all of the arguments in their promoted form to the arg list.
1142 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1143 Type *PTy = getPromotedType((*AI)->getType());
1144 if (PTy != (*AI)->getType()) {
1145 // Must promote to pass through va_arg area!
1146 Instruction::CastOps opcode =
1147 CastInst::getCastOpcode(*AI, false, PTy, false);
1148 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1150 Args.push_back(*AI);
1153 // Add any parameter attributes.
1154 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1155 if (PAttrs.hasAttributes())
1156 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1162 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1163 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1164 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1166 if (NewRetTy->isVoidTy())
1167 Caller->setName(""); // Void type should not have a name.
1169 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1173 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1174 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1175 II->getUnwindDest(), Args);
1177 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1178 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1180 CallInst *CI = cast<CallInst>(Caller);
1181 NC = Builder->CreateCall(Callee, Args);
1183 if (CI->isTailCall())
1184 cast<CallInst>(NC)->setTailCall();
1185 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1186 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1189 // Insert a cast of the return type as necessary.
1191 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1192 if (!NV->getType()->isVoidTy()) {
1193 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1194 NC->setDebugLoc(Caller->getDebugLoc());
1196 // If this is an invoke instruction, we should insert it after the first
1197 // non-phi, instruction in the normal successor block.
1198 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1199 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1200 InsertNewInstBefore(NC, *I);
1202 // Otherwise, it's a call, just insert cast right after the call.
1203 InsertNewInstBefore(NC, *Caller);
1205 Worklist.AddUsersToWorkList(*Caller);
1207 NV = UndefValue::get(Caller->getType());
1211 if (!Caller->use_empty())
1212 ReplaceInstUsesWith(*Caller, NV);
1213 else if (Caller->hasValueHandle())
1214 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1216 EraseInstFromFunction(*Caller);
1220 // transformCallThroughTrampoline - Turn a call to a function created by
1221 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1222 // underlying function.
1225 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1226 IntrinsicInst *Tramp) {
1227 Value *Callee = CS.getCalledValue();
1228 PointerType *PTy = cast<PointerType>(Callee->getType());
1229 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1230 const AttributeSet &Attrs = CS.getAttributes();
1232 // If the call already has the 'nest' attribute somewhere then give up -
1233 // otherwise 'nest' would occur twice after splicing in the chain.
1234 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1238 "transformCallThroughTrampoline called with incorrect CallSite.");
1240 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1241 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1242 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1244 const AttributeSet &NestAttrs = NestF->getAttributes();
1245 if (!NestAttrs.isEmpty()) {
1246 unsigned NestIdx = 1;
1248 AttributeSet NestAttr;
1250 // Look for a parameter marked with the 'nest' attribute.
1251 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1252 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1253 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1254 // Record the parameter type and any other attributes.
1256 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1261 Instruction *Caller = CS.getInstruction();
1262 std::vector<Value*> NewArgs;
1263 NewArgs.reserve(CS.arg_size() + 1);
1265 SmallVector<AttributeSet, 8> NewAttrs;
1266 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1268 // Insert the nest argument into the call argument list, which may
1269 // mean appending it. Likewise for attributes.
1271 // Add any result attributes.
1272 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1273 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1274 Attrs.getRetAttributes()));
1278 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1280 if (Idx == NestIdx) {
1281 // Add the chain argument and attributes.
1282 Value *NestVal = Tramp->getArgOperand(2);
1283 if (NestVal->getType() != NestTy)
1284 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1285 NewArgs.push_back(NestVal);
1286 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1293 // Add the original argument and attributes.
1294 NewArgs.push_back(*I);
1295 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1296 if (Attr.hasAttributes(Idx)) {
1297 AttrBuilder B(Attr, Idx);
1298 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1299 Idx + (Idx >= NestIdx), B));
1306 // Add any function attributes.
1307 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1308 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1309 Attrs.getFnAttributes()));
1311 // The trampoline may have been bitcast to a bogus type (FTy).
1312 // Handle this by synthesizing a new function type, equal to FTy
1313 // with the chain parameter inserted.
1315 std::vector<Type*> NewTypes;
1316 NewTypes.reserve(FTy->getNumParams()+1);
1318 // Insert the chain's type into the list of parameter types, which may
1319 // mean appending it.
1322 FunctionType::param_iterator I = FTy->param_begin(),
1323 E = FTy->param_end();
1327 // Add the chain's type.
1328 NewTypes.push_back(NestTy);
1333 // Add the original type.
1334 NewTypes.push_back(*I);
1340 // Replace the trampoline call with a direct call. Let the generic
1341 // code sort out any function type mismatches.
1342 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1344 Constant *NewCallee =
1345 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1346 NestF : ConstantExpr::getBitCast(NestF,
1347 PointerType::getUnqual(NewFTy));
1348 const AttributeSet &NewPAL =
1349 AttributeSet::get(FTy->getContext(), NewAttrs);
1351 Instruction *NewCaller;
1352 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1353 NewCaller = InvokeInst::Create(NewCallee,
1354 II->getNormalDest(), II->getUnwindDest(),
1356 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1357 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1359 NewCaller = CallInst::Create(NewCallee, NewArgs);
1360 if (cast<CallInst>(Caller)->isTailCall())
1361 cast<CallInst>(NewCaller)->setTailCall();
1362 cast<CallInst>(NewCaller)->
1363 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1364 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1371 // Replace the trampoline call with a direct call. Since there is no 'nest'
1372 // parameter, there is no need to adjust the argument list. Let the generic
1373 // code sort out any function type mismatches.
1374 Constant *NewCallee =
1375 NestF->getType() == PTy ? NestF :
1376 ConstantExpr::getBitCast(NestF, PTy);
1377 CS.setCalledFunction(NewCallee);
1378 return CS.getInstruction();