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/DataLayout.h"
18 #include "llvm/Support/CallSite.h"
19 #include "llvm/Support/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), TD);
60 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
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 (TD && SrcETy->isSized() && TD->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(), TD);
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, TD, 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, TD) >= 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, TD) >= 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, TD) >= 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), TD);
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 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth();
670 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) {
671 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
672 VectorType* VT = cast<VectorType>(CV0->getType());
673 SmallVector<Constant*, 4> NewElems;
674 for (unsigned i = 0; i < VT->getNumElements(); ++i) {
676 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue();
677 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth);
679 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue();
680 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth);
682 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E));
684 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems));
687 // Couldn't simplify - canonicalize constant to the RHS.
688 std::swap(Arg0, Arg1);
691 // Handle mul by one:
692 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
693 if (ConstantInt *Splat =
694 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) {
695 if (Splat->isOne()) {
697 return CastInst::CreateZExtOrBitCast(Arg0, II->getType());
699 return CastInst::CreateSExtOrBitCast(Arg0, II->getType());
707 case Intrinsic::stackrestore: {
708 // If the save is right next to the restore, remove the restore. This can
709 // happen when variable allocas are DCE'd.
710 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
711 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
712 BasicBlock::iterator BI = SS;
714 return EraseInstFromFunction(CI);
718 // Scan down this block to see if there is another stack restore in the
719 // same block without an intervening call/alloca.
720 BasicBlock::iterator BI = II;
721 TerminatorInst *TI = II->getParent()->getTerminator();
722 bool CannotRemove = false;
723 for (++BI; &*BI != TI; ++BI) {
724 if (isa<AllocaInst>(BI)) {
728 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
729 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
730 // If there is a stackrestore below this one, remove this one.
731 if (II->getIntrinsicID() == Intrinsic::stackrestore)
732 return EraseInstFromFunction(CI);
733 // Otherwise, ignore the intrinsic.
735 // If we found a non-intrinsic call, we can't remove the stack
743 // If the stack restore is in a return, resume, or unwind block and if there
744 // are no allocas or calls between the restore and the return, nuke the
746 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
747 return EraseInstFromFunction(CI);
752 return visitCallSite(II);
755 // InvokeInst simplification
757 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
758 return visitCallSite(&II);
761 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
762 /// passed through the varargs area, we can eliminate the use of the cast.
763 static bool isSafeToEliminateVarargsCast(const CallSite CS,
764 const CastInst * const CI,
765 const DataLayout * const TD,
767 if (!CI->isLosslessCast())
770 // The size of ByVal or InAlloca arguments is derived from the type, so we
771 // can't change to a type with a different size. If the size were
772 // passed explicitly we could avoid this check.
773 if (!CS.isByValOrInAllocaArgument(ix))
777 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
778 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
779 if (!SrcTy->isSized() || !DstTy->isSized())
781 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
786 // Try to fold some different type of calls here.
787 // Currently we're only working with the checking functions, memcpy_chk,
788 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
789 // strcat_chk and strncat_chk.
790 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *TD) {
791 if (CI->getCalledFunction() == 0) return 0;
793 if (Value *With = Simplifier->optimizeCall(CI)) {
795 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
801 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
802 // Strip off at most one level of pointer casts, looking for an alloca. This
803 // is good enough in practice and simpler than handling any number of casts.
804 Value *Underlying = TrampMem->stripPointerCasts();
805 if (Underlying != TrampMem &&
806 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
808 if (!isa<AllocaInst>(Underlying))
811 IntrinsicInst *InitTrampoline = 0;
812 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
814 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
817 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
819 // More than one init_trampoline writes to this value. Give up.
824 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
825 // Allow any number of calls to adjust.trampoline.
830 // No call to init.trampoline found.
834 // Check that the alloca is being used in the expected way.
835 if (InitTrampoline->getOperand(0) != TrampMem)
838 return InitTrampoline;
841 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
843 // Visit all the previous instructions in the basic block, and try to find a
844 // init.trampoline which has a direct path to the adjust.trampoline.
845 for (BasicBlock::iterator I = AdjustTramp,
846 E = AdjustTramp->getParent()->begin(); I != E; ) {
847 Instruction *Inst = --I;
848 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
849 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
850 II->getOperand(0) == TrampMem)
852 if (Inst->mayWriteToMemory())
858 // Given a call to llvm.adjust.trampoline, find and return the corresponding
859 // call to llvm.init.trampoline if the call to the trampoline can be optimized
860 // to a direct call to a function. Otherwise return NULL.
862 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
863 Callee = Callee->stripPointerCasts();
864 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
866 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
869 Value *TrampMem = AdjustTramp->getOperand(0);
871 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
873 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
878 // visitCallSite - Improvements for call and invoke instructions.
880 Instruction *InstCombiner::visitCallSite(CallSite CS) {
881 if (isAllocLikeFn(CS.getInstruction(), TLI))
882 return visitAllocSite(*CS.getInstruction());
884 bool Changed = false;
886 // If the callee is a pointer to a function, attempt to move any casts to the
887 // arguments of the call/invoke.
888 Value *Callee = CS.getCalledValue();
889 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
892 if (Function *CalleeF = dyn_cast<Function>(Callee))
893 // If the call and callee calling conventions don't match, this call must
894 // be unreachable, as the call is undefined.
895 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
896 // Only do this for calls to a function with a body. A prototype may
897 // not actually end up matching the implementation's calling conv for a
898 // variety of reasons (e.g. it may be written in assembly).
899 !CalleeF->isDeclaration()) {
900 Instruction *OldCall = CS.getInstruction();
901 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
902 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
904 // If OldCall does not return void then replaceAllUsesWith undef.
905 // This allows ValueHandlers and custom metadata to adjust itself.
906 if (!OldCall->getType()->isVoidTy())
907 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
908 if (isa<CallInst>(OldCall))
909 return EraseInstFromFunction(*OldCall);
911 // We cannot remove an invoke, because it would change the CFG, just
912 // change the callee to a null pointer.
913 cast<InvokeInst>(OldCall)->setCalledFunction(
914 Constant::getNullValue(CalleeF->getType()));
918 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
919 // If CS does not return void then replaceAllUsesWith undef.
920 // This allows ValueHandlers and custom metadata to adjust itself.
921 if (!CS.getInstruction()->getType()->isVoidTy())
922 ReplaceInstUsesWith(*CS.getInstruction(),
923 UndefValue::get(CS.getInstruction()->getType()));
925 if (isa<InvokeInst>(CS.getInstruction())) {
926 // Can't remove an invoke because we cannot change the CFG.
930 // This instruction is not reachable, just remove it. We insert a store to
931 // undef so that we know that this code is not reachable, despite the fact
932 // that we can't modify the CFG here.
933 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
934 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
935 CS.getInstruction());
937 return EraseInstFromFunction(*CS.getInstruction());
940 if (IntrinsicInst *II = FindInitTrampoline(Callee))
941 return transformCallThroughTrampoline(CS, II);
943 PointerType *PTy = cast<PointerType>(Callee->getType());
944 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
945 if (FTy->isVarArg()) {
946 int ix = FTy->getNumParams();
947 // See if we can optimize any arguments passed through the varargs area of
949 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
950 E = CS.arg_end(); I != E; ++I, ++ix) {
951 CastInst *CI = dyn_cast<CastInst>(*I);
952 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
953 *I = CI->getOperand(0);
959 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
960 // Inline asm calls cannot throw - mark them 'nounwind'.
961 CS.setDoesNotThrow();
965 // Try to optimize the call if possible, we require DataLayout for most of
966 // this. None of these calls are seen as possibly dead so go ahead and
967 // delete the instruction now.
968 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
969 Instruction *I = tryOptimizeCall(CI, TD);
970 // If we changed something return the result, etc. Otherwise let
971 // the fallthrough check.
972 if (I) return EraseInstFromFunction(*I);
975 return Changed ? CS.getInstruction() : 0;
978 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
979 // attempt to move the cast to the arguments of the call/invoke.
981 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
983 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
986 Instruction *Caller = CS.getInstruction();
987 const AttributeSet &CallerPAL = CS.getAttributes();
989 // Okay, this is a cast from a function to a different type. Unless doing so
990 // would cause a type conversion of one of our arguments, change this call to
991 // be a direct call with arguments casted to the appropriate types.
993 FunctionType *FT = Callee->getFunctionType();
994 Type *OldRetTy = Caller->getType();
995 Type *NewRetTy = FT->getReturnType();
997 // Check to see if we are changing the return type...
998 if (OldRetTy != NewRetTy) {
1000 if (NewRetTy->isStructTy())
1001 return false; // TODO: Handle multiple return values.
1003 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1004 if (Callee->isDeclaration())
1005 return false; // Cannot transform this return value.
1007 if (!Caller->use_empty() &&
1008 // void -> non-void is handled specially
1009 !NewRetTy->isVoidTy())
1010 return false; // Cannot transform this return value.
1013 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1014 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1016 hasAttributes(AttributeFuncs::
1017 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1018 AttributeSet::ReturnIndex))
1019 return false; // Attribute not compatible with transformed value.
1022 // If the callsite is an invoke instruction, and the return value is used by
1023 // a PHI node in a successor, we cannot change the return type of the call
1024 // because there is no place to put the cast instruction (without breaking
1025 // the critical edge). Bail out in this case.
1026 if (!Caller->use_empty())
1027 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1028 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1030 if (PHINode *PN = dyn_cast<PHINode>(*UI))
1031 if (PN->getParent() == II->getNormalDest() ||
1032 PN->getParent() == II->getUnwindDest())
1036 unsigned NumActualArgs = CS.arg_size();
1037 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1039 CallSite::arg_iterator AI = CS.arg_begin();
1040 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1041 Type *ParamTy = FT->getParamType(i);
1042 Type *ActTy = (*AI)->getType();
1044 if (!CastInst::isBitCastable(ActTy, ParamTy))
1045 return false; // Cannot transform this parameter value.
1047 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1048 hasAttributes(AttributeFuncs::
1049 typeIncompatible(ParamTy, i + 1), i + 1))
1050 return false; // Attribute not compatible with transformed value.
1052 if (CS.isInAllocaArgument(i))
1053 return false; // Cannot transform to and from inalloca.
1055 // If the parameter is passed as a byval argument, then we have to have a
1056 // sized type and the sized type has to have the same size as the old type.
1057 if (ParamTy != ActTy &&
1058 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1059 Attribute::ByVal)) {
1060 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1061 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1064 Type *CurElTy = ActTy->getPointerElementType();
1065 if (TD->getTypeAllocSize(CurElTy) !=
1066 TD->getTypeAllocSize(ParamPTy->getElementType()))
1071 if (Callee->isDeclaration()) {
1072 // Do not delete arguments unless we have a function body.
1073 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1076 // If the callee is just a declaration, don't change the varargsness of the
1077 // call. We don't want to introduce a varargs call where one doesn't
1079 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1080 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1083 // If both the callee and the cast type are varargs, we still have to make
1084 // sure the number of fixed parameters are the same or we have the same
1085 // ABI issues as if we introduce a varargs call.
1086 if (FT->isVarArg() &&
1087 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1088 FT->getNumParams() !=
1089 cast<FunctionType>(APTy->getElementType())->getNumParams())
1093 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1094 !CallerPAL.isEmpty())
1095 // In this case we have more arguments than the new function type, but we
1096 // won't be dropping them. Check that these extra arguments have attributes
1097 // that are compatible with being a vararg call argument.
1098 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1099 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1100 if (Index <= FT->getNumParams())
1103 // Check if it has an attribute that's incompatible with varargs.
1104 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1105 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1110 // Okay, we decided that this is a safe thing to do: go ahead and start
1111 // inserting cast instructions as necessary.
1112 std::vector<Value*> Args;
1113 Args.reserve(NumActualArgs);
1114 SmallVector<AttributeSet, 8> attrVec;
1115 attrVec.reserve(NumCommonArgs);
1117 // Get any return attributes.
1118 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1120 // If the return value is not being used, the type may not be compatible
1121 // with the existing attributes. Wipe out any problematic attributes.
1123 removeAttributes(AttributeFuncs::
1124 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1125 AttributeSet::ReturnIndex);
1127 // Add the new return attributes.
1128 if (RAttrs.hasAttributes())
1129 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1130 AttributeSet::ReturnIndex, RAttrs));
1132 AI = CS.arg_begin();
1133 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1134 Type *ParamTy = FT->getParamType(i);
1136 if ((*AI)->getType() == ParamTy) {
1137 Args.push_back(*AI);
1139 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1142 // Add any parameter attributes.
1143 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1144 if (PAttrs.hasAttributes())
1145 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1149 // If the function takes more arguments than the call was taking, add them
1151 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1152 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1154 // If we are removing arguments to the function, emit an obnoxious warning.
1155 if (FT->getNumParams() < NumActualArgs) {
1156 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1157 if (FT->isVarArg()) {
1158 // Add all of the arguments in their promoted form to the arg list.
1159 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1160 Type *PTy = getPromotedType((*AI)->getType());
1161 if (PTy != (*AI)->getType()) {
1162 // Must promote to pass through va_arg area!
1163 Instruction::CastOps opcode =
1164 CastInst::getCastOpcode(*AI, false, PTy, false);
1165 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1167 Args.push_back(*AI);
1170 // Add any parameter attributes.
1171 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1172 if (PAttrs.hasAttributes())
1173 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1179 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1180 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1181 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1183 if (NewRetTy->isVoidTy())
1184 Caller->setName(""); // Void type should not have a name.
1186 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1190 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1191 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1192 II->getUnwindDest(), Args);
1194 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1195 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1197 CallInst *CI = cast<CallInst>(Caller);
1198 NC = Builder->CreateCall(Callee, Args);
1200 if (CI->isTailCall())
1201 cast<CallInst>(NC)->setTailCall();
1202 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1203 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1206 // Insert a cast of the return type as necessary.
1208 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1209 if (!NV->getType()->isVoidTy()) {
1210 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1211 NC->setDebugLoc(Caller->getDebugLoc());
1213 // If this is an invoke instruction, we should insert it after the first
1214 // non-phi, instruction in the normal successor block.
1215 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1216 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1217 InsertNewInstBefore(NC, *I);
1219 // Otherwise, it's a call, just insert cast right after the call.
1220 InsertNewInstBefore(NC, *Caller);
1222 Worklist.AddUsersToWorkList(*Caller);
1224 NV = UndefValue::get(Caller->getType());
1228 if (!Caller->use_empty())
1229 ReplaceInstUsesWith(*Caller, NV);
1231 EraseInstFromFunction(*Caller);
1235 // transformCallThroughTrampoline - Turn a call to a function created by
1236 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1237 // underlying function.
1240 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1241 IntrinsicInst *Tramp) {
1242 Value *Callee = CS.getCalledValue();
1243 PointerType *PTy = cast<PointerType>(Callee->getType());
1244 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1245 const AttributeSet &Attrs = CS.getAttributes();
1247 // If the call already has the 'nest' attribute somewhere then give up -
1248 // otherwise 'nest' would occur twice after splicing in the chain.
1249 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1253 "transformCallThroughTrampoline called with incorrect CallSite.");
1255 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1256 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1257 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1259 const AttributeSet &NestAttrs = NestF->getAttributes();
1260 if (!NestAttrs.isEmpty()) {
1261 unsigned NestIdx = 1;
1263 AttributeSet NestAttr;
1265 // Look for a parameter marked with the 'nest' attribute.
1266 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1267 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1268 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1269 // Record the parameter type and any other attributes.
1271 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1276 Instruction *Caller = CS.getInstruction();
1277 std::vector<Value*> NewArgs;
1278 NewArgs.reserve(CS.arg_size() + 1);
1280 SmallVector<AttributeSet, 8> NewAttrs;
1281 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1283 // Insert the nest argument into the call argument list, which may
1284 // mean appending it. Likewise for attributes.
1286 // Add any result attributes.
1287 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1288 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1289 Attrs.getRetAttributes()));
1293 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1295 if (Idx == NestIdx) {
1296 // Add the chain argument and attributes.
1297 Value *NestVal = Tramp->getArgOperand(2);
1298 if (NestVal->getType() != NestTy)
1299 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1300 NewArgs.push_back(NestVal);
1301 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1308 // Add the original argument and attributes.
1309 NewArgs.push_back(*I);
1310 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1311 if (Attr.hasAttributes(Idx)) {
1312 AttrBuilder B(Attr, Idx);
1313 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1314 Idx + (Idx >= NestIdx), B));
1321 // Add any function attributes.
1322 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1323 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1324 Attrs.getFnAttributes()));
1326 // The trampoline may have been bitcast to a bogus type (FTy).
1327 // Handle this by synthesizing a new function type, equal to FTy
1328 // with the chain parameter inserted.
1330 std::vector<Type*> NewTypes;
1331 NewTypes.reserve(FTy->getNumParams()+1);
1333 // Insert the chain's type into the list of parameter types, which may
1334 // mean appending it.
1337 FunctionType::param_iterator I = FTy->param_begin(),
1338 E = FTy->param_end();
1342 // Add the chain's type.
1343 NewTypes.push_back(NestTy);
1348 // Add the original type.
1349 NewTypes.push_back(*I);
1355 // Replace the trampoline call with a direct call. Let the generic
1356 // code sort out any function type mismatches.
1357 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1359 Constant *NewCallee =
1360 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1361 NestF : ConstantExpr::getBitCast(NestF,
1362 PointerType::getUnqual(NewFTy));
1363 const AttributeSet &NewPAL =
1364 AttributeSet::get(FTy->getContext(), NewAttrs);
1366 Instruction *NewCaller;
1367 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1368 NewCaller = InvokeInst::Create(NewCallee,
1369 II->getNormalDest(), II->getUnwindDest(),
1371 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1372 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1374 NewCaller = CallInst::Create(NewCallee, NewArgs);
1375 if (cast<CallInst>(Caller)->isTailCall())
1376 cast<CallInst>(NewCaller)->setTailCall();
1377 cast<CallInst>(NewCaller)->
1378 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1379 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1386 // Replace the trampoline call with a direct call. Since there is no 'nest'
1387 // parameter, there is no need to adjust the argument list. Let the generic
1388 // code sort out any function type mismatches.
1389 Constant *NewCallee =
1390 NestF->getType() == PTy ? NestF :
1391 ConstantExpr::getBitCast(NestF, PTy);
1392 CS.setCalledFunction(NewCallee);
1393 return CS.getInstruction();