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/Dominators.h"
20 #include "llvm/IR/PatternMatch.h"
21 #include "llvm/IR/Statepoint.h"
22 #include "llvm/Transforms/Utils/BuildLibCalls.h"
23 #include "llvm/Transforms/Utils/Local.h"
25 using namespace PatternMatch;
27 #define DEBUG_TYPE "instcombine"
29 STATISTIC(NumSimplified, "Number of library calls simplified");
31 /// getPromotedType - Return the specified type promoted as it would be to pass
32 /// though a va_arg area.
33 static Type *getPromotedType(Type *Ty) {
34 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
35 if (ITy->getBitWidth() < 32)
36 return Type::getInt32Ty(Ty->getContext());
41 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
42 /// single scalar element, like {{{type}}} or [1 x type], return type.
43 static Type *reduceToSingleValueType(Type *T) {
44 while (!T->isSingleValueType()) {
45 if (StructType *STy = dyn_cast<StructType>(T)) {
46 if (STy->getNumElements() == 1)
47 T = STy->getElementType(0);
50 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
51 if (ATy->getNumElements() == 1)
52 T = ATy->getElementType();
62 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
63 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, AC, MI, DT);
64 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, AC, MI, DT);
65 unsigned MinAlign = std::min(DstAlign, SrcAlign);
66 unsigned CopyAlign = MI->getAlignment();
68 if (CopyAlign < MinAlign) {
69 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
74 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
76 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
77 if (!MemOpLength) return nullptr;
79 // Source and destination pointer types are always "i8*" for intrinsic. See
80 // if the size is something we can handle with a single primitive load/store.
81 // A single load+store correctly handles overlapping memory in the memmove
83 uint64_t Size = MemOpLength->getLimitedValue();
84 assert(Size && "0-sized memory transferring should be removed already.");
86 if (Size > 8 || (Size&(Size-1)))
87 return nullptr; // If not 1/2/4/8 bytes, exit.
89 // Use an integer load+store unless we can find something better.
91 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
93 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
95 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
96 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
97 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
99 // Memcpy forces the use of i8* for the source and destination. That means
100 // that if you're using memcpy to move one double around, you'll get a cast
101 // from double* to i8*. We'd much rather use a double load+store rather than
102 // an i64 load+store, here because this improves the odds that the source or
103 // dest address will be promotable. See if we can find a better type than the
105 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
106 MDNode *CopyMD = nullptr;
107 if (StrippedDest != MI->getArgOperand(0)) {
108 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
110 if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
111 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
112 // down through these levels if so.
113 SrcETy = reduceToSingleValueType(SrcETy);
115 if (SrcETy->isSingleValueType()) {
116 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
117 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
119 // If the memcpy has metadata describing the members, see if we can
120 // get the TBAA tag describing our copy.
121 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
122 if (M->getNumOperands() == 3 && M->getOperand(0) &&
123 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
124 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() &&
126 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
127 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
129 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
130 CopyMD = cast<MDNode>(M->getOperand(2));
136 // If the memcpy/memmove provides better alignment info than we can
138 SrcAlign = std::max(SrcAlign, CopyAlign);
139 DstAlign = std::max(DstAlign, CopyAlign);
141 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
142 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
143 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
144 L->setAlignment(SrcAlign);
146 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
147 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
148 S->setAlignment(DstAlign);
150 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
152 // Set the size of the copy to 0, it will be deleted on the next iteration.
153 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
157 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
158 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, AC, MI, DT);
159 if (MI->getAlignment() < Alignment) {
160 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
165 // Extract the length and alignment and fill if they are constant.
166 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
167 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
168 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
170 uint64_t Len = LenC->getLimitedValue();
171 Alignment = MI->getAlignment();
172 assert(Len && "0-sized memory setting should be removed already.");
174 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
175 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
176 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
178 Value *Dest = MI->getDest();
179 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
180 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
181 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
183 // Alignment 0 is identity for alignment 1 for memset, but not store.
184 if (Alignment == 0) Alignment = 1;
186 // Extract the fill value and store.
187 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
188 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
190 S->setAlignment(Alignment);
192 // Set the size of the copy to 0, it will be deleted on the next iteration.
193 MI->setLength(Constant::getNullValue(LenC->getType()));
200 /// visitCallInst - CallInst simplification. This mostly only handles folding
201 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
202 /// the heavy lifting.
204 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
205 if (isFreeCall(&CI, TLI))
206 return visitFree(CI);
208 // If the caller function is nounwind, mark the call as nounwind, even if the
210 if (CI.getParent()->getParent()->doesNotThrow() &&
211 !CI.doesNotThrow()) {
212 CI.setDoesNotThrow();
216 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
217 if (!II) return visitCallSite(&CI);
219 // Intrinsics cannot occur in an invoke, so handle them here instead of in
221 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
222 bool Changed = false;
224 // memmove/cpy/set of zero bytes is a noop.
225 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
226 if (NumBytes->isNullValue())
227 return EraseInstFromFunction(CI);
229 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
230 if (CI->getZExtValue() == 1) {
231 // Replace the instruction with just byte operations. We would
232 // transform other cases to loads/stores, but we don't know if
233 // alignment is sufficient.
237 // No other transformations apply to volatile transfers.
238 if (MI->isVolatile())
241 // If we have a memmove and the source operation is a constant global,
242 // then the source and dest pointers can't alias, so we can change this
243 // into a call to memcpy.
244 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
245 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
246 if (GVSrc->isConstant()) {
247 Module *M = CI.getParent()->getParent()->getParent();
248 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
249 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
250 CI.getArgOperand(1)->getType(),
251 CI.getArgOperand(2)->getType() };
252 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
257 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
258 // memmove(x,x,size) -> noop.
259 if (MTI->getSource() == MTI->getDest())
260 return EraseInstFromFunction(CI);
263 // If we can determine a pointer alignment that is bigger than currently
264 // set, update the alignment.
265 if (isa<MemTransferInst>(MI)) {
266 if (Instruction *I = SimplifyMemTransfer(MI))
268 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
269 if (Instruction *I = SimplifyMemSet(MSI))
273 if (Changed) return II;
276 switch (II->getIntrinsicID()) {
278 case Intrinsic::objectsize: {
280 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
281 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
284 case Intrinsic::bswap: {
285 Value *IIOperand = II->getArgOperand(0);
288 // bswap(bswap(x)) -> x
289 if (match(IIOperand, m_BSwap(m_Value(X))))
290 return ReplaceInstUsesWith(CI, X);
292 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
293 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
294 unsigned C = X->getType()->getPrimitiveSizeInBits() -
295 IIOperand->getType()->getPrimitiveSizeInBits();
296 Value *CV = ConstantInt::get(X->getType(), C);
297 Value *V = Builder->CreateLShr(X, CV);
298 return new TruncInst(V, IIOperand->getType());
303 case Intrinsic::powi:
304 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
307 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
310 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
311 // powi(x, -1) -> 1/x
312 if (Power->isAllOnesValue())
313 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
314 II->getArgOperand(0));
317 case Intrinsic::cttz: {
318 // If all bits below the first known one are known zero,
319 // this value is constant.
320 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
321 // FIXME: Try to simplify vectors of integers.
323 uint32_t BitWidth = IT->getBitWidth();
324 APInt KnownZero(BitWidth, 0);
325 APInt KnownOne(BitWidth, 0);
326 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
327 unsigned TrailingZeros = KnownOne.countTrailingZeros();
328 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
329 if ((Mask & KnownZero) == Mask)
330 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
331 APInt(BitWidth, TrailingZeros)));
335 case Intrinsic::ctlz: {
336 // If all bits above the first known one are known zero,
337 // this value is constant.
338 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
339 // FIXME: Try to simplify vectors of integers.
341 uint32_t BitWidth = IT->getBitWidth();
342 APInt KnownZero(BitWidth, 0);
343 APInt KnownOne(BitWidth, 0);
344 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
345 unsigned LeadingZeros = KnownOne.countLeadingZeros();
346 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
347 if ((Mask & KnownZero) == Mask)
348 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
349 APInt(BitWidth, LeadingZeros)));
353 case Intrinsic::uadd_with_overflow: {
354 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
355 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
356 uint32_t BitWidth = IT->getBitWidth();
357 APInt LHSKnownZero(BitWidth, 0);
358 APInt LHSKnownOne(BitWidth, 0);
359 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, II);
360 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
361 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
363 if (LHSKnownNegative || LHSKnownPositive) {
364 APInt RHSKnownZero(BitWidth, 0);
365 APInt RHSKnownOne(BitWidth, 0);
366 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, II);
367 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
368 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
369 if (LHSKnownNegative && RHSKnownNegative) {
370 // The sign bit is set in both cases: this MUST overflow.
371 // Create a simple add instruction, and insert it into the struct.
372 return CreateOverflowTuple(II, Builder->CreateAdd(LHS, RHS), true,
376 if (LHSKnownPositive && RHSKnownPositive) {
377 // The sign bit is clear in both cases: this CANNOT overflow.
378 // Create a simple add instruction, and insert it into the struct.
379 return CreateOverflowTuple(II, Builder->CreateNUWAdd(LHS, RHS), false);
383 // FALL THROUGH uadd into sadd
384 case Intrinsic::sadd_with_overflow:
385 // Canonicalize constants into the RHS.
386 if (isa<Constant>(II->getArgOperand(0)) &&
387 !isa<Constant>(II->getArgOperand(1))) {
388 Value *LHS = II->getArgOperand(0);
389 II->setArgOperand(0, II->getArgOperand(1));
390 II->setArgOperand(1, LHS);
394 // X + undef -> undef
395 if (isa<UndefValue>(II->getArgOperand(1)))
396 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
398 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
399 // X + 0 -> {X, false}
401 return CreateOverflowTuple(II, II->getArgOperand(0), false,
406 // We can strength reduce reduce this signed add into a regular add if we
407 // can prove that it will never overflow.
408 if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow) {
409 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
410 if (WillNotOverflowSignedAdd(LHS, RHS, II)) {
411 return CreateOverflowTuple(II, Builder->CreateNSWAdd(LHS, RHS), false);
416 case Intrinsic::usub_with_overflow:
417 case Intrinsic::ssub_with_overflow: {
418 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
419 // undef - X -> undef
420 // X - undef -> undef
421 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
422 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
424 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
425 // X - 0 -> {X, false}
426 if (ConstRHS->isZero()) {
427 return CreateOverflowTuple(II, LHS, false, /*ReUseName*/false);
430 if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
431 if (WillNotOverflowSignedSub(LHS, RHS, II)) {
432 return CreateOverflowTuple(II, Builder->CreateNSWSub(LHS, RHS), false);
435 if (WillNotOverflowUnsignedSub(LHS, RHS, II)) {
436 return CreateOverflowTuple(II, Builder->CreateNUWSub(LHS, RHS), false);
441 case Intrinsic::umul_with_overflow: {
442 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
443 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, II);
444 if (OR == OverflowResult::NeverOverflows) {
445 return CreateOverflowTuple(II, Builder->CreateNUWMul(LHS, RHS), false);
446 } else if (OR == OverflowResult::AlwaysOverflows) {
447 return CreateOverflowTuple(II, Builder->CreateMul(LHS, RHS), true);
450 case Intrinsic::smul_with_overflow:
451 // Canonicalize constants into the RHS.
452 if (isa<Constant>(II->getArgOperand(0)) &&
453 !isa<Constant>(II->getArgOperand(1))) {
454 Value *LHS = II->getArgOperand(0);
455 II->setArgOperand(0, II->getArgOperand(1));
456 II->setArgOperand(1, LHS);
460 // X * undef -> undef
461 if (isa<UndefValue>(II->getArgOperand(1)))
462 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
464 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
467 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
469 // X * 1 -> {X, false}
470 if (RHSI->equalsInt(1)) {
471 return CreateOverflowTuple(II, II->getArgOperand(0), false,
475 if (II->getIntrinsicID() == Intrinsic::smul_with_overflow) {
476 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
477 if (WillNotOverflowSignedMul(LHS, RHS, II)) {
478 return CreateOverflowTuple(II, Builder->CreateNSWMul(LHS, RHS), false);
482 case Intrinsic::minnum:
483 case Intrinsic::maxnum: {
484 Value *Arg0 = II->getArgOperand(0);
485 Value *Arg1 = II->getArgOperand(1);
489 return ReplaceInstUsesWith(CI, Arg0);
491 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
492 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
494 // Canonicalize constants into the RHS.
496 II->setArgOperand(0, Arg1);
497 II->setArgOperand(1, Arg0);
502 if (C1 && C1->isNaN())
503 return ReplaceInstUsesWith(CI, Arg0);
505 // This is the value because if undef were NaN, we would return the other
506 // value and cannot return a NaN unless both operands are.
508 // fmin(undef, x) -> x
509 if (isa<UndefValue>(Arg0))
510 return ReplaceInstUsesWith(CI, Arg1);
512 // fmin(x, undef) -> x
513 if (isa<UndefValue>(Arg1))
514 return ReplaceInstUsesWith(CI, Arg0);
518 if (II->getIntrinsicID() == Intrinsic::minnum) {
519 // fmin(x, fmin(x, y)) -> fmin(x, y)
520 // fmin(y, fmin(x, y)) -> fmin(x, y)
521 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
522 if (Arg0 == X || Arg0 == Y)
523 return ReplaceInstUsesWith(CI, Arg1);
526 // fmin(fmin(x, y), x) -> fmin(x, y)
527 // fmin(fmin(x, y), y) -> fmin(x, y)
528 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
529 if (Arg1 == X || Arg1 == Y)
530 return ReplaceInstUsesWith(CI, Arg0);
533 // TODO: fmin(nnan x, inf) -> x
534 // TODO: fmin(nnan ninf x, flt_max) -> x
535 if (C1 && C1->isInfinity()) {
536 // fmin(x, -inf) -> -inf
537 if (C1->isNegative())
538 return ReplaceInstUsesWith(CI, Arg1);
541 assert(II->getIntrinsicID() == Intrinsic::maxnum);
542 // fmax(x, fmax(x, y)) -> fmax(x, y)
543 // fmax(y, fmax(x, y)) -> fmax(x, y)
544 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
545 if (Arg0 == X || Arg0 == Y)
546 return ReplaceInstUsesWith(CI, Arg1);
549 // fmax(fmax(x, y), x) -> fmax(x, y)
550 // fmax(fmax(x, y), y) -> fmax(x, y)
551 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
552 if (Arg1 == X || Arg1 == Y)
553 return ReplaceInstUsesWith(CI, Arg0);
556 // TODO: fmax(nnan x, -inf) -> x
557 // TODO: fmax(nnan ninf x, -flt_max) -> x
558 if (C1 && C1->isInfinity()) {
559 // fmax(x, inf) -> inf
560 if (!C1->isNegative())
561 return ReplaceInstUsesWith(CI, Arg1);
566 case Intrinsic::ppc_altivec_lvx:
567 case Intrinsic::ppc_altivec_lvxl:
568 // Turn PPC lvx -> load if the pointer is known aligned.
569 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, AC, II, DT) >=
571 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
572 PointerType::getUnqual(II->getType()));
573 return new LoadInst(Ptr);
576 case Intrinsic::ppc_vsx_lxvw4x:
577 case Intrinsic::ppc_vsx_lxvd2x: {
578 // Turn PPC VSX loads into normal loads.
579 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
580 PointerType::getUnqual(II->getType()));
581 return new LoadInst(Ptr, Twine(""), false, 1);
583 case Intrinsic::ppc_altivec_stvx:
584 case Intrinsic::ppc_altivec_stvxl:
585 // Turn stvx -> store if the pointer is known aligned.
586 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, AC, II, DT) >=
589 PointerType::getUnqual(II->getArgOperand(0)->getType());
590 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
591 return new StoreInst(II->getArgOperand(0), Ptr);
594 case Intrinsic::ppc_vsx_stxvw4x:
595 case Intrinsic::ppc_vsx_stxvd2x: {
596 // Turn PPC VSX stores into normal stores.
597 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
598 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
599 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
601 case Intrinsic::x86_sse_storeu_ps:
602 case Intrinsic::x86_sse2_storeu_pd:
603 case Intrinsic::x86_sse2_storeu_dq:
604 // Turn X86 storeu -> store if the pointer is known aligned.
605 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, AC, II, DT) >=
608 PointerType::getUnqual(II->getArgOperand(1)->getType());
609 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
610 return new StoreInst(II->getArgOperand(1), Ptr);
614 case Intrinsic::x86_sse_cvtss2si:
615 case Intrinsic::x86_sse_cvtss2si64:
616 case Intrinsic::x86_sse_cvttss2si:
617 case Intrinsic::x86_sse_cvttss2si64:
618 case Intrinsic::x86_sse2_cvtsd2si:
619 case Intrinsic::x86_sse2_cvtsd2si64:
620 case Intrinsic::x86_sse2_cvttsd2si:
621 case Intrinsic::x86_sse2_cvttsd2si64: {
622 // These intrinsics only demand the 0th element of their input vectors. If
623 // we can simplify the input based on that, do so now.
625 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
626 APInt DemandedElts(VWidth, 1);
627 APInt UndefElts(VWidth, 0);
628 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
629 DemandedElts, UndefElts)) {
630 II->setArgOperand(0, V);
636 // Constant fold <A x Bi> << Ci.
637 // FIXME: We don't handle _dq because it's a shift of an i128, but is
638 // represented in the IR as <2 x i64>. A per element shift is wrong.
639 case Intrinsic::x86_sse2_psll_d:
640 case Intrinsic::x86_sse2_psll_q:
641 case Intrinsic::x86_sse2_psll_w:
642 case Intrinsic::x86_sse2_pslli_d:
643 case Intrinsic::x86_sse2_pslli_q:
644 case Intrinsic::x86_sse2_pslli_w:
645 case Intrinsic::x86_avx2_psll_d:
646 case Intrinsic::x86_avx2_psll_q:
647 case Intrinsic::x86_avx2_psll_w:
648 case Intrinsic::x86_avx2_pslli_d:
649 case Intrinsic::x86_avx2_pslli_q:
650 case Intrinsic::x86_avx2_pslli_w:
651 case Intrinsic::x86_sse2_psrl_d:
652 case Intrinsic::x86_sse2_psrl_q:
653 case Intrinsic::x86_sse2_psrl_w:
654 case Intrinsic::x86_sse2_psrli_d:
655 case Intrinsic::x86_sse2_psrli_q:
656 case Intrinsic::x86_sse2_psrli_w:
657 case Intrinsic::x86_avx2_psrl_d:
658 case Intrinsic::x86_avx2_psrl_q:
659 case Intrinsic::x86_avx2_psrl_w:
660 case Intrinsic::x86_avx2_psrli_d:
661 case Intrinsic::x86_avx2_psrli_q:
662 case Intrinsic::x86_avx2_psrli_w: {
663 // Simplify if count is constant. To 0 if >= BitWidth,
664 // otherwise to shl/lshr.
665 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
666 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
671 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
675 auto Vec = II->getArgOperand(0);
676 auto VT = cast<VectorType>(Vec->getType());
677 if (Count->getZExtValue() >
678 VT->getElementType()->getPrimitiveSizeInBits() - 1)
679 return ReplaceInstUsesWith(
680 CI, ConstantAggregateZero::get(Vec->getType()));
682 bool isPackedShiftLeft = true;
683 switch (II->getIntrinsicID()) {
685 case Intrinsic::x86_sse2_psrl_d:
686 case Intrinsic::x86_sse2_psrl_q:
687 case Intrinsic::x86_sse2_psrl_w:
688 case Intrinsic::x86_sse2_psrli_d:
689 case Intrinsic::x86_sse2_psrli_q:
690 case Intrinsic::x86_sse2_psrli_w:
691 case Intrinsic::x86_avx2_psrl_d:
692 case Intrinsic::x86_avx2_psrl_q:
693 case Intrinsic::x86_avx2_psrl_w:
694 case Intrinsic::x86_avx2_psrli_d:
695 case Intrinsic::x86_avx2_psrli_q:
696 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
699 unsigned VWidth = VT->getNumElements();
700 // Get a constant vector of the same type as the first operand.
701 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
702 if (isPackedShiftLeft)
703 return BinaryOperator::CreateShl(Vec,
704 Builder->CreateVectorSplat(VWidth, VTCI));
706 return BinaryOperator::CreateLShr(Vec,
707 Builder->CreateVectorSplat(VWidth, VTCI));
710 case Intrinsic::x86_sse41_pmovsxbw:
711 case Intrinsic::x86_sse41_pmovsxwd:
712 case Intrinsic::x86_sse41_pmovsxdq:
713 case Intrinsic::x86_sse41_pmovzxbw:
714 case Intrinsic::x86_sse41_pmovzxwd:
715 case Intrinsic::x86_sse41_pmovzxdq: {
716 // pmov{s|z}x ignores the upper half of their input vectors.
718 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
719 unsigned LowHalfElts = VWidth / 2;
720 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
721 APInt UndefElts(VWidth, 0);
722 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
725 II->setArgOperand(0, TmpV);
731 case Intrinsic::x86_sse4a_insertqi: {
732 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
734 // TODO: eventually we should lower this intrinsic to IR
735 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
736 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
737 unsigned Index = CIStart->getZExtValue();
738 // From AMD documentation: "a value of zero in the field length is
739 // defined as length of 64".
740 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
742 // From AMD documentation: "If the sum of the bit index + length field
743 // is greater than 64, the results are undefined".
745 // Note that both field index and field length are 8-bit quantities.
746 // Since variables 'Index' and 'Length' are unsigned values
747 // obtained from zero-extending field index and field length
748 // respectively, their sum should never wrap around.
749 if ((Index + Length) > 64)
750 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
752 if (Length == 64 && Index == 0) {
753 Value *Vec = II->getArgOperand(1);
754 Value *Undef = UndefValue::get(Vec->getType());
755 const uint32_t Mask[] = { 0, 2 };
756 return ReplaceInstUsesWith(
758 Builder->CreateShuffleVector(
759 Vec, Undef, ConstantDataVector::get(
760 II->getContext(), makeArrayRef(Mask))));
762 } else if (auto Source =
763 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
764 if (Source->hasOneUse() &&
765 Source->getArgOperand(1) == II->getArgOperand(1)) {
766 // If the source of the insert has only one use and it's another
767 // insert (and they're both inserting from the same vector), try to
768 // bundle both together.
770 dyn_cast<ConstantInt>(Source->getArgOperand(2));
772 dyn_cast<ConstantInt>(Source->getArgOperand(3));
773 if (CISourceStart && CISourceWidth) {
774 unsigned Start = CIStart->getZExtValue();
775 unsigned Width = CIWidth->getZExtValue();
776 unsigned End = Start + Width;
777 unsigned SourceStart = CISourceStart->getZExtValue();
778 unsigned SourceWidth = CISourceWidth->getZExtValue();
779 unsigned SourceEnd = SourceStart + SourceWidth;
780 unsigned NewStart, NewWidth;
781 bool ShouldReplace = false;
782 if (Start <= SourceStart && SourceStart <= End) {
784 NewWidth = std::max(End, SourceEnd) - NewStart;
785 ShouldReplace = true;
786 } else if (SourceStart <= Start && Start <= SourceEnd) {
787 NewStart = SourceStart;
788 NewWidth = std::max(SourceEnd, End) - NewStart;
789 ShouldReplace = true;
793 Constant *ConstantWidth = ConstantInt::get(
794 II->getArgOperand(2)->getType(), NewWidth, false);
795 Constant *ConstantStart = ConstantInt::get(
796 II->getArgOperand(3)->getType(), NewStart, false);
797 Value *Args[4] = { Source->getArgOperand(0),
798 II->getArgOperand(1), ConstantWidth,
800 Module *M = CI.getParent()->getParent()->getParent();
802 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
803 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
813 case Intrinsic::x86_sse41_pblendvb:
814 case Intrinsic::x86_sse41_blendvps:
815 case Intrinsic::x86_sse41_blendvpd:
816 case Intrinsic::x86_avx_blendv_ps_256:
817 case Intrinsic::x86_avx_blendv_pd_256:
818 case Intrinsic::x86_avx2_pblendvb: {
819 // Convert blendv* to vector selects if the mask is constant.
820 // This optimization is convoluted because the intrinsic is defined as
821 // getting a vector of floats or doubles for the ps and pd versions.
822 // FIXME: That should be changed.
823 Value *Mask = II->getArgOperand(2);
824 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
825 auto Tyi1 = Builder->getInt1Ty();
826 auto SelectorType = cast<VectorType>(Mask->getType());
827 auto EltTy = SelectorType->getElementType();
828 unsigned Size = SelectorType->getNumElements();
832 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
833 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
834 "Wrong arguments for variable blend intrinsic");
835 SmallVector<Constant *, 32> Selectors;
836 for (unsigned I = 0; I < Size; ++I) {
837 // The intrinsics only read the top bit
840 Selector = C->getElementAsInteger(I);
842 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
843 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
845 auto NewSelector = ConstantVector::get(Selectors);
846 return SelectInst::Create(NewSelector, II->getArgOperand(1),
847 II->getArgOperand(0), "blendv");
853 case Intrinsic::x86_avx_vpermilvar_ps:
854 case Intrinsic::x86_avx_vpermilvar_ps_256:
855 case Intrinsic::x86_avx_vpermilvar_pd:
856 case Intrinsic::x86_avx_vpermilvar_pd_256: {
857 // Convert vpermil* to shufflevector if the mask is constant.
858 Value *V = II->getArgOperand(1);
859 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
860 assert(Size == 8 || Size == 4 || Size == 2);
862 if (auto C = dyn_cast<ConstantDataVector>(V)) {
863 // The intrinsics only read one or two bits, clear the rest.
864 for (unsigned I = 0; I < Size; ++I) {
865 uint32_t Index = C->getElementAsInteger(I) & 0x3;
866 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
867 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
871 } else if (isa<ConstantAggregateZero>(V)) {
872 for (unsigned I = 0; I < Size; ++I)
877 // The _256 variants are a bit trickier since the mask bits always index
878 // into the corresponding 128 half. In order to convert to a generic
879 // shuffle, we have to make that explicit.
880 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
881 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
882 for (unsigned I = Size / 2; I < Size; ++I)
883 Indexes[I] += Size / 2;
886 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
887 auto V1 = II->getArgOperand(0);
888 auto V2 = UndefValue::get(V1->getType());
889 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
890 return ReplaceInstUsesWith(CI, Shuffle);
893 case Intrinsic::ppc_altivec_vperm:
894 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
895 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
896 // a vectorshuffle for little endian, we must undo the transformation
897 // performed on vec_perm in altivec.h. That is, we must complement
898 // the permutation mask with respect to 31 and reverse the order of
900 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
901 assert(Mask->getType()->getVectorNumElements() == 16 &&
902 "Bad type for intrinsic!");
904 // Check that all of the elements are integer constants or undefs.
905 bool AllEltsOk = true;
906 for (unsigned i = 0; i != 16; ++i) {
907 Constant *Elt = Mask->getAggregateElement(i);
908 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
915 // Cast the input vectors to byte vectors.
916 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
918 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
920 Value *Result = UndefValue::get(Op0->getType());
922 // Only extract each element once.
923 Value *ExtractedElts[32];
924 memset(ExtractedElts, 0, sizeof(ExtractedElts));
926 for (unsigned i = 0; i != 16; ++i) {
927 if (isa<UndefValue>(Mask->getAggregateElement(i)))
930 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
931 Idx &= 31; // Match the hardware behavior.
932 if (DL && DL->isLittleEndian())
935 if (!ExtractedElts[Idx]) {
936 Value *Op0ToUse = (DL && DL->isLittleEndian()) ? Op1 : Op0;
937 Value *Op1ToUse = (DL && DL->isLittleEndian()) ? Op0 : Op1;
939 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
940 Builder->getInt32(Idx&15));
943 // Insert this value into the result vector.
944 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
945 Builder->getInt32(i));
947 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
952 case Intrinsic::arm_neon_vld1:
953 case Intrinsic::arm_neon_vld2:
954 case Intrinsic::arm_neon_vld3:
955 case Intrinsic::arm_neon_vld4:
956 case Intrinsic::arm_neon_vld2lane:
957 case Intrinsic::arm_neon_vld3lane:
958 case Intrinsic::arm_neon_vld4lane:
959 case Intrinsic::arm_neon_vst1:
960 case Intrinsic::arm_neon_vst2:
961 case Intrinsic::arm_neon_vst3:
962 case Intrinsic::arm_neon_vst4:
963 case Intrinsic::arm_neon_vst2lane:
964 case Intrinsic::arm_neon_vst3lane:
965 case Intrinsic::arm_neon_vst4lane: {
966 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, AC, II, DT);
967 unsigned AlignArg = II->getNumArgOperands() - 1;
968 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
969 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
970 II->setArgOperand(AlignArg,
971 ConstantInt::get(Type::getInt32Ty(II->getContext()),
978 case Intrinsic::arm_neon_vmulls:
979 case Intrinsic::arm_neon_vmullu:
980 case Intrinsic::aarch64_neon_smull:
981 case Intrinsic::aarch64_neon_umull: {
982 Value *Arg0 = II->getArgOperand(0);
983 Value *Arg1 = II->getArgOperand(1);
985 // Handle mul by zero first:
986 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
987 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
990 // Check for constant LHS & RHS - in this case we just simplify.
991 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
992 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
993 VectorType *NewVT = cast<VectorType>(II->getType());
994 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
995 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
996 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
997 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
999 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1002 // Couldn't simplify - canonicalize constant to the RHS.
1003 std::swap(Arg0, Arg1);
1006 // Handle mul by one:
1007 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1008 if (ConstantInt *Splat =
1009 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1011 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1012 /*isSigned=*/!Zext);
1017 case Intrinsic::AMDGPU_rcp: {
1018 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1019 const APFloat &ArgVal = C->getValueAPF();
1020 APFloat Val(ArgVal.getSemantics(), 1.0);
1021 APFloat::opStatus Status = Val.divide(ArgVal,
1022 APFloat::rmNearestTiesToEven);
1023 // Only do this if it was exact and therefore not dependent on the
1025 if (Status == APFloat::opOK)
1026 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1031 case Intrinsic::stackrestore: {
1032 // If the save is right next to the restore, remove the restore. This can
1033 // happen when variable allocas are DCE'd.
1034 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1035 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1036 BasicBlock::iterator BI = SS;
1038 return EraseInstFromFunction(CI);
1042 // Scan down this block to see if there is another stack restore in the
1043 // same block without an intervening call/alloca.
1044 BasicBlock::iterator BI = II;
1045 TerminatorInst *TI = II->getParent()->getTerminator();
1046 bool CannotRemove = false;
1047 for (++BI; &*BI != TI; ++BI) {
1048 if (isa<AllocaInst>(BI)) {
1049 CannotRemove = true;
1052 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1053 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1054 // If there is a stackrestore below this one, remove this one.
1055 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1056 return EraseInstFromFunction(CI);
1057 // Otherwise, ignore the intrinsic.
1059 // If we found a non-intrinsic call, we can't remove the stack
1061 CannotRemove = true;
1067 // If the stack restore is in a return, resume, or unwind block and if there
1068 // are no allocas or calls between the restore and the return, nuke the
1070 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1071 return EraseInstFromFunction(CI);
1074 case Intrinsic::assume: {
1075 // Canonicalize assume(a && b) -> assume(a); assume(b);
1076 // Note: New assumption intrinsics created here are registered by
1077 // the InstCombineIRInserter object.
1078 Value *IIOperand = II->getArgOperand(0), *A, *B,
1079 *AssumeIntrinsic = II->getCalledValue();
1080 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1081 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1082 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1083 return EraseInstFromFunction(*II);
1085 // assume(!(a || b)) -> assume(!a); assume(!b);
1086 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1087 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1089 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1091 return EraseInstFromFunction(*II);
1094 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1095 // (if assume is valid at the load)
1096 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1097 Value *LHS = ICmp->getOperand(0);
1098 Value *RHS = ICmp->getOperand(1);
1099 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1100 isa<LoadInst>(LHS) &&
1101 isa<Constant>(RHS) &&
1102 RHS->getType()->isPointerTy() &&
1103 cast<Constant>(RHS)->isNullValue()) {
1104 LoadInst* LI = cast<LoadInst>(LHS);
1105 if (isValidAssumeForContext(II, LI, DL, DT)) {
1106 MDNode *MD = MDNode::get(II->getContext(), None);
1107 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1108 return EraseInstFromFunction(*II);
1111 // TODO: apply nonnull return attributes to calls and invokes
1112 // TODO: apply range metadata for range check patterns?
1114 // If there is a dominating assume with the same condition as this one,
1115 // then this one is redundant, and should be removed.
1116 APInt KnownZero(1, 0), KnownOne(1, 0);
1117 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1118 if (KnownOne.isAllOnesValue())
1119 return EraseInstFromFunction(*II);
1123 case Intrinsic::experimental_gc_relocate: {
1124 // Translate facts known about a pointer before relocating into
1125 // facts about the relocate value, while being careful to
1126 // preserve relocation semantics.
1127 GCRelocateOperands Operands(II);
1128 Value *DerivedPtr = Operands.derivedPtr();
1130 // Remove the relocation if unused, note that this check is required
1131 // to prevent the cases below from looping forever.
1132 if (II->use_empty())
1133 return EraseInstFromFunction(*II);
1135 // Undef is undef, even after relocation.
1136 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1137 // most practical collectors, but there was discussion in the review thread
1138 // about whether it was legal for all possible collectors.
1139 if (isa<UndefValue>(DerivedPtr))
1140 return ReplaceInstUsesWith(*II, DerivedPtr);
1142 // The relocation of null will be null for most any collector.
1143 // TODO: provide a hook for this in GCStrategy. There might be some weird
1144 // collector this property does not hold for.
1145 if (isa<ConstantPointerNull>(DerivedPtr))
1146 return ReplaceInstUsesWith(*II, DerivedPtr);
1148 // isKnownNonNull -> nonnull attribute
1149 if (isKnownNonNull(DerivedPtr))
1150 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1152 // TODO: dereferenceable -> deref attribute
1154 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1155 // Canonicalize on the type from the uses to the defs
1157 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1161 return visitCallSite(II);
1164 // InvokeInst simplification
1166 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1167 return visitCallSite(&II);
1170 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1171 /// passed through the varargs area, we can eliminate the use of the cast.
1172 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1173 const CastInst * const CI,
1174 const DataLayout * const DL,
1176 if (!CI->isLosslessCast())
1179 // If this is a GC intrinsic, avoid munging types. We need types for
1180 // statepoint reconstruction in SelectionDAG.
1181 // TODO: This is probably something which should be expanded to all
1182 // intrinsics since the entire point of intrinsics is that
1183 // they are understandable by the optimizer.
1184 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1187 // The size of ByVal or InAlloca arguments is derived from the type, so we
1188 // can't change to a type with a different size. If the size were
1189 // passed explicitly we could avoid this check.
1190 if (!CS.isByValOrInAllocaArgument(ix))
1194 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1195 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1196 if (!SrcTy->isSized() || !DstTy->isSized())
1198 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
1203 // Try to fold some different type of calls here.
1204 // Currently we're only working with the checking functions, memcpy_chk,
1205 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1206 // strcat_chk and strncat_chk.
1207 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
1208 if (!CI->getCalledFunction()) return nullptr;
1210 if (Value *With = Simplifier->optimizeCall(CI)) {
1212 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1218 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1219 // Strip off at most one level of pointer casts, looking for an alloca. This
1220 // is good enough in practice and simpler than handling any number of casts.
1221 Value *Underlying = TrampMem->stripPointerCasts();
1222 if (Underlying != TrampMem &&
1223 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1225 if (!isa<AllocaInst>(Underlying))
1228 IntrinsicInst *InitTrampoline = nullptr;
1229 for (User *U : TrampMem->users()) {
1230 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1233 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1235 // More than one init_trampoline writes to this value. Give up.
1237 InitTrampoline = II;
1240 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1241 // Allow any number of calls to adjust.trampoline.
1246 // No call to init.trampoline found.
1247 if (!InitTrampoline)
1250 // Check that the alloca is being used in the expected way.
1251 if (InitTrampoline->getOperand(0) != TrampMem)
1254 return InitTrampoline;
1257 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1259 // Visit all the previous instructions in the basic block, and try to find a
1260 // init.trampoline which has a direct path to the adjust.trampoline.
1261 for (BasicBlock::iterator I = AdjustTramp,
1262 E = AdjustTramp->getParent()->begin(); I != E; ) {
1263 Instruction *Inst = --I;
1264 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1265 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1266 II->getOperand(0) == TrampMem)
1268 if (Inst->mayWriteToMemory())
1274 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1275 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1276 // to a direct call to a function. Otherwise return NULL.
1278 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1279 Callee = Callee->stripPointerCasts();
1280 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1282 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1285 Value *TrampMem = AdjustTramp->getOperand(0);
1287 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1289 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1294 // visitCallSite - Improvements for call and invoke instructions.
1296 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1297 if (isAllocLikeFn(CS.getInstruction(), TLI))
1298 return visitAllocSite(*CS.getInstruction());
1300 bool Changed = false;
1302 // If the callee is a pointer to a function, attempt to move any casts to the
1303 // arguments of the call/invoke.
1304 Value *Callee = CS.getCalledValue();
1305 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1308 if (Function *CalleeF = dyn_cast<Function>(Callee))
1309 // If the call and callee calling conventions don't match, this call must
1310 // be unreachable, as the call is undefined.
1311 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1312 // Only do this for calls to a function with a body. A prototype may
1313 // not actually end up matching the implementation's calling conv for a
1314 // variety of reasons (e.g. it may be written in assembly).
1315 !CalleeF->isDeclaration()) {
1316 Instruction *OldCall = CS.getInstruction();
1317 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1318 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1320 // If OldCall does not return void then replaceAllUsesWith undef.
1321 // This allows ValueHandlers and custom metadata to adjust itself.
1322 if (!OldCall->getType()->isVoidTy())
1323 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1324 if (isa<CallInst>(OldCall))
1325 return EraseInstFromFunction(*OldCall);
1327 // We cannot remove an invoke, because it would change the CFG, just
1328 // change the callee to a null pointer.
1329 cast<InvokeInst>(OldCall)->setCalledFunction(
1330 Constant::getNullValue(CalleeF->getType()));
1334 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1335 // If CS does not return void then replaceAllUsesWith undef.
1336 // This allows ValueHandlers and custom metadata to adjust itself.
1337 if (!CS.getInstruction()->getType()->isVoidTy())
1338 ReplaceInstUsesWith(*CS.getInstruction(),
1339 UndefValue::get(CS.getInstruction()->getType()));
1341 if (isa<InvokeInst>(CS.getInstruction())) {
1342 // Can't remove an invoke because we cannot change the CFG.
1346 // This instruction is not reachable, just remove it. We insert a store to
1347 // undef so that we know that this code is not reachable, despite the fact
1348 // that we can't modify the CFG here.
1349 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1350 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1351 CS.getInstruction());
1353 return EraseInstFromFunction(*CS.getInstruction());
1356 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1357 return transformCallThroughTrampoline(CS, II);
1359 PointerType *PTy = cast<PointerType>(Callee->getType());
1360 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1361 if (FTy->isVarArg()) {
1362 int ix = FTy->getNumParams();
1363 // See if we can optimize any arguments passed through the varargs area of
1365 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1366 E = CS.arg_end(); I != E; ++I, ++ix) {
1367 CastInst *CI = dyn_cast<CastInst>(*I);
1368 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1369 *I = CI->getOperand(0);
1375 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1376 // Inline asm calls cannot throw - mark them 'nounwind'.
1377 CS.setDoesNotThrow();
1381 // Try to optimize the call if possible, we require DataLayout for most of
1382 // this. None of these calls are seen as possibly dead so go ahead and
1383 // delete the instruction now.
1384 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1385 Instruction *I = tryOptimizeCall(CI, DL);
1386 // If we changed something return the result, etc. Otherwise let
1387 // the fallthrough check.
1388 if (I) return EraseInstFromFunction(*I);
1391 return Changed ? CS.getInstruction() : nullptr;
1394 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1395 // attempt to move the cast to the arguments of the call/invoke.
1397 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1399 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1402 Instruction *Caller = CS.getInstruction();
1403 const AttributeSet &CallerPAL = CS.getAttributes();
1405 // Okay, this is a cast from a function to a different type. Unless doing so
1406 // would cause a type conversion of one of our arguments, change this call to
1407 // be a direct call with arguments casted to the appropriate types.
1409 FunctionType *FT = Callee->getFunctionType();
1410 Type *OldRetTy = Caller->getType();
1411 Type *NewRetTy = FT->getReturnType();
1413 // Check to see if we are changing the return type...
1414 if (OldRetTy != NewRetTy) {
1416 if (NewRetTy->isStructTy())
1417 return false; // TODO: Handle multiple return values.
1419 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1420 if (Callee->isDeclaration())
1421 return false; // Cannot transform this return value.
1423 if (!Caller->use_empty() &&
1424 // void -> non-void is handled specially
1425 !NewRetTy->isVoidTy())
1426 return false; // Cannot transform this return value.
1429 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1430 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1432 hasAttributes(AttributeFuncs::
1433 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1434 AttributeSet::ReturnIndex))
1435 return false; // Attribute not compatible with transformed value.
1438 // If the callsite is an invoke instruction, and the return value is used by
1439 // a PHI node in a successor, we cannot change the return type of the call
1440 // because there is no place to put the cast instruction (without breaking
1441 // the critical edge). Bail out in this case.
1442 if (!Caller->use_empty())
1443 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1444 for (User *U : II->users())
1445 if (PHINode *PN = dyn_cast<PHINode>(U))
1446 if (PN->getParent() == II->getNormalDest() ||
1447 PN->getParent() == II->getUnwindDest())
1451 unsigned NumActualArgs = CS.arg_size();
1452 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1454 CallSite::arg_iterator AI = CS.arg_begin();
1455 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1456 Type *ParamTy = FT->getParamType(i);
1457 Type *ActTy = (*AI)->getType();
1459 if (!CastInst::isBitCastable(ActTy, ParamTy))
1460 return false; // Cannot transform this parameter value.
1462 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1463 hasAttributes(AttributeFuncs::
1464 typeIncompatible(ParamTy, i + 1), i + 1))
1465 return false; // Attribute not compatible with transformed value.
1467 if (CS.isInAllocaArgument(i))
1468 return false; // Cannot transform to and from inalloca.
1470 // If the parameter is passed as a byval argument, then we have to have a
1471 // sized type and the sized type has to have the same size as the old type.
1472 if (ParamTy != ActTy &&
1473 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1474 Attribute::ByVal)) {
1475 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1476 if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1479 Type *CurElTy = ActTy->getPointerElementType();
1480 if (DL->getTypeAllocSize(CurElTy) !=
1481 DL->getTypeAllocSize(ParamPTy->getElementType()))
1486 if (Callee->isDeclaration()) {
1487 // Do not delete arguments unless we have a function body.
1488 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1491 // If the callee is just a declaration, don't change the varargsness of the
1492 // call. We don't want to introduce a varargs call where one doesn't
1494 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1495 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1498 // If both the callee and the cast type are varargs, we still have to make
1499 // sure the number of fixed parameters are the same or we have the same
1500 // ABI issues as if we introduce a varargs call.
1501 if (FT->isVarArg() &&
1502 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1503 FT->getNumParams() !=
1504 cast<FunctionType>(APTy->getElementType())->getNumParams())
1508 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1509 !CallerPAL.isEmpty())
1510 // In this case we have more arguments than the new function type, but we
1511 // won't be dropping them. Check that these extra arguments have attributes
1512 // that are compatible with being a vararg call argument.
1513 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1514 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1515 if (Index <= FT->getNumParams())
1518 // Check if it has an attribute that's incompatible with varargs.
1519 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1520 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1525 // Okay, we decided that this is a safe thing to do: go ahead and start
1526 // inserting cast instructions as necessary.
1527 std::vector<Value*> Args;
1528 Args.reserve(NumActualArgs);
1529 SmallVector<AttributeSet, 8> attrVec;
1530 attrVec.reserve(NumCommonArgs);
1532 // Get any return attributes.
1533 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1535 // If the return value is not being used, the type may not be compatible
1536 // with the existing attributes. Wipe out any problematic attributes.
1538 removeAttributes(AttributeFuncs::
1539 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1540 AttributeSet::ReturnIndex);
1542 // Add the new return attributes.
1543 if (RAttrs.hasAttributes())
1544 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1545 AttributeSet::ReturnIndex, RAttrs));
1547 AI = CS.arg_begin();
1548 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1549 Type *ParamTy = FT->getParamType(i);
1551 if ((*AI)->getType() == ParamTy) {
1552 Args.push_back(*AI);
1554 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1557 // Add any parameter attributes.
1558 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1559 if (PAttrs.hasAttributes())
1560 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1564 // If the function takes more arguments than the call was taking, add them
1566 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1567 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1569 // If we are removing arguments to the function, emit an obnoxious warning.
1570 if (FT->getNumParams() < NumActualArgs) {
1571 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1572 if (FT->isVarArg()) {
1573 // Add all of the arguments in their promoted form to the arg list.
1574 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1575 Type *PTy = getPromotedType((*AI)->getType());
1576 if (PTy != (*AI)->getType()) {
1577 // Must promote to pass through va_arg area!
1578 Instruction::CastOps opcode =
1579 CastInst::getCastOpcode(*AI, false, PTy, false);
1580 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1582 Args.push_back(*AI);
1585 // Add any parameter attributes.
1586 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1587 if (PAttrs.hasAttributes())
1588 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1594 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1595 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1596 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1598 if (NewRetTy->isVoidTy())
1599 Caller->setName(""); // Void type should not have a name.
1601 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1605 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1606 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1607 II->getUnwindDest(), Args);
1609 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1610 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1612 CallInst *CI = cast<CallInst>(Caller);
1613 NC = Builder->CreateCall(Callee, Args);
1615 if (CI->isTailCall())
1616 cast<CallInst>(NC)->setTailCall();
1617 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1618 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1621 // Insert a cast of the return type as necessary.
1623 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1624 if (!NV->getType()->isVoidTy()) {
1625 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1626 NC->setDebugLoc(Caller->getDebugLoc());
1628 // If this is an invoke instruction, we should insert it after the first
1629 // non-phi, instruction in the normal successor block.
1630 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1631 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1632 InsertNewInstBefore(NC, *I);
1634 // Otherwise, it's a call, just insert cast right after the call.
1635 InsertNewInstBefore(NC, *Caller);
1637 Worklist.AddUsersToWorkList(*Caller);
1639 NV = UndefValue::get(Caller->getType());
1643 if (!Caller->use_empty())
1644 ReplaceInstUsesWith(*Caller, NV);
1645 else if (Caller->hasValueHandle()) {
1646 if (OldRetTy == NV->getType())
1647 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1649 // We cannot call ValueIsRAUWd with a different type, and the
1650 // actual tracked value will disappear.
1651 ValueHandleBase::ValueIsDeleted(Caller);
1654 EraseInstFromFunction(*Caller);
1658 // transformCallThroughTrampoline - Turn a call to a function created by
1659 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1660 // underlying function.
1663 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1664 IntrinsicInst *Tramp) {
1665 Value *Callee = CS.getCalledValue();
1666 PointerType *PTy = cast<PointerType>(Callee->getType());
1667 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1668 const AttributeSet &Attrs = CS.getAttributes();
1670 // If the call already has the 'nest' attribute somewhere then give up -
1671 // otherwise 'nest' would occur twice after splicing in the chain.
1672 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1676 "transformCallThroughTrampoline called with incorrect CallSite.");
1678 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1679 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1680 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1682 const AttributeSet &NestAttrs = NestF->getAttributes();
1683 if (!NestAttrs.isEmpty()) {
1684 unsigned NestIdx = 1;
1685 Type *NestTy = nullptr;
1686 AttributeSet NestAttr;
1688 // Look for a parameter marked with the 'nest' attribute.
1689 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1690 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1691 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1692 // Record the parameter type and any other attributes.
1694 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1699 Instruction *Caller = CS.getInstruction();
1700 std::vector<Value*> NewArgs;
1701 NewArgs.reserve(CS.arg_size() + 1);
1703 SmallVector<AttributeSet, 8> NewAttrs;
1704 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1706 // Insert the nest argument into the call argument list, which may
1707 // mean appending it. Likewise for attributes.
1709 // Add any result attributes.
1710 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1711 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1712 Attrs.getRetAttributes()));
1716 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1718 if (Idx == NestIdx) {
1719 // Add the chain argument and attributes.
1720 Value *NestVal = Tramp->getArgOperand(2);
1721 if (NestVal->getType() != NestTy)
1722 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1723 NewArgs.push_back(NestVal);
1724 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1731 // Add the original argument and attributes.
1732 NewArgs.push_back(*I);
1733 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1734 if (Attr.hasAttributes(Idx)) {
1735 AttrBuilder B(Attr, Idx);
1736 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1737 Idx + (Idx >= NestIdx), B));
1744 // Add any function attributes.
1745 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1746 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1747 Attrs.getFnAttributes()));
1749 // The trampoline may have been bitcast to a bogus type (FTy).
1750 // Handle this by synthesizing a new function type, equal to FTy
1751 // with the chain parameter inserted.
1753 std::vector<Type*> NewTypes;
1754 NewTypes.reserve(FTy->getNumParams()+1);
1756 // Insert the chain's type into the list of parameter types, which may
1757 // mean appending it.
1760 FunctionType::param_iterator I = FTy->param_begin(),
1761 E = FTy->param_end();
1765 // Add the chain's type.
1766 NewTypes.push_back(NestTy);
1771 // Add the original type.
1772 NewTypes.push_back(*I);
1778 // Replace the trampoline call with a direct call. Let the generic
1779 // code sort out any function type mismatches.
1780 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1782 Constant *NewCallee =
1783 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1784 NestF : ConstantExpr::getBitCast(NestF,
1785 PointerType::getUnqual(NewFTy));
1786 const AttributeSet &NewPAL =
1787 AttributeSet::get(FTy->getContext(), NewAttrs);
1789 Instruction *NewCaller;
1790 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1791 NewCaller = InvokeInst::Create(NewCallee,
1792 II->getNormalDest(), II->getUnwindDest(),
1794 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1795 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1797 NewCaller = CallInst::Create(NewCallee, NewArgs);
1798 if (cast<CallInst>(Caller)->isTailCall())
1799 cast<CallInst>(NewCaller)->setTailCall();
1800 cast<CallInst>(NewCaller)->
1801 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1802 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1809 // Replace the trampoline call with a direct call. Since there is no 'nest'
1810 // parameter, there is no need to adjust the argument list. Let the generic
1811 // code sort out any function type mismatches.
1812 Constant *NewCallee =
1813 NestF->getType() == PTy ? NestF :
1814 ConstantExpr::getBitCast(NestF, PTy);
1815 CS.setCalledFunction(NewCallee);
1816 return CS.getInstruction();