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, AT, MI, DT);
64 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, AT, 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, AT, 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 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
445 APInt LHSKnownZero(BitWidth, 0);
446 APInt LHSKnownOne(BitWidth, 0);
447 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, 0, II);
448 APInt RHSKnownZero(BitWidth, 0);
449 APInt RHSKnownOne(BitWidth, 0);
450 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, 0, II);
452 // Get the largest possible values for each operand.
453 APInt LHSMax = ~LHSKnownZero;
454 APInt RHSMax = ~RHSKnownZero;
456 // If multiplying the maximum values does not overflow then we can turn
457 // this into a plain NUW mul.
459 LHSMax.umul_ov(RHSMax, Overflow);
461 return CreateOverflowTuple(II, Builder->CreateNUWMul(LHS, RHS), false);
464 case Intrinsic::smul_with_overflow:
465 // Canonicalize constants into the RHS.
466 if (isa<Constant>(II->getArgOperand(0)) &&
467 !isa<Constant>(II->getArgOperand(1))) {
468 Value *LHS = II->getArgOperand(0);
469 II->setArgOperand(0, II->getArgOperand(1));
470 II->setArgOperand(1, LHS);
474 // X * undef -> undef
475 if (isa<UndefValue>(II->getArgOperand(1)))
476 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
478 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
481 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
483 // X * 1 -> {X, false}
484 if (RHSI->equalsInt(1)) {
485 return CreateOverflowTuple(II, II->getArgOperand(0), false,
489 if (II->getIntrinsicID() == Intrinsic::smul_with_overflow) {
490 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
491 if (WillNotOverflowSignedMul(LHS, RHS, II)) {
492 return CreateOverflowTuple(II, Builder->CreateNSWMul(LHS, RHS), false);
496 case Intrinsic::minnum:
497 case Intrinsic::maxnum: {
498 Value *Arg0 = II->getArgOperand(0);
499 Value *Arg1 = II->getArgOperand(1);
503 return ReplaceInstUsesWith(CI, Arg0);
505 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
506 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
508 // Canonicalize constants into the RHS.
510 II->setArgOperand(0, Arg1);
511 II->setArgOperand(1, Arg0);
516 if (C1 && C1->isNaN())
517 return ReplaceInstUsesWith(CI, Arg0);
519 // This is the value because if undef were NaN, we would return the other
520 // value and cannot return a NaN unless both operands are.
522 // fmin(undef, x) -> x
523 if (isa<UndefValue>(Arg0))
524 return ReplaceInstUsesWith(CI, Arg1);
526 // fmin(x, undef) -> x
527 if (isa<UndefValue>(Arg1))
528 return ReplaceInstUsesWith(CI, Arg0);
532 if (II->getIntrinsicID() == Intrinsic::minnum) {
533 // fmin(x, fmin(x, y)) -> fmin(x, y)
534 // fmin(y, fmin(x, y)) -> fmin(x, y)
535 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
536 if (Arg0 == X || Arg0 == Y)
537 return ReplaceInstUsesWith(CI, Arg1);
540 // fmin(fmin(x, y), x) -> fmin(x, y)
541 // fmin(fmin(x, y), y) -> fmin(x, y)
542 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
543 if (Arg1 == X || Arg1 == Y)
544 return ReplaceInstUsesWith(CI, Arg0);
547 // TODO: fmin(nnan x, inf) -> x
548 // TODO: fmin(nnan ninf x, flt_max) -> x
549 if (C1 && C1->isInfinity()) {
550 // fmin(x, -inf) -> -inf
551 if (C1->isNegative())
552 return ReplaceInstUsesWith(CI, Arg1);
555 assert(II->getIntrinsicID() == Intrinsic::maxnum);
556 // fmax(x, fmax(x, y)) -> fmax(x, y)
557 // fmax(y, fmax(x, y)) -> fmax(x, y)
558 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
559 if (Arg0 == X || Arg0 == Y)
560 return ReplaceInstUsesWith(CI, Arg1);
563 // fmax(fmax(x, y), x) -> fmax(x, y)
564 // fmax(fmax(x, y), y) -> fmax(x, y)
565 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
566 if (Arg1 == X || Arg1 == Y)
567 return ReplaceInstUsesWith(CI, Arg0);
570 // TODO: fmax(nnan x, -inf) -> x
571 // TODO: fmax(nnan ninf x, -flt_max) -> x
572 if (C1 && C1->isInfinity()) {
573 // fmax(x, inf) -> inf
574 if (!C1->isNegative())
575 return ReplaceInstUsesWith(CI, Arg1);
580 case Intrinsic::ppc_altivec_lvx:
581 case Intrinsic::ppc_altivec_lvxl:
582 // Turn PPC lvx -> load if the pointer is known aligned.
583 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16,
584 DL, AT, II, DT) >= 16) {
585 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
586 PointerType::getUnqual(II->getType()));
587 return new LoadInst(Ptr);
590 case Intrinsic::ppc_vsx_lxvw4x:
591 case Intrinsic::ppc_vsx_lxvd2x: {
592 // Turn PPC VSX loads into normal loads.
593 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
594 PointerType::getUnqual(II->getType()));
595 return new LoadInst(Ptr, Twine(""), false, 1);
597 case Intrinsic::ppc_altivec_stvx:
598 case Intrinsic::ppc_altivec_stvxl:
599 // Turn stvx -> store if the pointer is known aligned.
600 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16,
601 DL, AT, II, DT) >= 16) {
603 PointerType::getUnqual(II->getArgOperand(0)->getType());
604 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
605 return new StoreInst(II->getArgOperand(0), Ptr);
608 case Intrinsic::ppc_vsx_stxvw4x:
609 case Intrinsic::ppc_vsx_stxvd2x: {
610 // Turn PPC VSX stores into normal stores.
611 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
612 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
613 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
615 case Intrinsic::x86_sse_storeu_ps:
616 case Intrinsic::x86_sse2_storeu_pd:
617 case Intrinsic::x86_sse2_storeu_dq:
618 // Turn X86 storeu -> store if the pointer is known aligned.
619 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16,
620 DL, AT, II, DT) >= 16) {
622 PointerType::getUnqual(II->getArgOperand(1)->getType());
623 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
624 return new StoreInst(II->getArgOperand(1), Ptr);
628 case Intrinsic::x86_sse_cvtss2si:
629 case Intrinsic::x86_sse_cvtss2si64:
630 case Intrinsic::x86_sse_cvttss2si:
631 case Intrinsic::x86_sse_cvttss2si64:
632 case Intrinsic::x86_sse2_cvtsd2si:
633 case Intrinsic::x86_sse2_cvtsd2si64:
634 case Intrinsic::x86_sse2_cvttsd2si:
635 case Intrinsic::x86_sse2_cvttsd2si64: {
636 // These intrinsics only demand the 0th element of their input vectors. If
637 // we can simplify the input based on that, do so now.
639 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
640 APInt DemandedElts(VWidth, 1);
641 APInt UndefElts(VWidth, 0);
642 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
643 DemandedElts, UndefElts)) {
644 II->setArgOperand(0, V);
650 // Constant fold <A x Bi> << Ci.
651 // FIXME: We don't handle _dq because it's a shift of an i128, but is
652 // represented in the IR as <2 x i64>. A per element shift is wrong.
653 case Intrinsic::x86_sse2_psll_d:
654 case Intrinsic::x86_sse2_psll_q:
655 case Intrinsic::x86_sse2_psll_w:
656 case Intrinsic::x86_sse2_pslli_d:
657 case Intrinsic::x86_sse2_pslli_q:
658 case Intrinsic::x86_sse2_pslli_w:
659 case Intrinsic::x86_avx2_psll_d:
660 case Intrinsic::x86_avx2_psll_q:
661 case Intrinsic::x86_avx2_psll_w:
662 case Intrinsic::x86_avx2_pslli_d:
663 case Intrinsic::x86_avx2_pslli_q:
664 case Intrinsic::x86_avx2_pslli_w:
665 case Intrinsic::x86_sse2_psrl_d:
666 case Intrinsic::x86_sse2_psrl_q:
667 case Intrinsic::x86_sse2_psrl_w:
668 case Intrinsic::x86_sse2_psrli_d:
669 case Intrinsic::x86_sse2_psrli_q:
670 case Intrinsic::x86_sse2_psrli_w:
671 case Intrinsic::x86_avx2_psrl_d:
672 case Intrinsic::x86_avx2_psrl_q:
673 case Intrinsic::x86_avx2_psrl_w:
674 case Intrinsic::x86_avx2_psrli_d:
675 case Intrinsic::x86_avx2_psrli_q:
676 case Intrinsic::x86_avx2_psrli_w: {
677 // Simplify if count is constant. To 0 if >= BitWidth,
678 // otherwise to shl/lshr.
679 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
680 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
685 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
689 auto Vec = II->getArgOperand(0);
690 auto VT = cast<VectorType>(Vec->getType());
691 if (Count->getZExtValue() >
692 VT->getElementType()->getPrimitiveSizeInBits() - 1)
693 return ReplaceInstUsesWith(
694 CI, ConstantAggregateZero::get(Vec->getType()));
696 bool isPackedShiftLeft = true;
697 switch (II->getIntrinsicID()) {
699 case Intrinsic::x86_sse2_psrl_d:
700 case Intrinsic::x86_sse2_psrl_q:
701 case Intrinsic::x86_sse2_psrl_w:
702 case Intrinsic::x86_sse2_psrli_d:
703 case Intrinsic::x86_sse2_psrli_q:
704 case Intrinsic::x86_sse2_psrli_w:
705 case Intrinsic::x86_avx2_psrl_d:
706 case Intrinsic::x86_avx2_psrl_q:
707 case Intrinsic::x86_avx2_psrl_w:
708 case Intrinsic::x86_avx2_psrli_d:
709 case Intrinsic::x86_avx2_psrli_q:
710 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
713 unsigned VWidth = VT->getNumElements();
714 // Get a constant vector of the same type as the first operand.
715 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
716 if (isPackedShiftLeft)
717 return BinaryOperator::CreateShl(Vec,
718 Builder->CreateVectorSplat(VWidth, VTCI));
720 return BinaryOperator::CreateLShr(Vec,
721 Builder->CreateVectorSplat(VWidth, VTCI));
724 case Intrinsic::x86_sse41_pmovsxbw:
725 case Intrinsic::x86_sse41_pmovsxwd:
726 case Intrinsic::x86_sse41_pmovsxdq:
727 case Intrinsic::x86_sse41_pmovzxbw:
728 case Intrinsic::x86_sse41_pmovzxwd:
729 case Intrinsic::x86_sse41_pmovzxdq: {
730 // pmov{s|z}x ignores the upper half of their input vectors.
732 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
733 unsigned LowHalfElts = VWidth / 2;
734 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
735 APInt UndefElts(VWidth, 0);
736 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
739 II->setArgOperand(0, TmpV);
745 case Intrinsic::x86_sse4a_insertqi: {
746 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
748 // TODO: eventually we should lower this intrinsic to IR
749 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
750 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
751 unsigned Index = CIStart->getZExtValue();
752 // From AMD documentation: "a value of zero in the field length is
753 // defined as length of 64".
754 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
756 // From AMD documentation: "If the sum of the bit index + length field
757 // is greater than 64, the results are undefined".
759 // Note that both field index and field length are 8-bit quantities.
760 // Since variables 'Index' and 'Length' are unsigned values
761 // obtained from zero-extending field index and field length
762 // respectively, their sum should never wrap around.
763 if ((Index + Length) > 64)
764 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
766 if (Length == 64 && Index == 0) {
767 Value *Vec = II->getArgOperand(1);
768 Value *Undef = UndefValue::get(Vec->getType());
769 const uint32_t Mask[] = { 0, 2 };
770 return ReplaceInstUsesWith(
772 Builder->CreateShuffleVector(
773 Vec, Undef, ConstantDataVector::get(
774 II->getContext(), makeArrayRef(Mask))));
776 } else if (auto Source =
777 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
778 if (Source->hasOneUse() &&
779 Source->getArgOperand(1) == II->getArgOperand(1)) {
780 // If the source of the insert has only one use and it's another
781 // insert (and they're both inserting from the same vector), try to
782 // bundle both together.
784 dyn_cast<ConstantInt>(Source->getArgOperand(2));
786 dyn_cast<ConstantInt>(Source->getArgOperand(3));
787 if (CISourceStart && CISourceWidth) {
788 unsigned Start = CIStart->getZExtValue();
789 unsigned Width = CIWidth->getZExtValue();
790 unsigned End = Start + Width;
791 unsigned SourceStart = CISourceStart->getZExtValue();
792 unsigned SourceWidth = CISourceWidth->getZExtValue();
793 unsigned SourceEnd = SourceStart + SourceWidth;
794 unsigned NewStart, NewWidth;
795 bool ShouldReplace = false;
796 if (Start <= SourceStart && SourceStart <= End) {
798 NewWidth = std::max(End, SourceEnd) - NewStart;
799 ShouldReplace = true;
800 } else if (SourceStart <= Start && Start <= SourceEnd) {
801 NewStart = SourceStart;
802 NewWidth = std::max(SourceEnd, End) - NewStart;
803 ShouldReplace = true;
807 Constant *ConstantWidth = ConstantInt::get(
808 II->getArgOperand(2)->getType(), NewWidth, false);
809 Constant *ConstantStart = ConstantInt::get(
810 II->getArgOperand(3)->getType(), NewStart, false);
811 Value *Args[4] = { Source->getArgOperand(0),
812 II->getArgOperand(1), ConstantWidth,
814 Module *M = CI.getParent()->getParent()->getParent();
816 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
817 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
827 case Intrinsic::x86_sse41_pblendvb:
828 case Intrinsic::x86_sse41_blendvps:
829 case Intrinsic::x86_sse41_blendvpd:
830 case Intrinsic::x86_avx_blendv_ps_256:
831 case Intrinsic::x86_avx_blendv_pd_256:
832 case Intrinsic::x86_avx2_pblendvb: {
833 // Convert blendv* to vector selects if the mask is constant.
834 // This optimization is convoluted because the intrinsic is defined as
835 // getting a vector of floats or doubles for the ps and pd versions.
836 // FIXME: That should be changed.
837 Value *Mask = II->getArgOperand(2);
838 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
839 auto Tyi1 = Builder->getInt1Ty();
840 auto SelectorType = cast<VectorType>(Mask->getType());
841 auto EltTy = SelectorType->getElementType();
842 unsigned Size = SelectorType->getNumElements();
846 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
847 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
848 "Wrong arguments for variable blend intrinsic");
849 SmallVector<Constant *, 32> Selectors;
850 for (unsigned I = 0; I < Size; ++I) {
851 // The intrinsics only read the top bit
854 Selector = C->getElementAsInteger(I);
856 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
857 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
859 auto NewSelector = ConstantVector::get(Selectors);
860 return SelectInst::Create(NewSelector, II->getArgOperand(1),
861 II->getArgOperand(0), "blendv");
867 case Intrinsic::x86_avx_vpermilvar_ps:
868 case Intrinsic::x86_avx_vpermilvar_ps_256:
869 case Intrinsic::x86_avx_vpermilvar_pd:
870 case Intrinsic::x86_avx_vpermilvar_pd_256: {
871 // Convert vpermil* to shufflevector if the mask is constant.
872 Value *V = II->getArgOperand(1);
873 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
874 assert(Size == 8 || Size == 4 || Size == 2);
876 if (auto C = dyn_cast<ConstantDataVector>(V)) {
877 // The intrinsics only read one or two bits, clear the rest.
878 for (unsigned I = 0; I < Size; ++I) {
879 uint32_t Index = C->getElementAsInteger(I) & 0x3;
880 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
881 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
885 } else if (isa<ConstantAggregateZero>(V)) {
886 for (unsigned I = 0; I < Size; ++I)
891 // The _256 variants are a bit trickier since the mask bits always index
892 // into the corresponding 128 half. In order to convert to a generic
893 // shuffle, we have to make that explicit.
894 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
895 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
896 for (unsigned I = Size / 2; I < Size; ++I)
897 Indexes[I] += Size / 2;
900 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
901 auto V1 = II->getArgOperand(0);
902 auto V2 = UndefValue::get(V1->getType());
903 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
904 return ReplaceInstUsesWith(CI, Shuffle);
907 case Intrinsic::ppc_altivec_vperm:
908 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
909 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
910 // a vectorshuffle for little endian, we must undo the transformation
911 // performed on vec_perm in altivec.h. That is, we must complement
912 // the permutation mask with respect to 31 and reverse the order of
914 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
915 assert(Mask->getType()->getVectorNumElements() == 16 &&
916 "Bad type for intrinsic!");
918 // Check that all of the elements are integer constants or undefs.
919 bool AllEltsOk = true;
920 for (unsigned i = 0; i != 16; ++i) {
921 Constant *Elt = Mask->getAggregateElement(i);
922 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
929 // Cast the input vectors to byte vectors.
930 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
932 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
934 Value *Result = UndefValue::get(Op0->getType());
936 // Only extract each element once.
937 Value *ExtractedElts[32];
938 memset(ExtractedElts, 0, sizeof(ExtractedElts));
940 for (unsigned i = 0; i != 16; ++i) {
941 if (isa<UndefValue>(Mask->getAggregateElement(i)))
944 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
945 Idx &= 31; // Match the hardware behavior.
946 if (DL && DL->isLittleEndian())
949 if (!ExtractedElts[Idx]) {
950 Value *Op0ToUse = (DL && DL->isLittleEndian()) ? Op1 : Op0;
951 Value *Op1ToUse = (DL && DL->isLittleEndian()) ? Op0 : Op1;
953 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
954 Builder->getInt32(Idx&15));
957 // Insert this value into the result vector.
958 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
959 Builder->getInt32(i));
961 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
966 case Intrinsic::arm_neon_vld1:
967 case Intrinsic::arm_neon_vld2:
968 case Intrinsic::arm_neon_vld3:
969 case Intrinsic::arm_neon_vld4:
970 case Intrinsic::arm_neon_vld2lane:
971 case Intrinsic::arm_neon_vld3lane:
972 case Intrinsic::arm_neon_vld4lane:
973 case Intrinsic::arm_neon_vst1:
974 case Intrinsic::arm_neon_vst2:
975 case Intrinsic::arm_neon_vst3:
976 case Intrinsic::arm_neon_vst4:
977 case Intrinsic::arm_neon_vst2lane:
978 case Intrinsic::arm_neon_vst3lane:
979 case Intrinsic::arm_neon_vst4lane: {
980 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, AT, II, DT);
981 unsigned AlignArg = II->getNumArgOperands() - 1;
982 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
983 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
984 II->setArgOperand(AlignArg,
985 ConstantInt::get(Type::getInt32Ty(II->getContext()),
992 case Intrinsic::arm_neon_vmulls:
993 case Intrinsic::arm_neon_vmullu:
994 case Intrinsic::aarch64_neon_smull:
995 case Intrinsic::aarch64_neon_umull: {
996 Value *Arg0 = II->getArgOperand(0);
997 Value *Arg1 = II->getArgOperand(1);
999 // Handle mul by zero first:
1000 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1001 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1004 // Check for constant LHS & RHS - in this case we just simplify.
1005 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1006 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1007 VectorType *NewVT = cast<VectorType>(II->getType());
1008 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1009 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1010 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1011 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1013 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1016 // Couldn't simplify - canonicalize constant to the RHS.
1017 std::swap(Arg0, Arg1);
1020 // Handle mul by one:
1021 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1022 if (ConstantInt *Splat =
1023 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1025 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1026 /*isSigned=*/!Zext);
1031 case Intrinsic::AMDGPU_rcp: {
1032 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1033 const APFloat &ArgVal = C->getValueAPF();
1034 APFloat Val(ArgVal.getSemantics(), 1.0);
1035 APFloat::opStatus Status = Val.divide(ArgVal,
1036 APFloat::rmNearestTiesToEven);
1037 // Only do this if it was exact and therefore not dependent on the
1039 if (Status == APFloat::opOK)
1040 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1045 case Intrinsic::stackrestore: {
1046 // If the save is right next to the restore, remove the restore. This can
1047 // happen when variable allocas are DCE'd.
1048 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1049 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1050 BasicBlock::iterator BI = SS;
1052 return EraseInstFromFunction(CI);
1056 // Scan down this block to see if there is another stack restore in the
1057 // same block without an intervening call/alloca.
1058 BasicBlock::iterator BI = II;
1059 TerminatorInst *TI = II->getParent()->getTerminator();
1060 bool CannotRemove = false;
1061 for (++BI; &*BI != TI; ++BI) {
1062 if (isa<AllocaInst>(BI)) {
1063 CannotRemove = true;
1066 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1067 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1068 // If there is a stackrestore below this one, remove this one.
1069 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1070 return EraseInstFromFunction(CI);
1071 // Otherwise, ignore the intrinsic.
1073 // If we found a non-intrinsic call, we can't remove the stack
1075 CannotRemove = true;
1081 // If the stack restore is in a return, resume, or unwind block and if there
1082 // are no allocas or calls between the restore and the return, nuke the
1084 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1085 return EraseInstFromFunction(CI);
1088 case Intrinsic::assume: {
1089 // Canonicalize assume(a && b) -> assume(a); assume(b);
1090 // Note: New assumption intrinsics created here are registered by
1091 // the InstCombineIRInserter object.
1092 Value *IIOperand = II->getArgOperand(0), *A, *B,
1093 *AssumeIntrinsic = II->getCalledValue();
1094 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1095 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1096 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1097 return EraseInstFromFunction(*II);
1099 // assume(!(a || b)) -> assume(!a); assume(!b);
1100 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1101 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1103 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1105 return EraseInstFromFunction(*II);
1108 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1109 // (if assume is valid at the load)
1110 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1111 Value *LHS = ICmp->getOperand(0);
1112 Value *RHS = ICmp->getOperand(1);
1113 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1114 isa<LoadInst>(LHS) &&
1115 isa<Constant>(RHS) &&
1116 RHS->getType()->isPointerTy() &&
1117 cast<Constant>(RHS)->isNullValue()) {
1118 LoadInst* LI = cast<LoadInst>(LHS);
1119 if (isValidAssumeForContext(II, LI, DL, DT)) {
1120 MDNode *MD = MDNode::get(II->getContext(), None);
1121 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1122 return EraseInstFromFunction(*II);
1125 // TODO: apply nonnull return attributes to calls and invokes
1126 // TODO: apply range metadata for range check patterns?
1128 // If there is a dominating assume with the same condition as this one,
1129 // then this one is redundant, and should be removed.
1130 APInt KnownZero(1, 0), KnownOne(1, 0);
1131 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1132 if (KnownOne.isAllOnesValue())
1133 return EraseInstFromFunction(*II);
1139 return visitCallSite(II);
1142 // InvokeInst simplification
1144 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1145 return visitCallSite(&II);
1148 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1149 /// passed through the varargs area, we can eliminate the use of the cast.
1150 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1151 const CastInst * const CI,
1152 const DataLayout * const DL,
1154 if (!CI->isLosslessCast())
1157 // If this is a GC intrinsic, avoid munging types. We need types for
1158 // statepoint reconstruction in SelectionDAG.
1159 // TODO: This is probably something which should be expanded to all
1160 // intrinsics since the entire point of intrinsics is that
1161 // they are understandable by the optimizer.
1162 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1165 // The size of ByVal or InAlloca arguments is derived from the type, so we
1166 // can't change to a type with a different size. If the size were
1167 // passed explicitly we could avoid this check.
1168 if (!CS.isByValOrInAllocaArgument(ix))
1172 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1173 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1174 if (!SrcTy->isSized() || !DstTy->isSized())
1176 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
1181 // Try to fold some different type of calls here.
1182 // Currently we're only working with the checking functions, memcpy_chk,
1183 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1184 // strcat_chk and strncat_chk.
1185 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
1186 if (!CI->getCalledFunction()) return nullptr;
1188 if (Value *With = Simplifier->optimizeCall(CI)) {
1190 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1196 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1197 // Strip off at most one level of pointer casts, looking for an alloca. This
1198 // is good enough in practice and simpler than handling any number of casts.
1199 Value *Underlying = TrampMem->stripPointerCasts();
1200 if (Underlying != TrampMem &&
1201 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1203 if (!isa<AllocaInst>(Underlying))
1206 IntrinsicInst *InitTrampoline = nullptr;
1207 for (User *U : TrampMem->users()) {
1208 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1211 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1213 // More than one init_trampoline writes to this value. Give up.
1215 InitTrampoline = II;
1218 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1219 // Allow any number of calls to adjust.trampoline.
1224 // No call to init.trampoline found.
1225 if (!InitTrampoline)
1228 // Check that the alloca is being used in the expected way.
1229 if (InitTrampoline->getOperand(0) != TrampMem)
1232 return InitTrampoline;
1235 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1237 // Visit all the previous instructions in the basic block, and try to find a
1238 // init.trampoline which has a direct path to the adjust.trampoline.
1239 for (BasicBlock::iterator I = AdjustTramp,
1240 E = AdjustTramp->getParent()->begin(); I != E; ) {
1241 Instruction *Inst = --I;
1242 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1243 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1244 II->getOperand(0) == TrampMem)
1246 if (Inst->mayWriteToMemory())
1252 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1253 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1254 // to a direct call to a function. Otherwise return NULL.
1256 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1257 Callee = Callee->stripPointerCasts();
1258 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1260 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1263 Value *TrampMem = AdjustTramp->getOperand(0);
1265 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1267 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1272 // visitCallSite - Improvements for call and invoke instructions.
1274 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1275 if (isAllocLikeFn(CS.getInstruction(), TLI))
1276 return visitAllocSite(*CS.getInstruction());
1278 bool Changed = false;
1280 // If the callee is a pointer to a function, attempt to move any casts to the
1281 // arguments of the call/invoke.
1282 Value *Callee = CS.getCalledValue();
1283 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1286 if (Function *CalleeF = dyn_cast<Function>(Callee))
1287 // If the call and callee calling conventions don't match, this call must
1288 // be unreachable, as the call is undefined.
1289 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1290 // Only do this for calls to a function with a body. A prototype may
1291 // not actually end up matching the implementation's calling conv for a
1292 // variety of reasons (e.g. it may be written in assembly).
1293 !CalleeF->isDeclaration()) {
1294 Instruction *OldCall = CS.getInstruction();
1295 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1296 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1298 // If OldCall does not return void then replaceAllUsesWith undef.
1299 // This allows ValueHandlers and custom metadata to adjust itself.
1300 if (!OldCall->getType()->isVoidTy())
1301 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1302 if (isa<CallInst>(OldCall))
1303 return EraseInstFromFunction(*OldCall);
1305 // We cannot remove an invoke, because it would change the CFG, just
1306 // change the callee to a null pointer.
1307 cast<InvokeInst>(OldCall)->setCalledFunction(
1308 Constant::getNullValue(CalleeF->getType()));
1312 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1313 // If CS does not return void then replaceAllUsesWith undef.
1314 // This allows ValueHandlers and custom metadata to adjust itself.
1315 if (!CS.getInstruction()->getType()->isVoidTy())
1316 ReplaceInstUsesWith(*CS.getInstruction(),
1317 UndefValue::get(CS.getInstruction()->getType()));
1319 if (isa<InvokeInst>(CS.getInstruction())) {
1320 // Can't remove an invoke because we cannot change the CFG.
1324 // This instruction is not reachable, just remove it. We insert a store to
1325 // undef so that we know that this code is not reachable, despite the fact
1326 // that we can't modify the CFG here.
1327 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1328 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1329 CS.getInstruction());
1331 return EraseInstFromFunction(*CS.getInstruction());
1334 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1335 return transformCallThroughTrampoline(CS, II);
1337 PointerType *PTy = cast<PointerType>(Callee->getType());
1338 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1339 if (FTy->isVarArg()) {
1340 int ix = FTy->getNumParams();
1341 // See if we can optimize any arguments passed through the varargs area of
1343 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1344 E = CS.arg_end(); I != E; ++I, ++ix) {
1345 CastInst *CI = dyn_cast<CastInst>(*I);
1346 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1347 *I = CI->getOperand(0);
1353 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1354 // Inline asm calls cannot throw - mark them 'nounwind'.
1355 CS.setDoesNotThrow();
1359 // Try to optimize the call if possible, we require DataLayout for most of
1360 // this. None of these calls are seen as possibly dead so go ahead and
1361 // delete the instruction now.
1362 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1363 Instruction *I = tryOptimizeCall(CI, DL);
1364 // If we changed something return the result, etc. Otherwise let
1365 // the fallthrough check.
1366 if (I) return EraseInstFromFunction(*I);
1369 return Changed ? CS.getInstruction() : nullptr;
1372 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1373 // attempt to move the cast to the arguments of the call/invoke.
1375 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1377 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1380 Instruction *Caller = CS.getInstruction();
1381 const AttributeSet &CallerPAL = CS.getAttributes();
1383 // Okay, this is a cast from a function to a different type. Unless doing so
1384 // would cause a type conversion of one of our arguments, change this call to
1385 // be a direct call with arguments casted to the appropriate types.
1387 FunctionType *FT = Callee->getFunctionType();
1388 Type *OldRetTy = Caller->getType();
1389 Type *NewRetTy = FT->getReturnType();
1391 // Check to see if we are changing the return type...
1392 if (OldRetTy != NewRetTy) {
1394 if (NewRetTy->isStructTy())
1395 return false; // TODO: Handle multiple return values.
1397 if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1398 if (Callee->isDeclaration())
1399 return false; // Cannot transform this return value.
1401 if (!Caller->use_empty() &&
1402 // void -> non-void is handled specially
1403 !NewRetTy->isVoidTy())
1404 return false; // Cannot transform this return value.
1407 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1408 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1410 hasAttributes(AttributeFuncs::
1411 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1412 AttributeSet::ReturnIndex))
1413 return false; // Attribute not compatible with transformed value.
1416 // If the callsite is an invoke instruction, and the return value is used by
1417 // a PHI node in a successor, we cannot change the return type of the call
1418 // because there is no place to put the cast instruction (without breaking
1419 // the critical edge). Bail out in this case.
1420 if (!Caller->use_empty())
1421 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1422 for (User *U : II->users())
1423 if (PHINode *PN = dyn_cast<PHINode>(U))
1424 if (PN->getParent() == II->getNormalDest() ||
1425 PN->getParent() == II->getUnwindDest())
1429 unsigned NumActualArgs = CS.arg_size();
1430 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1432 CallSite::arg_iterator AI = CS.arg_begin();
1433 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1434 Type *ParamTy = FT->getParamType(i);
1435 Type *ActTy = (*AI)->getType();
1437 if (!CastInst::isBitCastable(ActTy, ParamTy))
1438 return false; // Cannot transform this parameter value.
1440 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1441 hasAttributes(AttributeFuncs::
1442 typeIncompatible(ParamTy, i + 1), i + 1))
1443 return false; // Attribute not compatible with transformed value.
1445 if (CS.isInAllocaArgument(i))
1446 return false; // Cannot transform to and from inalloca.
1448 // If the parameter is passed as a byval argument, then we have to have a
1449 // sized type and the sized type has to have the same size as the old type.
1450 if (ParamTy != ActTy &&
1451 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1452 Attribute::ByVal)) {
1453 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1454 if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1457 Type *CurElTy = ActTy->getPointerElementType();
1458 if (DL->getTypeAllocSize(CurElTy) !=
1459 DL->getTypeAllocSize(ParamPTy->getElementType()))
1464 if (Callee->isDeclaration()) {
1465 // Do not delete arguments unless we have a function body.
1466 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1469 // If the callee is just a declaration, don't change the varargsness of the
1470 // call. We don't want to introduce a varargs call where one doesn't
1472 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1473 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1476 // If both the callee and the cast type are varargs, we still have to make
1477 // sure the number of fixed parameters are the same or we have the same
1478 // ABI issues as if we introduce a varargs call.
1479 if (FT->isVarArg() &&
1480 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1481 FT->getNumParams() !=
1482 cast<FunctionType>(APTy->getElementType())->getNumParams())
1486 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1487 !CallerPAL.isEmpty())
1488 // In this case we have more arguments than the new function type, but we
1489 // won't be dropping them. Check that these extra arguments have attributes
1490 // that are compatible with being a vararg call argument.
1491 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1492 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1493 if (Index <= FT->getNumParams())
1496 // Check if it has an attribute that's incompatible with varargs.
1497 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1498 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1503 // Okay, we decided that this is a safe thing to do: go ahead and start
1504 // inserting cast instructions as necessary.
1505 std::vector<Value*> Args;
1506 Args.reserve(NumActualArgs);
1507 SmallVector<AttributeSet, 8> attrVec;
1508 attrVec.reserve(NumCommonArgs);
1510 // Get any return attributes.
1511 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1513 // If the return value is not being used, the type may not be compatible
1514 // with the existing attributes. Wipe out any problematic attributes.
1516 removeAttributes(AttributeFuncs::
1517 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1518 AttributeSet::ReturnIndex);
1520 // Add the new return attributes.
1521 if (RAttrs.hasAttributes())
1522 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1523 AttributeSet::ReturnIndex, RAttrs));
1525 AI = CS.arg_begin();
1526 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1527 Type *ParamTy = FT->getParamType(i);
1529 if ((*AI)->getType() == ParamTy) {
1530 Args.push_back(*AI);
1532 Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1535 // Add any parameter attributes.
1536 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1537 if (PAttrs.hasAttributes())
1538 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1542 // If the function takes more arguments than the call was taking, add them
1544 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1545 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1547 // If we are removing arguments to the function, emit an obnoxious warning.
1548 if (FT->getNumParams() < NumActualArgs) {
1549 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1550 if (FT->isVarArg()) {
1551 // Add all of the arguments in their promoted form to the arg list.
1552 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1553 Type *PTy = getPromotedType((*AI)->getType());
1554 if (PTy != (*AI)->getType()) {
1555 // Must promote to pass through va_arg area!
1556 Instruction::CastOps opcode =
1557 CastInst::getCastOpcode(*AI, false, PTy, false);
1558 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1560 Args.push_back(*AI);
1563 // Add any parameter attributes.
1564 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1565 if (PAttrs.hasAttributes())
1566 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1572 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1573 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1574 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1576 if (NewRetTy->isVoidTy())
1577 Caller->setName(""); // Void type should not have a name.
1579 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1583 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1584 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1585 II->getUnwindDest(), Args);
1587 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1588 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1590 CallInst *CI = cast<CallInst>(Caller);
1591 NC = Builder->CreateCall(Callee, Args);
1593 if (CI->isTailCall())
1594 cast<CallInst>(NC)->setTailCall();
1595 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1596 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1599 // Insert a cast of the return type as necessary.
1601 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1602 if (!NV->getType()->isVoidTy()) {
1603 NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1604 NC->setDebugLoc(Caller->getDebugLoc());
1606 // If this is an invoke instruction, we should insert it after the first
1607 // non-phi, instruction in the normal successor block.
1608 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1609 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1610 InsertNewInstBefore(NC, *I);
1612 // Otherwise, it's a call, just insert cast right after the call.
1613 InsertNewInstBefore(NC, *Caller);
1615 Worklist.AddUsersToWorkList(*Caller);
1617 NV = UndefValue::get(Caller->getType());
1621 if (!Caller->use_empty())
1622 ReplaceInstUsesWith(*Caller, NV);
1623 else if (Caller->hasValueHandle()) {
1624 if (OldRetTy == NV->getType())
1625 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1627 // We cannot call ValueIsRAUWd with a different type, and the
1628 // actual tracked value will disappear.
1629 ValueHandleBase::ValueIsDeleted(Caller);
1632 EraseInstFromFunction(*Caller);
1636 // transformCallThroughTrampoline - Turn a call to a function created by
1637 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1638 // underlying function.
1641 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1642 IntrinsicInst *Tramp) {
1643 Value *Callee = CS.getCalledValue();
1644 PointerType *PTy = cast<PointerType>(Callee->getType());
1645 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1646 const AttributeSet &Attrs = CS.getAttributes();
1648 // If the call already has the 'nest' attribute somewhere then give up -
1649 // otherwise 'nest' would occur twice after splicing in the chain.
1650 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1654 "transformCallThroughTrampoline called with incorrect CallSite.");
1656 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1657 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1658 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1660 const AttributeSet &NestAttrs = NestF->getAttributes();
1661 if (!NestAttrs.isEmpty()) {
1662 unsigned NestIdx = 1;
1663 Type *NestTy = nullptr;
1664 AttributeSet NestAttr;
1666 // Look for a parameter marked with the 'nest' attribute.
1667 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1668 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1669 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1670 // Record the parameter type and any other attributes.
1672 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1677 Instruction *Caller = CS.getInstruction();
1678 std::vector<Value*> NewArgs;
1679 NewArgs.reserve(CS.arg_size() + 1);
1681 SmallVector<AttributeSet, 8> NewAttrs;
1682 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1684 // Insert the nest argument into the call argument list, which may
1685 // mean appending it. Likewise for attributes.
1687 // Add any result attributes.
1688 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1689 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1690 Attrs.getRetAttributes()));
1694 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1696 if (Idx == NestIdx) {
1697 // Add the chain argument and attributes.
1698 Value *NestVal = Tramp->getArgOperand(2);
1699 if (NestVal->getType() != NestTy)
1700 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1701 NewArgs.push_back(NestVal);
1702 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1709 // Add the original argument and attributes.
1710 NewArgs.push_back(*I);
1711 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1712 if (Attr.hasAttributes(Idx)) {
1713 AttrBuilder B(Attr, Idx);
1714 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1715 Idx + (Idx >= NestIdx), B));
1722 // Add any function attributes.
1723 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1724 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1725 Attrs.getFnAttributes()));
1727 // The trampoline may have been bitcast to a bogus type (FTy).
1728 // Handle this by synthesizing a new function type, equal to FTy
1729 // with the chain parameter inserted.
1731 std::vector<Type*> NewTypes;
1732 NewTypes.reserve(FTy->getNumParams()+1);
1734 // Insert the chain's type into the list of parameter types, which may
1735 // mean appending it.
1738 FunctionType::param_iterator I = FTy->param_begin(),
1739 E = FTy->param_end();
1743 // Add the chain's type.
1744 NewTypes.push_back(NestTy);
1749 // Add the original type.
1750 NewTypes.push_back(*I);
1756 // Replace the trampoline call with a direct call. Let the generic
1757 // code sort out any function type mismatches.
1758 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1760 Constant *NewCallee =
1761 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1762 NestF : ConstantExpr::getBitCast(NestF,
1763 PointerType::getUnqual(NewFTy));
1764 const AttributeSet &NewPAL =
1765 AttributeSet::get(FTy->getContext(), NewAttrs);
1767 Instruction *NewCaller;
1768 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1769 NewCaller = InvokeInst::Create(NewCallee,
1770 II->getNormalDest(), II->getUnwindDest(),
1772 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1773 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1775 NewCaller = CallInst::Create(NewCallee, NewArgs);
1776 if (cast<CallInst>(Caller)->isTailCall())
1777 cast<CallInst>(NewCaller)->setTailCall();
1778 cast<CallInst>(NewCaller)->
1779 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1780 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1787 // Replace the trampoline call with a direct call. Since there is no 'nest'
1788 // parameter, there is no need to adjust the argument list. Let the generic
1789 // code sort out any function type mismatches.
1790 Constant *NewCallee =
1791 NestF->getType() == PTy ? NestF :
1792 ConstantExpr::getBitCast(NestF, PTy);
1793 CS.setCalledFunction(NewCallee);
1794 return CS.getInstruction();