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 "InstCombineInternal.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/MemoryBuiltins.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/Dominators.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/IR/Statepoint.h"
21 #include "llvm/Transforms/Utils/BuildLibCalls.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/SimplifyLibCalls.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, MI, AC, DT);
64 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, AC, 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 (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, MI, AC, 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 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit
201 /// source vectors, unless a zero bit is set. If a zero bit is set,
202 /// then ignore that half of the mask and clear that half of the vector.
203 static Value *SimplifyX86vperm2(const IntrinsicInst &II,
204 InstCombiner::BuilderTy &Builder) {
205 if (auto CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
206 VectorType *VecTy = cast<VectorType>(II.getType());
207 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
209 // The immediate permute control byte looks like this:
210 // [1:0] - select 128 bits from sources for low half of destination
212 // [3] - zero low half of destination
213 // [5:4] - select 128 bits from sources for high half of destination
215 // [7] - zero high half of destination
217 uint8_t Imm = CInt->getZExtValue();
219 bool LowHalfZero = Imm & 0x08;
220 bool HighHalfZero = Imm & 0x80;
222 // If both zero mask bits are set, this was just a weird way to
223 // generate a zero vector.
224 if (LowHalfZero && HighHalfZero)
227 // If 0 or 1 zero mask bits are set, this is a simple shuffle.
228 unsigned NumElts = VecTy->getNumElements();
229 unsigned HalfSize = NumElts / 2;
230 SmallVector<int, 8> ShuffleMask(NumElts);
232 // The high bit of the selection field chooses the 1st or 2nd operand.
233 bool LowInputSelect = Imm & 0x02;
234 bool HighInputSelect = Imm & 0x20;
236 // The low bit of the selection field chooses the low or high half
237 // of the selected operand.
238 bool LowHalfSelect = Imm & 0x01;
239 bool HighHalfSelect = Imm & 0x10;
241 // Determine which operand(s) are actually in use for this instruction.
242 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
243 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0);
245 // If needed, replace operands based on zero mask.
246 V0 = LowHalfZero ? ZeroVector : V0;
247 V1 = HighHalfZero ? ZeroVector : V1;
249 // Permute low half of result.
250 unsigned StartIndex = LowHalfSelect ? HalfSize : 0;
251 for (unsigned i = 0; i < HalfSize; ++i)
252 ShuffleMask[i] = StartIndex + i;
254 // Permute high half of result.
255 StartIndex = HighHalfSelect ? HalfSize : 0;
256 StartIndex += NumElts;
257 for (unsigned i = 0; i < HalfSize; ++i)
258 ShuffleMask[i + HalfSize] = StartIndex + i;
260 return Builder.CreateShuffleVector(V0, V1, ShuffleMask);
265 /// visitCallInst - CallInst simplification. This mostly only handles folding
266 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
267 /// the heavy lifting.
269 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
270 if (isFreeCall(&CI, TLI))
271 return visitFree(CI);
273 // If the caller function is nounwind, mark the call as nounwind, even if the
275 if (CI.getParent()->getParent()->doesNotThrow() &&
276 !CI.doesNotThrow()) {
277 CI.setDoesNotThrow();
281 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
282 if (!II) return visitCallSite(&CI);
284 // Intrinsics cannot occur in an invoke, so handle them here instead of in
286 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
287 bool Changed = false;
289 // memmove/cpy/set of zero bytes is a noop.
290 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
291 if (NumBytes->isNullValue())
292 return EraseInstFromFunction(CI);
294 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
295 if (CI->getZExtValue() == 1) {
296 // Replace the instruction with just byte operations. We would
297 // transform other cases to loads/stores, but we don't know if
298 // alignment is sufficient.
302 // No other transformations apply to volatile transfers.
303 if (MI->isVolatile())
306 // If we have a memmove and the source operation is a constant global,
307 // then the source and dest pointers can't alias, so we can change this
308 // into a call to memcpy.
309 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
310 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
311 if (GVSrc->isConstant()) {
312 Module *M = CI.getParent()->getParent()->getParent();
313 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
314 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
315 CI.getArgOperand(1)->getType(),
316 CI.getArgOperand(2)->getType() };
317 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
322 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
323 // memmove(x,x,size) -> noop.
324 if (MTI->getSource() == MTI->getDest())
325 return EraseInstFromFunction(CI);
328 // If we can determine a pointer alignment that is bigger than currently
329 // set, update the alignment.
330 if (isa<MemTransferInst>(MI)) {
331 if (Instruction *I = SimplifyMemTransfer(MI))
333 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
334 if (Instruction *I = SimplifyMemSet(MSI))
338 if (Changed) return II;
341 switch (II->getIntrinsicID()) {
343 case Intrinsic::objectsize: {
345 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
346 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
349 case Intrinsic::bswap: {
350 Value *IIOperand = II->getArgOperand(0);
353 // bswap(bswap(x)) -> x
354 if (match(IIOperand, m_BSwap(m_Value(X))))
355 return ReplaceInstUsesWith(CI, X);
357 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
358 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
359 unsigned C = X->getType()->getPrimitiveSizeInBits() -
360 IIOperand->getType()->getPrimitiveSizeInBits();
361 Value *CV = ConstantInt::get(X->getType(), C);
362 Value *V = Builder->CreateLShr(X, CV);
363 return new TruncInst(V, IIOperand->getType());
368 case Intrinsic::powi:
369 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
372 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
375 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
376 // powi(x, -1) -> 1/x
377 if (Power->isAllOnesValue())
378 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
379 II->getArgOperand(0));
382 case Intrinsic::cttz: {
383 // If all bits below the first known one are known zero,
384 // this value is constant.
385 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
386 // FIXME: Try to simplify vectors of integers.
388 uint32_t BitWidth = IT->getBitWidth();
389 APInt KnownZero(BitWidth, 0);
390 APInt KnownOne(BitWidth, 0);
391 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
392 unsigned TrailingZeros = KnownOne.countTrailingZeros();
393 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
394 if ((Mask & KnownZero) == Mask)
395 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
396 APInt(BitWidth, TrailingZeros)));
400 case Intrinsic::ctlz: {
401 // If all bits above the first known one are known zero,
402 // this value is constant.
403 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
404 // FIXME: Try to simplify vectors of integers.
406 uint32_t BitWidth = IT->getBitWidth();
407 APInt KnownZero(BitWidth, 0);
408 APInt KnownOne(BitWidth, 0);
409 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
410 unsigned LeadingZeros = KnownOne.countLeadingZeros();
411 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
412 if ((Mask & KnownZero) == Mask)
413 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
414 APInt(BitWidth, LeadingZeros)));
418 case Intrinsic::uadd_with_overflow: {
419 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
420 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, II);
421 if (OR == OverflowResult::NeverOverflows)
422 return CreateOverflowTuple(II, Builder->CreateNUWAdd(LHS, RHS), false);
423 if (OR == OverflowResult::AlwaysOverflows)
424 return CreateOverflowTuple(II, Builder->CreateAdd(LHS, RHS), true);
426 // FALL THROUGH uadd into sadd
427 case Intrinsic::sadd_with_overflow:
428 // Canonicalize constants into the RHS.
429 if (isa<Constant>(II->getArgOperand(0)) &&
430 !isa<Constant>(II->getArgOperand(1))) {
431 Value *LHS = II->getArgOperand(0);
432 II->setArgOperand(0, II->getArgOperand(1));
433 II->setArgOperand(1, LHS);
437 // X + undef -> undef
438 if (isa<UndefValue>(II->getArgOperand(1)))
439 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
441 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
442 // X + 0 -> {X, false}
444 return CreateOverflowTuple(II, II->getArgOperand(0), false,
449 // We can strength reduce reduce this signed add into a regular add if we
450 // can prove that it will never overflow.
451 if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow) {
452 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
453 if (WillNotOverflowSignedAdd(LHS, RHS, *II)) {
454 return CreateOverflowTuple(II, Builder->CreateNSWAdd(LHS, RHS), false);
459 case Intrinsic::usub_with_overflow:
460 case Intrinsic::ssub_with_overflow: {
461 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
462 // undef - X -> undef
463 // X - undef -> undef
464 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
465 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
467 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
468 // X - 0 -> {X, false}
469 if (ConstRHS->isZero()) {
470 return CreateOverflowTuple(II, LHS, false, /*ReUseName*/false);
473 if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
474 if (WillNotOverflowSignedSub(LHS, RHS, *II)) {
475 return CreateOverflowTuple(II, Builder->CreateNSWSub(LHS, RHS), false);
478 if (WillNotOverflowUnsignedSub(LHS, RHS, *II)) {
479 return CreateOverflowTuple(II, Builder->CreateNUWSub(LHS, RHS), false);
484 case Intrinsic::umul_with_overflow: {
485 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
486 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, II);
487 if (OR == OverflowResult::NeverOverflows)
488 return CreateOverflowTuple(II, Builder->CreateNUWMul(LHS, RHS), false);
489 if (OR == OverflowResult::AlwaysOverflows)
490 return CreateOverflowTuple(II, Builder->CreateMul(LHS, RHS), true);
492 case Intrinsic::smul_with_overflow:
493 // Canonicalize constants into the RHS.
494 if (isa<Constant>(II->getArgOperand(0)) &&
495 !isa<Constant>(II->getArgOperand(1))) {
496 Value *LHS = II->getArgOperand(0);
497 II->setArgOperand(0, II->getArgOperand(1));
498 II->setArgOperand(1, LHS);
502 // X * undef -> undef
503 if (isa<UndefValue>(II->getArgOperand(1)))
504 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
506 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
509 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
511 // X * 1 -> {X, false}
512 if (RHSI->equalsInt(1)) {
513 return CreateOverflowTuple(II, II->getArgOperand(0), false,
517 if (II->getIntrinsicID() == Intrinsic::smul_with_overflow) {
518 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
519 if (WillNotOverflowSignedMul(LHS, RHS, *II)) {
520 return CreateOverflowTuple(II, Builder->CreateNSWMul(LHS, RHS), false);
524 case Intrinsic::minnum:
525 case Intrinsic::maxnum: {
526 Value *Arg0 = II->getArgOperand(0);
527 Value *Arg1 = II->getArgOperand(1);
531 return ReplaceInstUsesWith(CI, Arg0);
533 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
534 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
536 // Canonicalize constants into the RHS.
538 II->setArgOperand(0, Arg1);
539 II->setArgOperand(1, Arg0);
544 if (C1 && C1->isNaN())
545 return ReplaceInstUsesWith(CI, Arg0);
547 // This is the value because if undef were NaN, we would return the other
548 // value and cannot return a NaN unless both operands are.
550 // fmin(undef, x) -> x
551 if (isa<UndefValue>(Arg0))
552 return ReplaceInstUsesWith(CI, Arg1);
554 // fmin(x, undef) -> x
555 if (isa<UndefValue>(Arg1))
556 return ReplaceInstUsesWith(CI, Arg0);
560 if (II->getIntrinsicID() == Intrinsic::minnum) {
561 // fmin(x, fmin(x, y)) -> fmin(x, y)
562 // fmin(y, fmin(x, y)) -> fmin(x, y)
563 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
564 if (Arg0 == X || Arg0 == Y)
565 return ReplaceInstUsesWith(CI, Arg1);
568 // fmin(fmin(x, y), x) -> fmin(x, y)
569 // fmin(fmin(x, y), y) -> fmin(x, y)
570 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
571 if (Arg1 == X || Arg1 == Y)
572 return ReplaceInstUsesWith(CI, Arg0);
575 // TODO: fmin(nnan x, inf) -> x
576 // TODO: fmin(nnan ninf x, flt_max) -> x
577 if (C1 && C1->isInfinity()) {
578 // fmin(x, -inf) -> -inf
579 if (C1->isNegative())
580 return ReplaceInstUsesWith(CI, Arg1);
583 assert(II->getIntrinsicID() == Intrinsic::maxnum);
584 // fmax(x, fmax(x, y)) -> fmax(x, y)
585 // fmax(y, fmax(x, y)) -> fmax(x, y)
586 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
587 if (Arg0 == X || Arg0 == Y)
588 return ReplaceInstUsesWith(CI, Arg1);
591 // fmax(fmax(x, y), x) -> fmax(x, y)
592 // fmax(fmax(x, y), y) -> fmax(x, y)
593 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
594 if (Arg1 == X || Arg1 == Y)
595 return ReplaceInstUsesWith(CI, Arg0);
598 // TODO: fmax(nnan x, -inf) -> x
599 // TODO: fmax(nnan ninf x, -flt_max) -> x
600 if (C1 && C1->isInfinity()) {
601 // fmax(x, inf) -> inf
602 if (!C1->isNegative())
603 return ReplaceInstUsesWith(CI, Arg1);
608 case Intrinsic::ppc_altivec_lvx:
609 case Intrinsic::ppc_altivec_lvxl:
610 // Turn PPC lvx -> load if the pointer is known aligned.
611 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
613 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
614 PointerType::getUnqual(II->getType()));
615 return new LoadInst(Ptr);
618 case Intrinsic::ppc_vsx_lxvw4x:
619 case Intrinsic::ppc_vsx_lxvd2x: {
620 // Turn PPC VSX loads into normal loads.
621 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
622 PointerType::getUnqual(II->getType()));
623 return new LoadInst(Ptr, Twine(""), false, 1);
625 case Intrinsic::ppc_altivec_stvx:
626 case Intrinsic::ppc_altivec_stvxl:
627 // Turn stvx -> store if the pointer is known aligned.
628 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
631 PointerType::getUnqual(II->getArgOperand(0)->getType());
632 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
633 return new StoreInst(II->getArgOperand(0), Ptr);
636 case Intrinsic::ppc_vsx_stxvw4x:
637 case Intrinsic::ppc_vsx_stxvd2x: {
638 // Turn PPC VSX stores into normal stores.
639 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
640 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
641 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
643 case Intrinsic::ppc_qpx_qvlfs:
644 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
645 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
647 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
648 PointerType::getUnqual(II->getType()));
649 return new LoadInst(Ptr);
652 case Intrinsic::ppc_qpx_qvlfd:
653 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
654 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >=
656 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
657 PointerType::getUnqual(II->getType()));
658 return new LoadInst(Ptr);
661 case Intrinsic::ppc_qpx_qvstfs:
662 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
663 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >=
666 PointerType::getUnqual(II->getArgOperand(0)->getType());
667 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
668 return new StoreInst(II->getArgOperand(0), Ptr);
671 case Intrinsic::ppc_qpx_qvstfd:
672 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
673 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >=
676 PointerType::getUnqual(II->getArgOperand(0)->getType());
677 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
678 return new StoreInst(II->getArgOperand(0), Ptr);
681 case Intrinsic::x86_sse_storeu_ps:
682 case Intrinsic::x86_sse2_storeu_pd:
683 case Intrinsic::x86_sse2_storeu_dq:
684 // Turn X86 storeu -> store if the pointer is known aligned.
685 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >=
688 PointerType::getUnqual(II->getArgOperand(1)->getType());
689 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
690 return new StoreInst(II->getArgOperand(1), Ptr);
694 case Intrinsic::x86_sse_cvtss2si:
695 case Intrinsic::x86_sse_cvtss2si64:
696 case Intrinsic::x86_sse_cvttss2si:
697 case Intrinsic::x86_sse_cvttss2si64:
698 case Intrinsic::x86_sse2_cvtsd2si:
699 case Intrinsic::x86_sse2_cvtsd2si64:
700 case Intrinsic::x86_sse2_cvttsd2si:
701 case Intrinsic::x86_sse2_cvttsd2si64: {
702 // These intrinsics only demand the 0th element of their input vectors. If
703 // we can simplify the input based on that, do so now.
705 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
706 APInt DemandedElts(VWidth, 1);
707 APInt UndefElts(VWidth, 0);
708 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
709 DemandedElts, UndefElts)) {
710 II->setArgOperand(0, V);
716 // Constant fold <A x Bi> << Ci.
717 // FIXME: We don't handle _dq because it's a shift of an i128, but is
718 // represented in the IR as <2 x i64>. A per element shift is wrong.
719 case Intrinsic::x86_sse2_psll_d:
720 case Intrinsic::x86_sse2_psll_q:
721 case Intrinsic::x86_sse2_psll_w:
722 case Intrinsic::x86_sse2_pslli_d:
723 case Intrinsic::x86_sse2_pslli_q:
724 case Intrinsic::x86_sse2_pslli_w:
725 case Intrinsic::x86_avx2_psll_d:
726 case Intrinsic::x86_avx2_psll_q:
727 case Intrinsic::x86_avx2_psll_w:
728 case Intrinsic::x86_avx2_pslli_d:
729 case Intrinsic::x86_avx2_pslli_q:
730 case Intrinsic::x86_avx2_pslli_w:
731 case Intrinsic::x86_sse2_psrl_d:
732 case Intrinsic::x86_sse2_psrl_q:
733 case Intrinsic::x86_sse2_psrl_w:
734 case Intrinsic::x86_sse2_psrli_d:
735 case Intrinsic::x86_sse2_psrli_q:
736 case Intrinsic::x86_sse2_psrli_w:
737 case Intrinsic::x86_avx2_psrl_d:
738 case Intrinsic::x86_avx2_psrl_q:
739 case Intrinsic::x86_avx2_psrl_w:
740 case Intrinsic::x86_avx2_psrli_d:
741 case Intrinsic::x86_avx2_psrli_q:
742 case Intrinsic::x86_avx2_psrli_w: {
743 // Simplify if count is constant. To 0 if >= BitWidth,
744 // otherwise to shl/lshr.
745 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
746 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
751 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
755 auto Vec = II->getArgOperand(0);
756 auto VT = cast<VectorType>(Vec->getType());
757 if (Count->getZExtValue() >
758 VT->getElementType()->getPrimitiveSizeInBits() - 1)
759 return ReplaceInstUsesWith(
760 CI, ConstantAggregateZero::get(Vec->getType()));
762 bool isPackedShiftLeft = true;
763 switch (II->getIntrinsicID()) {
765 case Intrinsic::x86_sse2_psrl_d:
766 case Intrinsic::x86_sse2_psrl_q:
767 case Intrinsic::x86_sse2_psrl_w:
768 case Intrinsic::x86_sse2_psrli_d:
769 case Intrinsic::x86_sse2_psrli_q:
770 case Intrinsic::x86_sse2_psrli_w:
771 case Intrinsic::x86_avx2_psrl_d:
772 case Intrinsic::x86_avx2_psrl_q:
773 case Intrinsic::x86_avx2_psrl_w:
774 case Intrinsic::x86_avx2_psrli_d:
775 case Intrinsic::x86_avx2_psrli_q:
776 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
779 unsigned VWidth = VT->getNumElements();
780 // Get a constant vector of the same type as the first operand.
781 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
782 if (isPackedShiftLeft)
783 return BinaryOperator::CreateShl(Vec,
784 Builder->CreateVectorSplat(VWidth, VTCI));
786 return BinaryOperator::CreateLShr(Vec,
787 Builder->CreateVectorSplat(VWidth, VTCI));
790 case Intrinsic::x86_sse41_pmovsxbw:
791 case Intrinsic::x86_sse41_pmovsxwd:
792 case Intrinsic::x86_sse41_pmovsxdq:
793 case Intrinsic::x86_sse41_pmovzxbw:
794 case Intrinsic::x86_sse41_pmovzxwd:
795 case Intrinsic::x86_sse41_pmovzxdq: {
796 // pmov{s|z}x ignores the upper half of their input vectors.
798 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
799 unsigned LowHalfElts = VWidth / 2;
800 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
801 APInt UndefElts(VWidth, 0);
802 if (Value *TmpV = SimplifyDemandedVectorElts(
803 II->getArgOperand(0), InputDemandedElts, UndefElts)) {
804 II->setArgOperand(0, TmpV);
810 case Intrinsic::x86_sse4a_insertqi: {
811 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
813 // TODO: eventually we should lower this intrinsic to IR
814 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
815 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
816 unsigned Index = CIStart->getZExtValue();
817 // From AMD documentation: "a value of zero in the field length is
818 // defined as length of 64".
819 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
821 // From AMD documentation: "If the sum of the bit index + length field
822 // is greater than 64, the results are undefined".
824 // Note that both field index and field length are 8-bit quantities.
825 // Since variables 'Index' and 'Length' are unsigned values
826 // obtained from zero-extending field index and field length
827 // respectively, their sum should never wrap around.
828 if ((Index + Length) > 64)
829 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
831 if (Length == 64 && Index == 0) {
832 Value *Vec = II->getArgOperand(1);
833 Value *Undef = UndefValue::get(Vec->getType());
834 const uint32_t Mask[] = { 0, 2 };
835 return ReplaceInstUsesWith(
837 Builder->CreateShuffleVector(
838 Vec, Undef, ConstantDataVector::get(
839 II->getContext(), makeArrayRef(Mask))));
841 } else if (auto Source =
842 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
843 if (Source->hasOneUse() &&
844 Source->getArgOperand(1) == II->getArgOperand(1)) {
845 // If the source of the insert has only one use and it's another
846 // insert (and they're both inserting from the same vector), try to
847 // bundle both together.
849 dyn_cast<ConstantInt>(Source->getArgOperand(2));
851 dyn_cast<ConstantInt>(Source->getArgOperand(3));
852 if (CISourceStart && CISourceWidth) {
853 unsigned Start = CIStart->getZExtValue();
854 unsigned Width = CIWidth->getZExtValue();
855 unsigned End = Start + Width;
856 unsigned SourceStart = CISourceStart->getZExtValue();
857 unsigned SourceWidth = CISourceWidth->getZExtValue();
858 unsigned SourceEnd = SourceStart + SourceWidth;
859 unsigned NewStart, NewWidth;
860 bool ShouldReplace = false;
861 if (Start <= SourceStart && SourceStart <= End) {
863 NewWidth = std::max(End, SourceEnd) - NewStart;
864 ShouldReplace = true;
865 } else if (SourceStart <= Start && Start <= SourceEnd) {
866 NewStart = SourceStart;
867 NewWidth = std::max(SourceEnd, End) - NewStart;
868 ShouldReplace = true;
872 Constant *ConstantWidth = ConstantInt::get(
873 II->getArgOperand(2)->getType(), NewWidth, false);
874 Constant *ConstantStart = ConstantInt::get(
875 II->getArgOperand(3)->getType(), NewStart, false);
876 Value *Args[4] = { Source->getArgOperand(0),
877 II->getArgOperand(1), ConstantWidth,
879 Module *M = CI.getParent()->getParent()->getParent();
881 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
882 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
892 case Intrinsic::x86_sse41_pblendvb:
893 case Intrinsic::x86_sse41_blendvps:
894 case Intrinsic::x86_sse41_blendvpd:
895 case Intrinsic::x86_avx_blendv_ps_256:
896 case Intrinsic::x86_avx_blendv_pd_256:
897 case Intrinsic::x86_avx2_pblendvb: {
898 // Convert blendv* to vector selects if the mask is constant.
899 // This optimization is convoluted because the intrinsic is defined as
900 // getting a vector of floats or doubles for the ps and pd versions.
901 // FIXME: That should be changed.
902 Value *Mask = II->getArgOperand(2);
903 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
904 auto Tyi1 = Builder->getInt1Ty();
905 auto SelectorType = cast<VectorType>(Mask->getType());
906 auto EltTy = SelectorType->getElementType();
907 unsigned Size = SelectorType->getNumElements();
911 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
912 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
913 "Wrong arguments for variable blend intrinsic");
914 SmallVector<Constant *, 32> Selectors;
915 for (unsigned I = 0; I < Size; ++I) {
916 // The intrinsics only read the top bit
919 Selector = C->getElementAsInteger(I);
921 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
922 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
924 auto NewSelector = ConstantVector::get(Selectors);
925 return SelectInst::Create(NewSelector, II->getArgOperand(1),
926 II->getArgOperand(0), "blendv");
932 case Intrinsic::x86_avx_vpermilvar_ps:
933 case Intrinsic::x86_avx_vpermilvar_ps_256:
934 case Intrinsic::x86_avx_vpermilvar_pd:
935 case Intrinsic::x86_avx_vpermilvar_pd_256: {
936 // Convert vpermil* to shufflevector if the mask is constant.
937 Value *V = II->getArgOperand(1);
938 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
939 assert(Size == 8 || Size == 4 || Size == 2);
941 if (auto C = dyn_cast<ConstantDataVector>(V)) {
942 // The intrinsics only read one or two bits, clear the rest.
943 for (unsigned I = 0; I < Size; ++I) {
944 uint32_t Index = C->getElementAsInteger(I) & 0x3;
945 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
946 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
950 } else if (isa<ConstantAggregateZero>(V)) {
951 for (unsigned I = 0; I < Size; ++I)
956 // The _256 variants are a bit trickier since the mask bits always index
957 // into the corresponding 128 half. In order to convert to a generic
958 // shuffle, we have to make that explicit.
959 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
960 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
961 for (unsigned I = Size / 2; I < Size; ++I)
962 Indexes[I] += Size / 2;
965 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
966 auto V1 = II->getArgOperand(0);
967 auto V2 = UndefValue::get(V1->getType());
968 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
969 return ReplaceInstUsesWith(CI, Shuffle);
972 case Intrinsic::x86_avx_vperm2f128_pd_256:
973 case Intrinsic::x86_avx_vperm2f128_ps_256:
974 case Intrinsic::x86_avx_vperm2f128_si_256:
975 // TODO: Add the AVX2 version of this instruction.
976 if (Value *V = SimplifyX86vperm2(*II, *Builder))
977 return ReplaceInstUsesWith(*II, V);
980 case Intrinsic::ppc_altivec_vperm:
981 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
982 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
983 // a vectorshuffle for little endian, we must undo the transformation
984 // performed on vec_perm in altivec.h. That is, we must complement
985 // the permutation mask with respect to 31 and reverse the order of
987 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
988 assert(Mask->getType()->getVectorNumElements() == 16 &&
989 "Bad type for intrinsic!");
991 // Check that all of the elements are integer constants or undefs.
992 bool AllEltsOk = true;
993 for (unsigned i = 0; i != 16; ++i) {
994 Constant *Elt = Mask->getAggregateElement(i);
995 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
1002 // Cast the input vectors to byte vectors.
1003 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
1005 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
1007 Value *Result = UndefValue::get(Op0->getType());
1009 // Only extract each element once.
1010 Value *ExtractedElts[32];
1011 memset(ExtractedElts, 0, sizeof(ExtractedElts));
1013 for (unsigned i = 0; i != 16; ++i) {
1014 if (isa<UndefValue>(Mask->getAggregateElement(i)))
1017 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
1018 Idx &= 31; // Match the hardware behavior.
1019 if (DL.isLittleEndian())
1022 if (!ExtractedElts[Idx]) {
1023 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
1024 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
1025 ExtractedElts[Idx] =
1026 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
1027 Builder->getInt32(Idx&15));
1030 // Insert this value into the result vector.
1031 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
1032 Builder->getInt32(i));
1034 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
1039 case Intrinsic::arm_neon_vld1:
1040 case Intrinsic::arm_neon_vld2:
1041 case Intrinsic::arm_neon_vld3:
1042 case Intrinsic::arm_neon_vld4:
1043 case Intrinsic::arm_neon_vld2lane:
1044 case Intrinsic::arm_neon_vld3lane:
1045 case Intrinsic::arm_neon_vld4lane:
1046 case Intrinsic::arm_neon_vst1:
1047 case Intrinsic::arm_neon_vst2:
1048 case Intrinsic::arm_neon_vst3:
1049 case Intrinsic::arm_neon_vst4:
1050 case Intrinsic::arm_neon_vst2lane:
1051 case Intrinsic::arm_neon_vst3lane:
1052 case Intrinsic::arm_neon_vst4lane: {
1053 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT);
1054 unsigned AlignArg = II->getNumArgOperands() - 1;
1055 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
1056 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
1057 II->setArgOperand(AlignArg,
1058 ConstantInt::get(Type::getInt32Ty(II->getContext()),
1065 case Intrinsic::arm_neon_vmulls:
1066 case Intrinsic::arm_neon_vmullu:
1067 case Intrinsic::aarch64_neon_smull:
1068 case Intrinsic::aarch64_neon_umull: {
1069 Value *Arg0 = II->getArgOperand(0);
1070 Value *Arg1 = II->getArgOperand(1);
1072 // Handle mul by zero first:
1073 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1074 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1077 // Check for constant LHS & RHS - in this case we just simplify.
1078 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1079 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1080 VectorType *NewVT = cast<VectorType>(II->getType());
1081 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1082 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1083 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1084 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1086 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1089 // Couldn't simplify - canonicalize constant to the RHS.
1090 std::swap(Arg0, Arg1);
1093 // Handle mul by one:
1094 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1095 if (ConstantInt *Splat =
1096 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1098 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1099 /*isSigned=*/!Zext);
1104 case Intrinsic::AMDGPU_rcp: {
1105 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1106 const APFloat &ArgVal = C->getValueAPF();
1107 APFloat Val(ArgVal.getSemantics(), 1.0);
1108 APFloat::opStatus Status = Val.divide(ArgVal,
1109 APFloat::rmNearestTiesToEven);
1110 // Only do this if it was exact and therefore not dependent on the
1112 if (Status == APFloat::opOK)
1113 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1118 case Intrinsic::stackrestore: {
1119 // If the save is right next to the restore, remove the restore. This can
1120 // happen when variable allocas are DCE'd.
1121 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1122 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1123 BasicBlock::iterator BI = SS;
1125 return EraseInstFromFunction(CI);
1129 // Scan down this block to see if there is another stack restore in the
1130 // same block without an intervening call/alloca.
1131 BasicBlock::iterator BI = II;
1132 TerminatorInst *TI = II->getParent()->getTerminator();
1133 bool CannotRemove = false;
1134 for (++BI; &*BI != TI; ++BI) {
1135 if (isa<AllocaInst>(BI)) {
1136 CannotRemove = true;
1139 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1140 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1141 // If there is a stackrestore below this one, remove this one.
1142 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1143 return EraseInstFromFunction(CI);
1144 // Otherwise, ignore the intrinsic.
1146 // If we found a non-intrinsic call, we can't remove the stack
1148 CannotRemove = true;
1154 // If the stack restore is in a return, resume, or unwind block and if there
1155 // are no allocas or calls between the restore and the return, nuke the
1157 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1158 return EraseInstFromFunction(CI);
1161 case Intrinsic::assume: {
1162 // Canonicalize assume(a && b) -> assume(a); assume(b);
1163 // Note: New assumption intrinsics created here are registered by
1164 // the InstCombineIRInserter object.
1165 Value *IIOperand = II->getArgOperand(0), *A, *B,
1166 *AssumeIntrinsic = II->getCalledValue();
1167 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1168 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1169 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1170 return EraseInstFromFunction(*II);
1172 // assume(!(a || b)) -> assume(!a); assume(!b);
1173 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1174 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1176 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1178 return EraseInstFromFunction(*II);
1181 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1182 // (if assume is valid at the load)
1183 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1184 Value *LHS = ICmp->getOperand(0);
1185 Value *RHS = ICmp->getOperand(1);
1186 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1187 isa<LoadInst>(LHS) &&
1188 isa<Constant>(RHS) &&
1189 RHS->getType()->isPointerTy() &&
1190 cast<Constant>(RHS)->isNullValue()) {
1191 LoadInst* LI = cast<LoadInst>(LHS);
1192 if (isValidAssumeForContext(II, LI, DT)) {
1193 MDNode *MD = MDNode::get(II->getContext(), None);
1194 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1195 return EraseInstFromFunction(*II);
1198 // TODO: apply nonnull return attributes to calls and invokes
1199 // TODO: apply range metadata for range check patterns?
1201 // If there is a dominating assume with the same condition as this one,
1202 // then this one is redundant, and should be removed.
1203 APInt KnownZero(1, 0), KnownOne(1, 0);
1204 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1205 if (KnownOne.isAllOnesValue())
1206 return EraseInstFromFunction(*II);
1210 case Intrinsic::experimental_gc_relocate: {
1211 // Translate facts known about a pointer before relocating into
1212 // facts about the relocate value, while being careful to
1213 // preserve relocation semantics.
1214 GCRelocateOperands Operands(II);
1215 Value *DerivedPtr = Operands.derivedPtr();
1217 // Remove the relocation if unused, note that this check is required
1218 // to prevent the cases below from looping forever.
1219 if (II->use_empty())
1220 return EraseInstFromFunction(*II);
1222 // Undef is undef, even after relocation.
1223 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1224 // most practical collectors, but there was discussion in the review thread
1225 // about whether it was legal for all possible collectors.
1226 if (isa<UndefValue>(DerivedPtr))
1227 return ReplaceInstUsesWith(*II, DerivedPtr);
1229 // The relocation of null will be null for most any collector.
1230 // TODO: provide a hook for this in GCStrategy. There might be some weird
1231 // collector this property does not hold for.
1232 if (isa<ConstantPointerNull>(DerivedPtr))
1233 return ReplaceInstUsesWith(*II, DerivedPtr);
1235 // isKnownNonNull -> nonnull attribute
1236 if (isKnownNonNull(DerivedPtr))
1237 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1239 // isDereferenceablePointer -> deref attribute
1240 if (DerivedPtr->isDereferenceablePointer(DL)) {
1241 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1242 uint64_t Bytes = A->getDereferenceableBytes();
1243 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1247 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1248 // Canonicalize on the type from the uses to the defs
1250 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1254 return visitCallSite(II);
1257 // InvokeInst simplification
1259 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1260 return visitCallSite(&II);
1263 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1264 /// passed through the varargs area, we can eliminate the use of the cast.
1265 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1266 const DataLayout &DL,
1267 const CastInst *const CI,
1269 if (!CI->isLosslessCast())
1272 // If this is a GC intrinsic, avoid munging types. We need types for
1273 // statepoint reconstruction in SelectionDAG.
1274 // TODO: This is probably something which should be expanded to all
1275 // intrinsics since the entire point of intrinsics is that
1276 // they are understandable by the optimizer.
1277 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1280 // The size of ByVal or InAlloca arguments is derived from the type, so we
1281 // can't change to a type with a different size. If the size were
1282 // passed explicitly we could avoid this check.
1283 if (!CS.isByValOrInAllocaArgument(ix))
1287 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1288 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1289 if (!SrcTy->isSized() || !DstTy->isSized())
1291 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
1296 // Try to fold some different type of calls here.
1297 // Currently we're only working with the checking functions, memcpy_chk,
1298 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1299 // strcat_chk and strncat_chk.
1300 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
1301 if (!CI->getCalledFunction()) return nullptr;
1303 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1304 ReplaceInstUsesWith(*From, With);
1306 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1307 if (Value *With = Simplifier.optimizeCall(CI)) {
1309 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1315 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1316 // Strip off at most one level of pointer casts, looking for an alloca. This
1317 // is good enough in practice and simpler than handling any number of casts.
1318 Value *Underlying = TrampMem->stripPointerCasts();
1319 if (Underlying != TrampMem &&
1320 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1322 if (!isa<AllocaInst>(Underlying))
1325 IntrinsicInst *InitTrampoline = nullptr;
1326 for (User *U : TrampMem->users()) {
1327 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1330 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1332 // More than one init_trampoline writes to this value. Give up.
1334 InitTrampoline = II;
1337 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1338 // Allow any number of calls to adjust.trampoline.
1343 // No call to init.trampoline found.
1344 if (!InitTrampoline)
1347 // Check that the alloca is being used in the expected way.
1348 if (InitTrampoline->getOperand(0) != TrampMem)
1351 return InitTrampoline;
1354 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1356 // Visit all the previous instructions in the basic block, and try to find a
1357 // init.trampoline which has a direct path to the adjust.trampoline.
1358 for (BasicBlock::iterator I = AdjustTramp,
1359 E = AdjustTramp->getParent()->begin(); I != E; ) {
1360 Instruction *Inst = --I;
1361 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1362 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1363 II->getOperand(0) == TrampMem)
1365 if (Inst->mayWriteToMemory())
1371 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1372 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1373 // to a direct call to a function. Otherwise return NULL.
1375 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1376 Callee = Callee->stripPointerCasts();
1377 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1379 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1382 Value *TrampMem = AdjustTramp->getOperand(0);
1384 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1386 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1391 // visitCallSite - Improvements for call and invoke instructions.
1393 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1394 if (isAllocLikeFn(CS.getInstruction(), TLI))
1395 return visitAllocSite(*CS.getInstruction());
1397 bool Changed = false;
1399 // If the callee is a pointer to a function, attempt to move any casts to the
1400 // arguments of the call/invoke.
1401 Value *Callee = CS.getCalledValue();
1402 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1405 if (Function *CalleeF = dyn_cast<Function>(Callee))
1406 // If the call and callee calling conventions don't match, this call must
1407 // be unreachable, as the call is undefined.
1408 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1409 // Only do this for calls to a function with a body. A prototype may
1410 // not actually end up matching the implementation's calling conv for a
1411 // variety of reasons (e.g. it may be written in assembly).
1412 !CalleeF->isDeclaration()) {
1413 Instruction *OldCall = CS.getInstruction();
1414 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1415 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1417 // If OldCall does not return void then replaceAllUsesWith undef.
1418 // This allows ValueHandlers and custom metadata to adjust itself.
1419 if (!OldCall->getType()->isVoidTy())
1420 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1421 if (isa<CallInst>(OldCall))
1422 return EraseInstFromFunction(*OldCall);
1424 // We cannot remove an invoke, because it would change the CFG, just
1425 // change the callee to a null pointer.
1426 cast<InvokeInst>(OldCall)->setCalledFunction(
1427 Constant::getNullValue(CalleeF->getType()));
1431 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1432 // If CS does not return void then replaceAllUsesWith undef.
1433 // This allows ValueHandlers and custom metadata to adjust itself.
1434 if (!CS.getInstruction()->getType()->isVoidTy())
1435 ReplaceInstUsesWith(*CS.getInstruction(),
1436 UndefValue::get(CS.getInstruction()->getType()));
1438 if (isa<InvokeInst>(CS.getInstruction())) {
1439 // Can't remove an invoke because we cannot change the CFG.
1443 // This instruction is not reachable, just remove it. We insert a store to
1444 // undef so that we know that this code is not reachable, despite the fact
1445 // that we can't modify the CFG here.
1446 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1447 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1448 CS.getInstruction());
1450 return EraseInstFromFunction(*CS.getInstruction());
1453 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1454 return transformCallThroughTrampoline(CS, II);
1456 PointerType *PTy = cast<PointerType>(Callee->getType());
1457 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1458 if (FTy->isVarArg()) {
1459 int ix = FTy->getNumParams();
1460 // See if we can optimize any arguments passed through the varargs area of
1462 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1463 E = CS.arg_end(); I != E; ++I, ++ix) {
1464 CastInst *CI = dyn_cast<CastInst>(*I);
1465 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
1466 *I = CI->getOperand(0);
1472 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1473 // Inline asm calls cannot throw - mark them 'nounwind'.
1474 CS.setDoesNotThrow();
1478 // Try to optimize the call if possible, we require DataLayout for most of
1479 // this. None of these calls are seen as possibly dead so go ahead and
1480 // delete the instruction now.
1481 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1482 Instruction *I = tryOptimizeCall(CI);
1483 // If we changed something return the result, etc. Otherwise let
1484 // the fallthrough check.
1485 if (I) return EraseInstFromFunction(*I);
1488 return Changed ? CS.getInstruction() : nullptr;
1491 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1492 // attempt to move the cast to the arguments of the call/invoke.
1494 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1496 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1499 // The prototype of thunks are a lie, don't try to directly call such
1501 if (Callee->hasFnAttribute("thunk"))
1503 Instruction *Caller = CS.getInstruction();
1504 const AttributeSet &CallerPAL = CS.getAttributes();
1506 // Okay, this is a cast from a function to a different type. Unless doing so
1507 // would cause a type conversion of one of our arguments, change this call to
1508 // be a direct call with arguments casted to the appropriate types.
1510 FunctionType *FT = Callee->getFunctionType();
1511 Type *OldRetTy = Caller->getType();
1512 Type *NewRetTy = FT->getReturnType();
1514 // Check to see if we are changing the return type...
1515 if (OldRetTy != NewRetTy) {
1517 if (NewRetTy->isStructTy())
1518 return false; // TODO: Handle multiple return values.
1520 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1521 if (Callee->isDeclaration())
1522 return false; // Cannot transform this return value.
1524 if (!Caller->use_empty() &&
1525 // void -> non-void is handled specially
1526 !NewRetTy->isVoidTy())
1527 return false; // Cannot transform this return value.
1530 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1531 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1533 hasAttributes(AttributeFuncs::
1534 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1535 AttributeSet::ReturnIndex))
1536 return false; // Attribute not compatible with transformed value.
1539 // If the callsite is an invoke instruction, and the return value is used by
1540 // a PHI node in a successor, we cannot change the return type of the call
1541 // because there is no place to put the cast instruction (without breaking
1542 // the critical edge). Bail out in this case.
1543 if (!Caller->use_empty())
1544 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1545 for (User *U : II->users())
1546 if (PHINode *PN = dyn_cast<PHINode>(U))
1547 if (PN->getParent() == II->getNormalDest() ||
1548 PN->getParent() == II->getUnwindDest())
1552 unsigned NumActualArgs = CS.arg_size();
1553 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1555 // Prevent us turning:
1556 // declare void @takes_i32_inalloca(i32* inalloca)
1557 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1560 // call void @takes_i32_inalloca(i32* null)
1562 // Similarly, avoid folding away bitcasts of byval calls.
1563 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
1564 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
1567 CallSite::arg_iterator AI = CS.arg_begin();
1568 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1569 Type *ParamTy = FT->getParamType(i);
1570 Type *ActTy = (*AI)->getType();
1572 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1573 return false; // Cannot transform this parameter value.
1575 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1576 hasAttributes(AttributeFuncs::
1577 typeIncompatible(ParamTy, i + 1), i + 1))
1578 return false; // Attribute not compatible with transformed value.
1580 if (CS.isInAllocaArgument(i))
1581 return false; // Cannot transform to and from inalloca.
1583 // If the parameter is passed as a byval argument, then we have to have a
1584 // sized type and the sized type has to have the same size as the old type.
1585 if (ParamTy != ActTy &&
1586 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1587 Attribute::ByVal)) {
1588 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1589 if (!ParamPTy || !ParamPTy->getElementType()->isSized())
1592 Type *CurElTy = ActTy->getPointerElementType();
1593 if (DL.getTypeAllocSize(CurElTy) !=
1594 DL.getTypeAllocSize(ParamPTy->getElementType()))
1599 if (Callee->isDeclaration()) {
1600 // Do not delete arguments unless we have a function body.
1601 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1604 // If the callee is just a declaration, don't change the varargsness of the
1605 // call. We don't want to introduce a varargs call where one doesn't
1607 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1608 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1611 // If both the callee and the cast type are varargs, we still have to make
1612 // sure the number of fixed parameters are the same or we have the same
1613 // ABI issues as if we introduce a varargs call.
1614 if (FT->isVarArg() &&
1615 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1616 FT->getNumParams() !=
1617 cast<FunctionType>(APTy->getElementType())->getNumParams())
1621 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1622 !CallerPAL.isEmpty())
1623 // In this case we have more arguments than the new function type, but we
1624 // won't be dropping them. Check that these extra arguments have attributes
1625 // that are compatible with being a vararg call argument.
1626 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1627 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1628 if (Index <= FT->getNumParams())
1631 // Check if it has an attribute that's incompatible with varargs.
1632 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1633 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1638 // Okay, we decided that this is a safe thing to do: go ahead and start
1639 // inserting cast instructions as necessary.
1640 std::vector<Value*> Args;
1641 Args.reserve(NumActualArgs);
1642 SmallVector<AttributeSet, 8> attrVec;
1643 attrVec.reserve(NumCommonArgs);
1645 // Get any return attributes.
1646 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1648 // If the return value is not being used, the type may not be compatible
1649 // with the existing attributes. Wipe out any problematic attributes.
1651 removeAttributes(AttributeFuncs::
1652 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1653 AttributeSet::ReturnIndex);
1655 // Add the new return attributes.
1656 if (RAttrs.hasAttributes())
1657 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1658 AttributeSet::ReturnIndex, RAttrs));
1660 AI = CS.arg_begin();
1661 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1662 Type *ParamTy = FT->getParamType(i);
1664 if ((*AI)->getType() == ParamTy) {
1665 Args.push_back(*AI);
1667 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1670 // Add any parameter attributes.
1671 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1672 if (PAttrs.hasAttributes())
1673 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1677 // If the function takes more arguments than the call was taking, add them
1679 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1680 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1682 // If we are removing arguments to the function, emit an obnoxious warning.
1683 if (FT->getNumParams() < NumActualArgs) {
1684 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1685 if (FT->isVarArg()) {
1686 // Add all of the arguments in their promoted form to the arg list.
1687 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1688 Type *PTy = getPromotedType((*AI)->getType());
1689 if (PTy != (*AI)->getType()) {
1690 // Must promote to pass through va_arg area!
1691 Instruction::CastOps opcode =
1692 CastInst::getCastOpcode(*AI, false, PTy, false);
1693 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1695 Args.push_back(*AI);
1698 // Add any parameter attributes.
1699 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1700 if (PAttrs.hasAttributes())
1701 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1707 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1708 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1709 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1711 if (NewRetTy->isVoidTy())
1712 Caller->setName(""); // Void type should not have a name.
1714 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1718 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1719 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1720 II->getUnwindDest(), Args);
1722 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1723 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1725 CallInst *CI = cast<CallInst>(Caller);
1726 NC = Builder->CreateCall(Callee, Args);
1728 if (CI->isTailCall())
1729 cast<CallInst>(NC)->setTailCall();
1730 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1731 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1734 // Insert a cast of the return type as necessary.
1736 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1737 if (!NV->getType()->isVoidTy()) {
1738 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1739 NC->setDebugLoc(Caller->getDebugLoc());
1741 // If this is an invoke instruction, we should insert it after the first
1742 // non-phi, instruction in the normal successor block.
1743 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1744 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1745 InsertNewInstBefore(NC, *I);
1747 // Otherwise, it's a call, just insert cast right after the call.
1748 InsertNewInstBefore(NC, *Caller);
1750 Worklist.AddUsersToWorkList(*Caller);
1752 NV = UndefValue::get(Caller->getType());
1756 if (!Caller->use_empty())
1757 ReplaceInstUsesWith(*Caller, NV);
1758 else if (Caller->hasValueHandle()) {
1759 if (OldRetTy == NV->getType())
1760 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1762 // We cannot call ValueIsRAUWd with a different type, and the
1763 // actual tracked value will disappear.
1764 ValueHandleBase::ValueIsDeleted(Caller);
1767 EraseInstFromFunction(*Caller);
1771 // transformCallThroughTrampoline - Turn a call to a function created by
1772 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1773 // underlying function.
1776 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1777 IntrinsicInst *Tramp) {
1778 Value *Callee = CS.getCalledValue();
1779 PointerType *PTy = cast<PointerType>(Callee->getType());
1780 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1781 const AttributeSet &Attrs = CS.getAttributes();
1783 // If the call already has the 'nest' attribute somewhere then give up -
1784 // otherwise 'nest' would occur twice after splicing in the chain.
1785 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1789 "transformCallThroughTrampoline called with incorrect CallSite.");
1791 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1792 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1793 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1795 const AttributeSet &NestAttrs = NestF->getAttributes();
1796 if (!NestAttrs.isEmpty()) {
1797 unsigned NestIdx = 1;
1798 Type *NestTy = nullptr;
1799 AttributeSet NestAttr;
1801 // Look for a parameter marked with the 'nest' attribute.
1802 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1803 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1804 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1805 // Record the parameter type and any other attributes.
1807 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1812 Instruction *Caller = CS.getInstruction();
1813 std::vector<Value*> NewArgs;
1814 NewArgs.reserve(CS.arg_size() + 1);
1816 SmallVector<AttributeSet, 8> NewAttrs;
1817 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1819 // Insert the nest argument into the call argument list, which may
1820 // mean appending it. Likewise for attributes.
1822 // Add any result attributes.
1823 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1824 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1825 Attrs.getRetAttributes()));
1829 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1831 if (Idx == NestIdx) {
1832 // Add the chain argument and attributes.
1833 Value *NestVal = Tramp->getArgOperand(2);
1834 if (NestVal->getType() != NestTy)
1835 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1836 NewArgs.push_back(NestVal);
1837 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1844 // Add the original argument and attributes.
1845 NewArgs.push_back(*I);
1846 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1847 if (Attr.hasAttributes(Idx)) {
1848 AttrBuilder B(Attr, Idx);
1849 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1850 Idx + (Idx >= NestIdx), B));
1857 // Add any function attributes.
1858 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1859 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1860 Attrs.getFnAttributes()));
1862 // The trampoline may have been bitcast to a bogus type (FTy).
1863 // Handle this by synthesizing a new function type, equal to FTy
1864 // with the chain parameter inserted.
1866 std::vector<Type*> NewTypes;
1867 NewTypes.reserve(FTy->getNumParams()+1);
1869 // Insert the chain's type into the list of parameter types, which may
1870 // mean appending it.
1873 FunctionType::param_iterator I = FTy->param_begin(),
1874 E = FTy->param_end();
1878 // Add the chain's type.
1879 NewTypes.push_back(NestTy);
1884 // Add the original type.
1885 NewTypes.push_back(*I);
1891 // Replace the trampoline call with a direct call. Let the generic
1892 // code sort out any function type mismatches.
1893 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1895 Constant *NewCallee =
1896 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1897 NestF : ConstantExpr::getBitCast(NestF,
1898 PointerType::getUnqual(NewFTy));
1899 const AttributeSet &NewPAL =
1900 AttributeSet::get(FTy->getContext(), NewAttrs);
1902 Instruction *NewCaller;
1903 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1904 NewCaller = InvokeInst::Create(NewCallee,
1905 II->getNormalDest(), II->getUnwindDest(),
1907 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1908 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1910 NewCaller = CallInst::Create(NewCallee, NewArgs);
1911 if (cast<CallInst>(Caller)->isTailCall())
1912 cast<CallInst>(NewCaller)->setTailCall();
1913 cast<CallInst>(NewCaller)->
1914 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1915 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1922 // Replace the trampoline call with a direct call. Since there is no 'nest'
1923 // parameter, there is no need to adjust the argument list. Let the generic
1924 // code sort out any function type mismatches.
1925 Constant *NewCallee =
1926 NestF->getType() == PTy ? NestF :
1927 ConstantExpr::getBitCast(NestF, PTy);
1928 CS.setCalledFunction(NewCallee);
1929 return CS.getInstruction();