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/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"
24 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
26 using namespace PatternMatch;
28 #define DEBUG_TYPE "instcombine"
30 STATISTIC(NumSimplified, "Number of library calls simplified");
32 /// getPromotedType - Return the specified type promoted as it would be to pass
33 /// though a va_arg area.
34 static Type *getPromotedType(Type *Ty) {
35 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
36 if (ITy->getBitWidth() < 32)
37 return Type::getInt32Ty(Ty->getContext());
42 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a
43 /// single scalar element, like {{{type}}} or [1 x type], return type.
44 static Type *reduceToSingleValueType(Type *T) {
45 while (!T->isSingleValueType()) {
46 if (StructType *STy = dyn_cast<StructType>(T)) {
47 if (STy->getNumElements() == 1)
48 T = STy->getElementType(0);
51 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
52 if (ATy->getNumElements() == 1)
53 T = ATy->getElementType();
63 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
64 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, AC, MI, DT);
65 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, AC, MI, DT);
66 unsigned MinAlign = std::min(DstAlign, SrcAlign);
67 unsigned CopyAlign = MI->getAlignment();
69 if (CopyAlign < MinAlign) {
70 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
75 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
77 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
78 if (!MemOpLength) return nullptr;
80 // Source and destination pointer types are always "i8*" for intrinsic. See
81 // if the size is something we can handle with a single primitive load/store.
82 // A single load+store correctly handles overlapping memory in the memmove
84 uint64_t Size = MemOpLength->getLimitedValue();
85 assert(Size && "0-sized memory transferring should be removed already.");
87 if (Size > 8 || (Size&(Size-1)))
88 return nullptr; // If not 1/2/4/8 bytes, exit.
90 // Use an integer load+store unless we can find something better.
92 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
94 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
96 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
97 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
98 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
100 // Memcpy forces the use of i8* for the source and destination. That means
101 // that if you're using memcpy to move one double around, you'll get a cast
102 // from double* to i8*. We'd much rather use a double load+store rather than
103 // an i64 load+store, here because this improves the odds that the source or
104 // dest address will be promotable. See if we can find a better type than the
106 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
107 MDNode *CopyMD = nullptr;
108 if (StrippedDest != MI->getArgOperand(0)) {
109 Type *SrcETy = cast<PointerType>(StrippedDest->getType())
111 if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
112 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
113 // down through these levels if so.
114 SrcETy = reduceToSingleValueType(SrcETy);
116 if (SrcETy->isSingleValueType()) {
117 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
118 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
120 // If the memcpy has metadata describing the members, see if we can
121 // get the TBAA tag describing our copy.
122 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
123 if (M->getNumOperands() == 3 && M->getOperand(0) &&
124 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
125 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() &&
127 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
128 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
130 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
131 CopyMD = cast<MDNode>(M->getOperand(2));
137 // If the memcpy/memmove provides better alignment info than we can
139 SrcAlign = std::max(SrcAlign, CopyAlign);
140 DstAlign = std::max(DstAlign, CopyAlign);
142 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
143 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
144 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
145 L->setAlignment(SrcAlign);
147 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
148 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
149 S->setAlignment(DstAlign);
151 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
153 // Set the size of the copy to 0, it will be deleted on the next iteration.
154 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
158 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
159 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, AC, MI, DT);
160 if (MI->getAlignment() < Alignment) {
161 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
166 // Extract the length and alignment and fill if they are constant.
167 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
168 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
169 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
171 uint64_t Len = LenC->getLimitedValue();
172 Alignment = MI->getAlignment();
173 assert(Len && "0-sized memory setting should be removed already.");
175 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
176 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
177 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
179 Value *Dest = MI->getDest();
180 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
181 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
182 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
184 // Alignment 0 is identity for alignment 1 for memset, but not store.
185 if (Alignment == 0) Alignment = 1;
187 // Extract the fill value and store.
188 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
189 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
191 S->setAlignment(Alignment);
193 // Set the size of the copy to 0, it will be deleted on the next iteration.
194 MI->setLength(Constant::getNullValue(LenC->getType()));
201 /// visitCallInst - CallInst simplification. This mostly only handles folding
202 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
203 /// the heavy lifting.
205 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
206 if (isFreeCall(&CI, TLI))
207 return visitFree(CI);
209 // If the caller function is nounwind, mark the call as nounwind, even if the
211 if (CI.getParent()->getParent()->doesNotThrow() &&
212 !CI.doesNotThrow()) {
213 CI.setDoesNotThrow();
217 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
218 if (!II) return visitCallSite(&CI);
220 // Intrinsics cannot occur in an invoke, so handle them here instead of in
222 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
223 bool Changed = false;
225 // memmove/cpy/set of zero bytes is a noop.
226 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
227 if (NumBytes->isNullValue())
228 return EraseInstFromFunction(CI);
230 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
231 if (CI->getZExtValue() == 1) {
232 // Replace the instruction with just byte operations. We would
233 // transform other cases to loads/stores, but we don't know if
234 // alignment is sufficient.
238 // No other transformations apply to volatile transfers.
239 if (MI->isVolatile())
242 // If we have a memmove and the source operation is a constant global,
243 // then the source and dest pointers can't alias, so we can change this
244 // into a call to memcpy.
245 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
246 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
247 if (GVSrc->isConstant()) {
248 Module *M = CI.getParent()->getParent()->getParent();
249 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
250 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
251 CI.getArgOperand(1)->getType(),
252 CI.getArgOperand(2)->getType() };
253 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
258 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
259 // memmove(x,x,size) -> noop.
260 if (MTI->getSource() == MTI->getDest())
261 return EraseInstFromFunction(CI);
264 // If we can determine a pointer alignment that is bigger than currently
265 // set, update the alignment.
266 if (isa<MemTransferInst>(MI)) {
267 if (Instruction *I = SimplifyMemTransfer(MI))
269 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
270 if (Instruction *I = SimplifyMemSet(MSI))
274 if (Changed) return II;
277 switch (II->getIntrinsicID()) {
279 case Intrinsic::objectsize: {
281 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
282 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
285 case Intrinsic::bswap: {
286 Value *IIOperand = II->getArgOperand(0);
289 // bswap(bswap(x)) -> x
290 if (match(IIOperand, m_BSwap(m_Value(X))))
291 return ReplaceInstUsesWith(CI, X);
293 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
294 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
295 unsigned C = X->getType()->getPrimitiveSizeInBits() -
296 IIOperand->getType()->getPrimitiveSizeInBits();
297 Value *CV = ConstantInt::get(X->getType(), C);
298 Value *V = Builder->CreateLShr(X, CV);
299 return new TruncInst(V, IIOperand->getType());
304 case Intrinsic::powi:
305 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
308 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
311 return ReplaceInstUsesWith(CI, II->getArgOperand(0));
312 // powi(x, -1) -> 1/x
313 if (Power->isAllOnesValue())
314 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
315 II->getArgOperand(0));
318 case Intrinsic::cttz: {
319 // If all bits below the first known one are known zero,
320 // this value is constant.
321 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
322 // FIXME: Try to simplify vectors of integers.
324 uint32_t BitWidth = IT->getBitWidth();
325 APInt KnownZero(BitWidth, 0);
326 APInt KnownOne(BitWidth, 0);
327 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
328 unsigned TrailingZeros = KnownOne.countTrailingZeros();
329 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
330 if ((Mask & KnownZero) == Mask)
331 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
332 APInt(BitWidth, TrailingZeros)));
336 case Intrinsic::ctlz: {
337 // If all bits above the first known one are known zero,
338 // this value is constant.
339 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
340 // FIXME: Try to simplify vectors of integers.
342 uint32_t BitWidth = IT->getBitWidth();
343 APInt KnownZero(BitWidth, 0);
344 APInt KnownOne(BitWidth, 0);
345 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II);
346 unsigned LeadingZeros = KnownOne.countLeadingZeros();
347 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
348 if ((Mask & KnownZero) == Mask)
349 return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
350 APInt(BitWidth, LeadingZeros)));
354 case Intrinsic::uadd_with_overflow: {
355 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
356 OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, II);
357 if (OR == OverflowResult::NeverOverflows)
358 return CreateOverflowTuple(II, Builder->CreateNUWAdd(LHS, RHS), false);
359 if (OR == OverflowResult::AlwaysOverflows)
360 return CreateOverflowTuple(II, Builder->CreateAdd(LHS, RHS), true);
362 // FALL THROUGH uadd into sadd
363 case Intrinsic::sadd_with_overflow:
364 // Canonicalize constants into the RHS.
365 if (isa<Constant>(II->getArgOperand(0)) &&
366 !isa<Constant>(II->getArgOperand(1))) {
367 Value *LHS = II->getArgOperand(0);
368 II->setArgOperand(0, II->getArgOperand(1));
369 II->setArgOperand(1, LHS);
373 // X + undef -> undef
374 if (isa<UndefValue>(II->getArgOperand(1)))
375 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
377 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
378 // X + 0 -> {X, false}
380 return CreateOverflowTuple(II, II->getArgOperand(0), false,
385 // We can strength reduce reduce this signed add into a regular add if we
386 // can prove that it will never overflow.
387 if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow) {
388 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
389 if (WillNotOverflowSignedAdd(LHS, RHS, II)) {
390 return CreateOverflowTuple(II, Builder->CreateNSWAdd(LHS, RHS), false);
395 case Intrinsic::usub_with_overflow:
396 case Intrinsic::ssub_with_overflow: {
397 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
398 // undef - X -> undef
399 // X - undef -> undef
400 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
401 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
403 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
404 // X - 0 -> {X, false}
405 if (ConstRHS->isZero()) {
406 return CreateOverflowTuple(II, LHS, false, /*ReUseName*/false);
409 if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
410 if (WillNotOverflowSignedSub(LHS, RHS, II)) {
411 return CreateOverflowTuple(II, Builder->CreateNSWSub(LHS, RHS), false);
414 if (WillNotOverflowUnsignedSub(LHS, RHS, II)) {
415 return CreateOverflowTuple(II, Builder->CreateNUWSub(LHS, RHS), false);
420 case Intrinsic::umul_with_overflow: {
421 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
422 OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, II);
423 if (OR == OverflowResult::NeverOverflows)
424 return CreateOverflowTuple(II, Builder->CreateNUWMul(LHS, RHS), false);
425 if (OR == OverflowResult::AlwaysOverflows)
426 return CreateOverflowTuple(II, Builder->CreateMul(LHS, RHS), true);
428 case Intrinsic::smul_with_overflow:
429 // Canonicalize constants into the RHS.
430 if (isa<Constant>(II->getArgOperand(0)) &&
431 !isa<Constant>(II->getArgOperand(1))) {
432 Value *LHS = II->getArgOperand(0);
433 II->setArgOperand(0, II->getArgOperand(1));
434 II->setArgOperand(1, LHS);
438 // X * undef -> undef
439 if (isa<UndefValue>(II->getArgOperand(1)))
440 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
442 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
445 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
447 // X * 1 -> {X, false}
448 if (RHSI->equalsInt(1)) {
449 return CreateOverflowTuple(II, II->getArgOperand(0), false,
453 if (II->getIntrinsicID() == Intrinsic::smul_with_overflow) {
454 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
455 if (WillNotOverflowSignedMul(LHS, RHS, II)) {
456 return CreateOverflowTuple(II, Builder->CreateNSWMul(LHS, RHS), false);
460 case Intrinsic::minnum:
461 case Intrinsic::maxnum: {
462 Value *Arg0 = II->getArgOperand(0);
463 Value *Arg1 = II->getArgOperand(1);
467 return ReplaceInstUsesWith(CI, Arg0);
469 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0);
470 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1);
472 // Canonicalize constants into the RHS.
474 II->setArgOperand(0, Arg1);
475 II->setArgOperand(1, Arg0);
480 if (C1 && C1->isNaN())
481 return ReplaceInstUsesWith(CI, Arg0);
483 // This is the value because if undef were NaN, we would return the other
484 // value and cannot return a NaN unless both operands are.
486 // fmin(undef, x) -> x
487 if (isa<UndefValue>(Arg0))
488 return ReplaceInstUsesWith(CI, Arg1);
490 // fmin(x, undef) -> x
491 if (isa<UndefValue>(Arg1))
492 return ReplaceInstUsesWith(CI, Arg0);
496 if (II->getIntrinsicID() == Intrinsic::minnum) {
497 // fmin(x, fmin(x, y)) -> fmin(x, y)
498 // fmin(y, fmin(x, y)) -> fmin(x, y)
499 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) {
500 if (Arg0 == X || Arg0 == Y)
501 return ReplaceInstUsesWith(CI, Arg1);
504 // fmin(fmin(x, y), x) -> fmin(x, y)
505 // fmin(fmin(x, y), y) -> fmin(x, y)
506 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) {
507 if (Arg1 == X || Arg1 == Y)
508 return ReplaceInstUsesWith(CI, Arg0);
511 // TODO: fmin(nnan x, inf) -> x
512 // TODO: fmin(nnan ninf x, flt_max) -> x
513 if (C1 && C1->isInfinity()) {
514 // fmin(x, -inf) -> -inf
515 if (C1->isNegative())
516 return ReplaceInstUsesWith(CI, Arg1);
519 assert(II->getIntrinsicID() == Intrinsic::maxnum);
520 // fmax(x, fmax(x, y)) -> fmax(x, y)
521 // fmax(y, fmax(x, y)) -> fmax(x, y)
522 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) {
523 if (Arg0 == X || Arg0 == Y)
524 return ReplaceInstUsesWith(CI, Arg1);
527 // fmax(fmax(x, y), x) -> fmax(x, y)
528 // fmax(fmax(x, y), y) -> fmax(x, y)
529 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) {
530 if (Arg1 == X || Arg1 == Y)
531 return ReplaceInstUsesWith(CI, Arg0);
534 // TODO: fmax(nnan x, -inf) -> x
535 // TODO: fmax(nnan ninf x, -flt_max) -> x
536 if (C1 && C1->isInfinity()) {
537 // fmax(x, inf) -> inf
538 if (!C1->isNegative())
539 return ReplaceInstUsesWith(CI, Arg1);
544 case Intrinsic::ppc_altivec_lvx:
545 case Intrinsic::ppc_altivec_lvxl:
546 // Turn PPC lvx -> load if the pointer is known aligned.
547 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, AC, II, DT) >=
549 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
550 PointerType::getUnqual(II->getType()));
551 return new LoadInst(Ptr);
554 case Intrinsic::ppc_vsx_lxvw4x:
555 case Intrinsic::ppc_vsx_lxvd2x: {
556 // Turn PPC VSX loads into normal loads.
557 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
558 PointerType::getUnqual(II->getType()));
559 return new LoadInst(Ptr, Twine(""), false, 1);
561 case Intrinsic::ppc_altivec_stvx:
562 case Intrinsic::ppc_altivec_stvxl:
563 // Turn stvx -> store if the pointer is known aligned.
564 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, AC, II, DT) >=
567 PointerType::getUnqual(II->getArgOperand(0)->getType());
568 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
569 return new StoreInst(II->getArgOperand(0), Ptr);
572 case Intrinsic::ppc_vsx_stxvw4x:
573 case Intrinsic::ppc_vsx_stxvd2x: {
574 // Turn PPC VSX stores into normal stores.
575 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
576 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
577 return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
579 case Intrinsic::ppc_qpx_qvlfs:
580 // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
581 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, AC, II, DT) >=
583 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
584 PointerType::getUnqual(II->getType()));
585 return new LoadInst(Ptr);
588 case Intrinsic::ppc_qpx_qvlfd:
589 // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
590 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, AC, II, DT) >=
592 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
593 PointerType::getUnqual(II->getType()));
594 return new LoadInst(Ptr);
597 case Intrinsic::ppc_qpx_qvstfs:
598 // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
599 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, AC, II, DT) >=
602 PointerType::getUnqual(II->getArgOperand(0)->getType());
603 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
604 return new StoreInst(II->getArgOperand(0), Ptr);
607 case Intrinsic::ppc_qpx_qvstfd:
608 // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
609 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, AC, II, DT) >=
612 PointerType::getUnqual(II->getArgOperand(0)->getType());
613 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
614 return new StoreInst(II->getArgOperand(0), Ptr);
617 case Intrinsic::x86_sse_storeu_ps:
618 case Intrinsic::x86_sse2_storeu_pd:
619 case Intrinsic::x86_sse2_storeu_dq:
620 // Turn X86 storeu -> store if the pointer is known aligned.
621 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, AC, II, DT) >=
624 PointerType::getUnqual(II->getArgOperand(1)->getType());
625 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
626 return new StoreInst(II->getArgOperand(1), Ptr);
630 case Intrinsic::x86_sse_cvtss2si:
631 case Intrinsic::x86_sse_cvtss2si64:
632 case Intrinsic::x86_sse_cvttss2si:
633 case Intrinsic::x86_sse_cvttss2si64:
634 case Intrinsic::x86_sse2_cvtsd2si:
635 case Intrinsic::x86_sse2_cvtsd2si64:
636 case Intrinsic::x86_sse2_cvttsd2si:
637 case Intrinsic::x86_sse2_cvttsd2si64: {
638 // These intrinsics only demand the 0th element of their input vectors. If
639 // we can simplify the input based on that, do so now.
641 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
642 APInt DemandedElts(VWidth, 1);
643 APInt UndefElts(VWidth, 0);
644 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
645 DemandedElts, UndefElts)) {
646 II->setArgOperand(0, V);
652 // Constant fold <A x Bi> << Ci.
653 // FIXME: We don't handle _dq because it's a shift of an i128, but is
654 // represented in the IR as <2 x i64>. A per element shift is wrong.
655 case Intrinsic::x86_sse2_psll_d:
656 case Intrinsic::x86_sse2_psll_q:
657 case Intrinsic::x86_sse2_psll_w:
658 case Intrinsic::x86_sse2_pslli_d:
659 case Intrinsic::x86_sse2_pslli_q:
660 case Intrinsic::x86_sse2_pslli_w:
661 case Intrinsic::x86_avx2_psll_d:
662 case Intrinsic::x86_avx2_psll_q:
663 case Intrinsic::x86_avx2_psll_w:
664 case Intrinsic::x86_avx2_pslli_d:
665 case Intrinsic::x86_avx2_pslli_q:
666 case Intrinsic::x86_avx2_pslli_w:
667 case Intrinsic::x86_sse2_psrl_d:
668 case Intrinsic::x86_sse2_psrl_q:
669 case Intrinsic::x86_sse2_psrl_w:
670 case Intrinsic::x86_sse2_psrli_d:
671 case Intrinsic::x86_sse2_psrli_q:
672 case Intrinsic::x86_sse2_psrli_w:
673 case Intrinsic::x86_avx2_psrl_d:
674 case Intrinsic::x86_avx2_psrl_q:
675 case Intrinsic::x86_avx2_psrl_w:
676 case Intrinsic::x86_avx2_psrli_d:
677 case Intrinsic::x86_avx2_psrli_q:
678 case Intrinsic::x86_avx2_psrli_w: {
679 // Simplify if count is constant. To 0 if >= BitWidth,
680 // otherwise to shl/lshr.
681 auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
682 auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
687 Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
691 auto Vec = II->getArgOperand(0);
692 auto VT = cast<VectorType>(Vec->getType());
693 if (Count->getZExtValue() >
694 VT->getElementType()->getPrimitiveSizeInBits() - 1)
695 return ReplaceInstUsesWith(
696 CI, ConstantAggregateZero::get(Vec->getType()));
698 bool isPackedShiftLeft = true;
699 switch (II->getIntrinsicID()) {
701 case Intrinsic::x86_sse2_psrl_d:
702 case Intrinsic::x86_sse2_psrl_q:
703 case Intrinsic::x86_sse2_psrl_w:
704 case Intrinsic::x86_sse2_psrli_d:
705 case Intrinsic::x86_sse2_psrli_q:
706 case Intrinsic::x86_sse2_psrli_w:
707 case Intrinsic::x86_avx2_psrl_d:
708 case Intrinsic::x86_avx2_psrl_q:
709 case Intrinsic::x86_avx2_psrl_w:
710 case Intrinsic::x86_avx2_psrli_d:
711 case Intrinsic::x86_avx2_psrli_q:
712 case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
715 unsigned VWidth = VT->getNumElements();
716 // Get a constant vector of the same type as the first operand.
717 auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
718 if (isPackedShiftLeft)
719 return BinaryOperator::CreateShl(Vec,
720 Builder->CreateVectorSplat(VWidth, VTCI));
722 return BinaryOperator::CreateLShr(Vec,
723 Builder->CreateVectorSplat(VWidth, VTCI));
726 case Intrinsic::x86_sse41_pmovsxbw:
727 case Intrinsic::x86_sse41_pmovsxwd:
728 case Intrinsic::x86_sse41_pmovsxdq:
729 case Intrinsic::x86_sse41_pmovzxbw:
730 case Intrinsic::x86_sse41_pmovzxwd:
731 case Intrinsic::x86_sse41_pmovzxdq: {
732 // pmov{s|z}x ignores the upper half of their input vectors.
734 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
735 unsigned LowHalfElts = VWidth / 2;
736 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
737 APInt UndefElts(VWidth, 0);
738 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
741 II->setArgOperand(0, TmpV);
747 case Intrinsic::x86_sse4a_insertqi: {
748 // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
750 // TODO: eventually we should lower this intrinsic to IR
751 if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
752 if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
753 unsigned Index = CIStart->getZExtValue();
754 // From AMD documentation: "a value of zero in the field length is
755 // defined as length of 64".
756 unsigned Length = CIWidth->equalsInt(0) ? 64 : CIWidth->getZExtValue();
758 // From AMD documentation: "If the sum of the bit index + length field
759 // is greater than 64, the results are undefined".
761 // Note that both field index and field length are 8-bit quantities.
762 // Since variables 'Index' and 'Length' are unsigned values
763 // obtained from zero-extending field index and field length
764 // respectively, their sum should never wrap around.
765 if ((Index + Length) > 64)
766 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
768 if (Length == 64 && Index == 0) {
769 Value *Vec = II->getArgOperand(1);
770 Value *Undef = UndefValue::get(Vec->getType());
771 const uint32_t Mask[] = { 0, 2 };
772 return ReplaceInstUsesWith(
774 Builder->CreateShuffleVector(
775 Vec, Undef, ConstantDataVector::get(
776 II->getContext(), makeArrayRef(Mask))));
778 } else if (auto Source =
779 dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
780 if (Source->hasOneUse() &&
781 Source->getArgOperand(1) == II->getArgOperand(1)) {
782 // If the source of the insert has only one use and it's another
783 // insert (and they're both inserting from the same vector), try to
784 // bundle both together.
786 dyn_cast<ConstantInt>(Source->getArgOperand(2));
788 dyn_cast<ConstantInt>(Source->getArgOperand(3));
789 if (CISourceStart && CISourceWidth) {
790 unsigned Start = CIStart->getZExtValue();
791 unsigned Width = CIWidth->getZExtValue();
792 unsigned End = Start + Width;
793 unsigned SourceStart = CISourceStart->getZExtValue();
794 unsigned SourceWidth = CISourceWidth->getZExtValue();
795 unsigned SourceEnd = SourceStart + SourceWidth;
796 unsigned NewStart, NewWidth;
797 bool ShouldReplace = false;
798 if (Start <= SourceStart && SourceStart <= End) {
800 NewWidth = std::max(End, SourceEnd) - NewStart;
801 ShouldReplace = true;
802 } else if (SourceStart <= Start && Start <= SourceEnd) {
803 NewStart = SourceStart;
804 NewWidth = std::max(SourceEnd, End) - NewStart;
805 ShouldReplace = true;
809 Constant *ConstantWidth = ConstantInt::get(
810 II->getArgOperand(2)->getType(), NewWidth, false);
811 Constant *ConstantStart = ConstantInt::get(
812 II->getArgOperand(3)->getType(), NewStart, false);
813 Value *Args[4] = { Source->getArgOperand(0),
814 II->getArgOperand(1), ConstantWidth,
816 Module *M = CI.getParent()->getParent()->getParent();
818 Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
819 return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
829 case Intrinsic::x86_sse41_pblendvb:
830 case Intrinsic::x86_sse41_blendvps:
831 case Intrinsic::x86_sse41_blendvpd:
832 case Intrinsic::x86_avx_blendv_ps_256:
833 case Intrinsic::x86_avx_blendv_pd_256:
834 case Intrinsic::x86_avx2_pblendvb: {
835 // Convert blendv* to vector selects if the mask is constant.
836 // This optimization is convoluted because the intrinsic is defined as
837 // getting a vector of floats or doubles for the ps and pd versions.
838 // FIXME: That should be changed.
839 Value *Mask = II->getArgOperand(2);
840 if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
841 auto Tyi1 = Builder->getInt1Ty();
842 auto SelectorType = cast<VectorType>(Mask->getType());
843 auto EltTy = SelectorType->getElementType();
844 unsigned Size = SelectorType->getNumElements();
848 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
849 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
850 "Wrong arguments for variable blend intrinsic");
851 SmallVector<Constant *, 32> Selectors;
852 for (unsigned I = 0; I < Size; ++I) {
853 // The intrinsics only read the top bit
856 Selector = C->getElementAsInteger(I);
858 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
859 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
861 auto NewSelector = ConstantVector::get(Selectors);
862 return SelectInst::Create(NewSelector, II->getArgOperand(1),
863 II->getArgOperand(0), "blendv");
869 case Intrinsic::x86_avx_vpermilvar_ps:
870 case Intrinsic::x86_avx_vpermilvar_ps_256:
871 case Intrinsic::x86_avx_vpermilvar_pd:
872 case Intrinsic::x86_avx_vpermilvar_pd_256: {
873 // Convert vpermil* to shufflevector if the mask is constant.
874 Value *V = II->getArgOperand(1);
875 unsigned Size = cast<VectorType>(V->getType())->getNumElements();
876 assert(Size == 8 || Size == 4 || Size == 2);
878 if (auto C = dyn_cast<ConstantDataVector>(V)) {
879 // The intrinsics only read one or two bits, clear the rest.
880 for (unsigned I = 0; I < Size; ++I) {
881 uint32_t Index = C->getElementAsInteger(I) & 0x3;
882 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
883 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
887 } else if (isa<ConstantAggregateZero>(V)) {
888 for (unsigned I = 0; I < Size; ++I)
893 // The _256 variants are a bit trickier since the mask bits always index
894 // into the corresponding 128 half. In order to convert to a generic
895 // shuffle, we have to make that explicit.
896 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
897 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
898 for (unsigned I = Size / 2; I < Size; ++I)
899 Indexes[I] += Size / 2;
902 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
903 auto V1 = II->getArgOperand(0);
904 auto V2 = UndefValue::get(V1->getType());
905 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
906 return ReplaceInstUsesWith(CI, Shuffle);
909 case Intrinsic::ppc_altivec_vperm:
910 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
911 // Note that ppc_altivec_vperm has a big-endian bias, so when creating
912 // a vectorshuffle for little endian, we must undo the transformation
913 // performed on vec_perm in altivec.h. That is, we must complement
914 // the permutation mask with respect to 31 and reverse the order of
916 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
917 assert(Mask->getType()->getVectorNumElements() == 16 &&
918 "Bad type for intrinsic!");
920 // Check that all of the elements are integer constants or undefs.
921 bool AllEltsOk = true;
922 for (unsigned i = 0; i != 16; ++i) {
923 Constant *Elt = Mask->getAggregateElement(i);
924 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
931 // Cast the input vectors to byte vectors.
932 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
934 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
936 Value *Result = UndefValue::get(Op0->getType());
938 // Only extract each element once.
939 Value *ExtractedElts[32];
940 memset(ExtractedElts, 0, sizeof(ExtractedElts));
942 for (unsigned i = 0; i != 16; ++i) {
943 if (isa<UndefValue>(Mask->getAggregateElement(i)))
946 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
947 Idx &= 31; // Match the hardware behavior.
948 if (DL && DL->isLittleEndian())
951 if (!ExtractedElts[Idx]) {
952 Value *Op0ToUse = (DL && DL->isLittleEndian()) ? Op1 : Op0;
953 Value *Op1ToUse = (DL && DL->isLittleEndian()) ? Op0 : Op1;
955 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
956 Builder->getInt32(Idx&15));
959 // Insert this value into the result vector.
960 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
961 Builder->getInt32(i));
963 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
968 case Intrinsic::arm_neon_vld1:
969 case Intrinsic::arm_neon_vld2:
970 case Intrinsic::arm_neon_vld3:
971 case Intrinsic::arm_neon_vld4:
972 case Intrinsic::arm_neon_vld2lane:
973 case Intrinsic::arm_neon_vld3lane:
974 case Intrinsic::arm_neon_vld4lane:
975 case Intrinsic::arm_neon_vst1:
976 case Intrinsic::arm_neon_vst2:
977 case Intrinsic::arm_neon_vst3:
978 case Intrinsic::arm_neon_vst4:
979 case Intrinsic::arm_neon_vst2lane:
980 case Intrinsic::arm_neon_vst3lane:
981 case Intrinsic::arm_neon_vst4lane: {
982 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, AC, II, DT);
983 unsigned AlignArg = II->getNumArgOperands() - 1;
984 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
985 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
986 II->setArgOperand(AlignArg,
987 ConstantInt::get(Type::getInt32Ty(II->getContext()),
994 case Intrinsic::arm_neon_vmulls:
995 case Intrinsic::arm_neon_vmullu:
996 case Intrinsic::aarch64_neon_smull:
997 case Intrinsic::aarch64_neon_umull: {
998 Value *Arg0 = II->getArgOperand(0);
999 Value *Arg1 = II->getArgOperand(1);
1001 // Handle mul by zero first:
1002 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
1003 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
1006 // Check for constant LHS & RHS - in this case we just simplify.
1007 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
1008 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
1009 VectorType *NewVT = cast<VectorType>(II->getType());
1010 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
1011 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
1012 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
1013 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
1015 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
1018 // Couldn't simplify - canonicalize constant to the RHS.
1019 std::swap(Arg0, Arg1);
1022 // Handle mul by one:
1023 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
1024 if (ConstantInt *Splat =
1025 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
1027 return CastInst::CreateIntegerCast(Arg0, II->getType(),
1028 /*isSigned=*/!Zext);
1033 case Intrinsic::AMDGPU_rcp: {
1034 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
1035 const APFloat &ArgVal = C->getValueAPF();
1036 APFloat Val(ArgVal.getSemantics(), 1.0);
1037 APFloat::opStatus Status = Val.divide(ArgVal,
1038 APFloat::rmNearestTiesToEven);
1039 // Only do this if it was exact and therefore not dependent on the
1041 if (Status == APFloat::opOK)
1042 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
1047 case Intrinsic::stackrestore: {
1048 // If the save is right next to the restore, remove the restore. This can
1049 // happen when variable allocas are DCE'd.
1050 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
1051 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
1052 BasicBlock::iterator BI = SS;
1054 return EraseInstFromFunction(CI);
1058 // Scan down this block to see if there is another stack restore in the
1059 // same block without an intervening call/alloca.
1060 BasicBlock::iterator BI = II;
1061 TerminatorInst *TI = II->getParent()->getTerminator();
1062 bool CannotRemove = false;
1063 for (++BI; &*BI != TI; ++BI) {
1064 if (isa<AllocaInst>(BI)) {
1065 CannotRemove = true;
1068 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
1069 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
1070 // If there is a stackrestore below this one, remove this one.
1071 if (II->getIntrinsicID() == Intrinsic::stackrestore)
1072 return EraseInstFromFunction(CI);
1073 // Otherwise, ignore the intrinsic.
1075 // If we found a non-intrinsic call, we can't remove the stack
1077 CannotRemove = true;
1083 // If the stack restore is in a return, resume, or unwind block and if there
1084 // are no allocas or calls between the restore and the return, nuke the
1086 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
1087 return EraseInstFromFunction(CI);
1090 case Intrinsic::assume: {
1091 // Canonicalize assume(a && b) -> assume(a); assume(b);
1092 // Note: New assumption intrinsics created here are registered by
1093 // the InstCombineIRInserter object.
1094 Value *IIOperand = II->getArgOperand(0), *A, *B,
1095 *AssumeIntrinsic = II->getCalledValue();
1096 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
1097 Builder->CreateCall(AssumeIntrinsic, A, II->getName());
1098 Builder->CreateCall(AssumeIntrinsic, B, II->getName());
1099 return EraseInstFromFunction(*II);
1101 // assume(!(a || b)) -> assume(!a); assume(!b);
1102 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
1103 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A),
1105 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B),
1107 return EraseInstFromFunction(*II);
1110 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
1111 // (if assume is valid at the load)
1112 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) {
1113 Value *LHS = ICmp->getOperand(0);
1114 Value *RHS = ICmp->getOperand(1);
1115 if (ICmpInst::ICMP_NE == ICmp->getPredicate() &&
1116 isa<LoadInst>(LHS) &&
1117 isa<Constant>(RHS) &&
1118 RHS->getType()->isPointerTy() &&
1119 cast<Constant>(RHS)->isNullValue()) {
1120 LoadInst* LI = cast<LoadInst>(LHS);
1121 if (isValidAssumeForContext(II, LI, DL, DT)) {
1122 MDNode *MD = MDNode::get(II->getContext(), None);
1123 LI->setMetadata(LLVMContext::MD_nonnull, MD);
1124 return EraseInstFromFunction(*II);
1127 // TODO: apply nonnull return attributes to calls and invokes
1128 // TODO: apply range metadata for range check patterns?
1130 // If there is a dominating assume with the same condition as this one,
1131 // then this one is redundant, and should be removed.
1132 APInt KnownZero(1, 0), KnownOne(1, 0);
1133 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II);
1134 if (KnownOne.isAllOnesValue())
1135 return EraseInstFromFunction(*II);
1139 case Intrinsic::experimental_gc_relocate: {
1140 // Translate facts known about a pointer before relocating into
1141 // facts about the relocate value, while being careful to
1142 // preserve relocation semantics.
1143 GCRelocateOperands Operands(II);
1144 Value *DerivedPtr = Operands.derivedPtr();
1146 // Remove the relocation if unused, note that this check is required
1147 // to prevent the cases below from looping forever.
1148 if (II->use_empty())
1149 return EraseInstFromFunction(*II);
1151 // Undef is undef, even after relocation.
1152 // TODO: provide a hook for this in GCStrategy. This is clearly legal for
1153 // most practical collectors, but there was discussion in the review thread
1154 // about whether it was legal for all possible collectors.
1155 if (isa<UndefValue>(DerivedPtr))
1156 return ReplaceInstUsesWith(*II, DerivedPtr);
1158 // The relocation of null will be null for most any collector.
1159 // TODO: provide a hook for this in GCStrategy. There might be some weird
1160 // collector this property does not hold for.
1161 if (isa<ConstantPointerNull>(DerivedPtr))
1162 return ReplaceInstUsesWith(*II, DerivedPtr);
1164 // isKnownNonNull -> nonnull attribute
1165 if (isKnownNonNull(DerivedPtr))
1166 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
1168 // isDereferenceablePointer -> deref attribute
1169 if (DerivedPtr->isDereferenceablePointer(DL)) {
1170 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) {
1171 uint64_t Bytes = A->getDereferenceableBytes();
1172 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes);
1176 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
1177 // Canonicalize on the type from the uses to the defs
1179 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
1183 return visitCallSite(II);
1186 // InvokeInst simplification
1188 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1189 return visitCallSite(&II);
1192 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
1193 /// passed through the varargs area, we can eliminate the use of the cast.
1194 static bool isSafeToEliminateVarargsCast(const CallSite CS,
1195 const CastInst * const CI,
1196 const DataLayout * const DL,
1198 if (!CI->isLosslessCast())
1201 // If this is a GC intrinsic, avoid munging types. We need types for
1202 // statepoint reconstruction in SelectionDAG.
1203 // TODO: This is probably something which should be expanded to all
1204 // intrinsics since the entire point of intrinsics is that
1205 // they are understandable by the optimizer.
1206 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
1209 // The size of ByVal or InAlloca arguments is derived from the type, so we
1210 // can't change to a type with a different size. If the size were
1211 // passed explicitly we could avoid this check.
1212 if (!CS.isByValOrInAllocaArgument(ix))
1216 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1217 Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1218 if (!SrcTy->isSized() || !DstTy->isSized())
1220 if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
1225 // Try to fold some different type of calls here.
1226 // Currently we're only working with the checking functions, memcpy_chk,
1227 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1228 // strcat_chk and strncat_chk.
1229 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
1230 if (!CI->getCalledFunction()) return nullptr;
1232 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
1233 ReplaceInstUsesWith(*From, With);
1235 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW);
1236 if (Value *With = Simplifier.optimizeCall(CI)) {
1238 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1244 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1245 // Strip off at most one level of pointer casts, looking for an alloca. This
1246 // is good enough in practice and simpler than handling any number of casts.
1247 Value *Underlying = TrampMem->stripPointerCasts();
1248 if (Underlying != TrampMem &&
1249 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1251 if (!isa<AllocaInst>(Underlying))
1254 IntrinsicInst *InitTrampoline = nullptr;
1255 for (User *U : TrampMem->users()) {
1256 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1259 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1261 // More than one init_trampoline writes to this value. Give up.
1263 InitTrampoline = II;
1266 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1267 // Allow any number of calls to adjust.trampoline.
1272 // No call to init.trampoline found.
1273 if (!InitTrampoline)
1276 // Check that the alloca is being used in the expected way.
1277 if (InitTrampoline->getOperand(0) != TrampMem)
1280 return InitTrampoline;
1283 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1285 // Visit all the previous instructions in the basic block, and try to find a
1286 // init.trampoline which has a direct path to the adjust.trampoline.
1287 for (BasicBlock::iterator I = AdjustTramp,
1288 E = AdjustTramp->getParent()->begin(); I != E; ) {
1289 Instruction *Inst = --I;
1290 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1291 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1292 II->getOperand(0) == TrampMem)
1294 if (Inst->mayWriteToMemory())
1300 // Given a call to llvm.adjust.trampoline, find and return the corresponding
1301 // call to llvm.init.trampoline if the call to the trampoline can be optimized
1302 // to a direct call to a function. Otherwise return NULL.
1304 static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1305 Callee = Callee->stripPointerCasts();
1306 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1308 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1311 Value *TrampMem = AdjustTramp->getOperand(0);
1313 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1315 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1320 // visitCallSite - Improvements for call and invoke instructions.
1322 Instruction *InstCombiner::visitCallSite(CallSite CS) {
1323 if (isAllocLikeFn(CS.getInstruction(), TLI))
1324 return visitAllocSite(*CS.getInstruction());
1326 bool Changed = false;
1328 // If the callee is a pointer to a function, attempt to move any casts to the
1329 // arguments of the call/invoke.
1330 Value *Callee = CS.getCalledValue();
1331 if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1334 if (Function *CalleeF = dyn_cast<Function>(Callee))
1335 // If the call and callee calling conventions don't match, this call must
1336 // be unreachable, as the call is undefined.
1337 if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1338 // Only do this for calls to a function with a body. A prototype may
1339 // not actually end up matching the implementation's calling conv for a
1340 // variety of reasons (e.g. it may be written in assembly).
1341 !CalleeF->isDeclaration()) {
1342 Instruction *OldCall = CS.getInstruction();
1343 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1344 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1346 // If OldCall does not return void then replaceAllUsesWith undef.
1347 // This allows ValueHandlers and custom metadata to adjust itself.
1348 if (!OldCall->getType()->isVoidTy())
1349 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1350 if (isa<CallInst>(OldCall))
1351 return EraseInstFromFunction(*OldCall);
1353 // We cannot remove an invoke, because it would change the CFG, just
1354 // change the callee to a null pointer.
1355 cast<InvokeInst>(OldCall)->setCalledFunction(
1356 Constant::getNullValue(CalleeF->getType()));
1360 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1361 // If CS does not return void then replaceAllUsesWith undef.
1362 // This allows ValueHandlers and custom metadata to adjust itself.
1363 if (!CS.getInstruction()->getType()->isVoidTy())
1364 ReplaceInstUsesWith(*CS.getInstruction(),
1365 UndefValue::get(CS.getInstruction()->getType()));
1367 if (isa<InvokeInst>(CS.getInstruction())) {
1368 // Can't remove an invoke because we cannot change the CFG.
1372 // This instruction is not reachable, just remove it. We insert a store to
1373 // undef so that we know that this code is not reachable, despite the fact
1374 // that we can't modify the CFG here.
1375 new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1376 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1377 CS.getInstruction());
1379 return EraseInstFromFunction(*CS.getInstruction());
1382 if (IntrinsicInst *II = FindInitTrampoline(Callee))
1383 return transformCallThroughTrampoline(CS, II);
1385 PointerType *PTy = cast<PointerType>(Callee->getType());
1386 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1387 if (FTy->isVarArg()) {
1388 int ix = FTy->getNumParams();
1389 // See if we can optimize any arguments passed through the varargs area of
1391 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1392 E = CS.arg_end(); I != E; ++I, ++ix) {
1393 CastInst *CI = dyn_cast<CastInst>(*I);
1394 if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1395 *I = CI->getOperand(0);
1401 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1402 // Inline asm calls cannot throw - mark them 'nounwind'.
1403 CS.setDoesNotThrow();
1407 // Try to optimize the call if possible, we require DataLayout for most of
1408 // this. None of these calls are seen as possibly dead so go ahead and
1409 // delete the instruction now.
1410 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1411 Instruction *I = tryOptimizeCall(CI, DL);
1412 // If we changed something return the result, etc. Otherwise let
1413 // the fallthrough check.
1414 if (I) return EraseInstFromFunction(*I);
1417 return Changed ? CS.getInstruction() : nullptr;
1420 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
1421 // attempt to move the cast to the arguments of the call/invoke.
1423 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1425 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1428 // The prototype of thunks are a lie, don't try to directly call such
1430 if (Callee->hasFnAttribute("thunk"))
1432 Instruction *Caller = CS.getInstruction();
1433 const AttributeSet &CallerPAL = CS.getAttributes();
1435 // Okay, this is a cast from a function to a different type. Unless doing so
1436 // would cause a type conversion of one of our arguments, change this call to
1437 // be a direct call with arguments casted to the appropriate types.
1439 FunctionType *FT = Callee->getFunctionType();
1440 Type *OldRetTy = Caller->getType();
1441 Type *NewRetTy = FT->getReturnType();
1443 // Check to see if we are changing the return type...
1444 if (OldRetTy != NewRetTy) {
1446 if (NewRetTy->isStructTy())
1447 return false; // TODO: Handle multiple return values.
1449 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
1450 if (Callee->isDeclaration())
1451 return false; // Cannot transform this return value.
1453 if (!Caller->use_empty() &&
1454 // void -> non-void is handled specially
1455 !NewRetTy->isVoidTy())
1456 return false; // Cannot transform this return value.
1459 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1460 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1462 hasAttributes(AttributeFuncs::
1463 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1464 AttributeSet::ReturnIndex))
1465 return false; // Attribute not compatible with transformed value.
1468 // If the callsite is an invoke instruction, and the return value is used by
1469 // a PHI node in a successor, we cannot change the return type of the call
1470 // because there is no place to put the cast instruction (without breaking
1471 // the critical edge). Bail out in this case.
1472 if (!Caller->use_empty())
1473 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1474 for (User *U : II->users())
1475 if (PHINode *PN = dyn_cast<PHINode>(U))
1476 if (PN->getParent() == II->getNormalDest() ||
1477 PN->getParent() == II->getUnwindDest())
1481 unsigned NumActualArgs = CS.arg_size();
1482 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1484 // Prevent us turning:
1485 // declare void @takes_i32_inalloca(i32* inalloca)
1486 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
1489 // call void @takes_i32_inalloca(i32* null)
1490 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca))
1493 CallSite::arg_iterator AI = CS.arg_begin();
1494 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1495 Type *ParamTy = FT->getParamType(i);
1496 Type *ActTy = (*AI)->getType();
1498 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
1499 return false; // Cannot transform this parameter value.
1501 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1502 hasAttributes(AttributeFuncs::
1503 typeIncompatible(ParamTy, i + 1), i + 1))
1504 return false; // Attribute not compatible with transformed value.
1506 if (CS.isInAllocaArgument(i))
1507 return false; // Cannot transform to and from inalloca.
1509 // If the parameter is passed as a byval argument, then we have to have a
1510 // sized type and the sized type has to have the same size as the old type.
1511 if (ParamTy != ActTy &&
1512 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1513 Attribute::ByVal)) {
1514 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1515 if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1518 Type *CurElTy = ActTy->getPointerElementType();
1519 if (DL->getTypeAllocSize(CurElTy) !=
1520 DL->getTypeAllocSize(ParamPTy->getElementType()))
1525 if (Callee->isDeclaration()) {
1526 // Do not delete arguments unless we have a function body.
1527 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1530 // If the callee is just a declaration, don't change the varargsness of the
1531 // call. We don't want to introduce a varargs call where one doesn't
1533 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1534 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1537 // If both the callee and the cast type are varargs, we still have to make
1538 // sure the number of fixed parameters are the same or we have the same
1539 // ABI issues as if we introduce a varargs call.
1540 if (FT->isVarArg() &&
1541 cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1542 FT->getNumParams() !=
1543 cast<FunctionType>(APTy->getElementType())->getNumParams())
1547 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1548 !CallerPAL.isEmpty())
1549 // In this case we have more arguments than the new function type, but we
1550 // won't be dropping them. Check that these extra arguments have attributes
1551 // that are compatible with being a vararg call argument.
1552 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1553 unsigned Index = CallerPAL.getSlotIndex(i - 1);
1554 if (Index <= FT->getNumParams())
1557 // Check if it has an attribute that's incompatible with varargs.
1558 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1559 if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1564 // Okay, we decided that this is a safe thing to do: go ahead and start
1565 // inserting cast instructions as necessary.
1566 std::vector<Value*> Args;
1567 Args.reserve(NumActualArgs);
1568 SmallVector<AttributeSet, 8> attrVec;
1569 attrVec.reserve(NumCommonArgs);
1571 // Get any return attributes.
1572 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1574 // If the return value is not being used, the type may not be compatible
1575 // with the existing attributes. Wipe out any problematic attributes.
1577 removeAttributes(AttributeFuncs::
1578 typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1579 AttributeSet::ReturnIndex);
1581 // Add the new return attributes.
1582 if (RAttrs.hasAttributes())
1583 attrVec.push_back(AttributeSet::get(Caller->getContext(),
1584 AttributeSet::ReturnIndex, RAttrs));
1586 AI = CS.arg_begin();
1587 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1588 Type *ParamTy = FT->getParamType(i);
1590 if ((*AI)->getType() == ParamTy) {
1591 Args.push_back(*AI);
1593 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy));
1596 // Add any parameter attributes.
1597 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1598 if (PAttrs.hasAttributes())
1599 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1603 // If the function takes more arguments than the call was taking, add them
1605 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1606 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1608 // If we are removing arguments to the function, emit an obnoxious warning.
1609 if (FT->getNumParams() < NumActualArgs) {
1610 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1611 if (FT->isVarArg()) {
1612 // Add all of the arguments in their promoted form to the arg list.
1613 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1614 Type *PTy = getPromotedType((*AI)->getType());
1615 if (PTy != (*AI)->getType()) {
1616 // Must promote to pass through va_arg area!
1617 Instruction::CastOps opcode =
1618 CastInst::getCastOpcode(*AI, false, PTy, false);
1619 Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1621 Args.push_back(*AI);
1624 // Add any parameter attributes.
1625 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1626 if (PAttrs.hasAttributes())
1627 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1633 AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1634 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1635 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1637 if (NewRetTy->isVoidTy())
1638 Caller->setName(""); // Void type should not have a name.
1640 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1644 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1645 NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1646 II->getUnwindDest(), Args);
1648 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1649 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1651 CallInst *CI = cast<CallInst>(Caller);
1652 NC = Builder->CreateCall(Callee, Args);
1654 if (CI->isTailCall())
1655 cast<CallInst>(NC)->setTailCall();
1656 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1657 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1660 // Insert a cast of the return type as necessary.
1662 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1663 if (!NV->getType()->isVoidTy()) {
1664 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
1665 NC->setDebugLoc(Caller->getDebugLoc());
1667 // If this is an invoke instruction, we should insert it after the first
1668 // non-phi, instruction in the normal successor block.
1669 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1670 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1671 InsertNewInstBefore(NC, *I);
1673 // Otherwise, it's a call, just insert cast right after the call.
1674 InsertNewInstBefore(NC, *Caller);
1676 Worklist.AddUsersToWorkList(*Caller);
1678 NV = UndefValue::get(Caller->getType());
1682 if (!Caller->use_empty())
1683 ReplaceInstUsesWith(*Caller, NV);
1684 else if (Caller->hasValueHandle()) {
1685 if (OldRetTy == NV->getType())
1686 ValueHandleBase::ValueIsRAUWd(Caller, NV);
1688 // We cannot call ValueIsRAUWd with a different type, and the
1689 // actual tracked value will disappear.
1690 ValueHandleBase::ValueIsDeleted(Caller);
1693 EraseInstFromFunction(*Caller);
1697 // transformCallThroughTrampoline - Turn a call to a function created by
1698 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1699 // underlying function.
1702 InstCombiner::transformCallThroughTrampoline(CallSite CS,
1703 IntrinsicInst *Tramp) {
1704 Value *Callee = CS.getCalledValue();
1705 PointerType *PTy = cast<PointerType>(Callee->getType());
1706 FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1707 const AttributeSet &Attrs = CS.getAttributes();
1709 // If the call already has the 'nest' attribute somewhere then give up -
1710 // otherwise 'nest' would occur twice after splicing in the chain.
1711 if (Attrs.hasAttrSomewhere(Attribute::Nest))
1715 "transformCallThroughTrampoline called with incorrect CallSite.");
1717 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1718 PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1719 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1721 const AttributeSet &NestAttrs = NestF->getAttributes();
1722 if (!NestAttrs.isEmpty()) {
1723 unsigned NestIdx = 1;
1724 Type *NestTy = nullptr;
1725 AttributeSet NestAttr;
1727 // Look for a parameter marked with the 'nest' attribute.
1728 for (FunctionType::param_iterator I = NestFTy->param_begin(),
1729 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1730 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1731 // Record the parameter type and any other attributes.
1733 NestAttr = NestAttrs.getParamAttributes(NestIdx);
1738 Instruction *Caller = CS.getInstruction();
1739 std::vector<Value*> NewArgs;
1740 NewArgs.reserve(CS.arg_size() + 1);
1742 SmallVector<AttributeSet, 8> NewAttrs;
1743 NewAttrs.reserve(Attrs.getNumSlots() + 1);
1745 // Insert the nest argument into the call argument list, which may
1746 // mean appending it. Likewise for attributes.
1748 // Add any result attributes.
1749 if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1750 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1751 Attrs.getRetAttributes()));
1755 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1757 if (Idx == NestIdx) {
1758 // Add the chain argument and attributes.
1759 Value *NestVal = Tramp->getArgOperand(2);
1760 if (NestVal->getType() != NestTy)
1761 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1762 NewArgs.push_back(NestVal);
1763 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1770 // Add the original argument and attributes.
1771 NewArgs.push_back(*I);
1772 AttributeSet Attr = Attrs.getParamAttributes(Idx);
1773 if (Attr.hasAttributes(Idx)) {
1774 AttrBuilder B(Attr, Idx);
1775 NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1776 Idx + (Idx >= NestIdx), B));
1783 // Add any function attributes.
1784 if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1785 NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1786 Attrs.getFnAttributes()));
1788 // The trampoline may have been bitcast to a bogus type (FTy).
1789 // Handle this by synthesizing a new function type, equal to FTy
1790 // with the chain parameter inserted.
1792 std::vector<Type*> NewTypes;
1793 NewTypes.reserve(FTy->getNumParams()+1);
1795 // Insert the chain's type into the list of parameter types, which may
1796 // mean appending it.
1799 FunctionType::param_iterator I = FTy->param_begin(),
1800 E = FTy->param_end();
1804 // Add the chain's type.
1805 NewTypes.push_back(NestTy);
1810 // Add the original type.
1811 NewTypes.push_back(*I);
1817 // Replace the trampoline call with a direct call. Let the generic
1818 // code sort out any function type mismatches.
1819 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1821 Constant *NewCallee =
1822 NestF->getType() == PointerType::getUnqual(NewFTy) ?
1823 NestF : ConstantExpr::getBitCast(NestF,
1824 PointerType::getUnqual(NewFTy));
1825 const AttributeSet &NewPAL =
1826 AttributeSet::get(FTy->getContext(), NewAttrs);
1828 Instruction *NewCaller;
1829 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1830 NewCaller = InvokeInst::Create(NewCallee,
1831 II->getNormalDest(), II->getUnwindDest(),
1833 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1834 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1836 NewCaller = CallInst::Create(NewCallee, NewArgs);
1837 if (cast<CallInst>(Caller)->isTailCall())
1838 cast<CallInst>(NewCaller)->setTailCall();
1839 cast<CallInst>(NewCaller)->
1840 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1841 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1848 // Replace the trampoline call with a direct call. Since there is no 'nest'
1849 // parameter, there is no need to adjust the argument list. Let the generic
1850 // code sort out any function type mismatches.
1851 Constant *NewCallee =
1852 NestF->getType() == PTy ? NestF :
1853 ConstantExpr::getBitCast(NestF, PTy);
1854 CS.setCalledFunction(NewCallee);
1855 return CS.getInstruction();